My Feed, My Way: The Strategy Comes Into View — And So Does Its Theory of Change Problem

~ alexiamaddox ~ Leave a comment

I want to be careful about what kind of post this is, because it isn’t the same kind of post as “Another Day, Another Declaration.” That one was about a newsletter dressed up as a victory lap for evidence that didn’t support it. This one is different. This is the moment the actual strategy becomes visible — not another announcement to catalogue, but the point where you can finally see the shape of the thing the government has been building toward, and ask whether it hangs together.

It doesn’t, and I want to show my working on why.

What actually landed

On 8 September, Anthony Albanese and Anika Wells announced the draft legislation for the Digital Duty of Care bill including an “Australian first” called My Feed, My Way.

Social media platforms will have to notify users of a choice, according to the press release: opt in to an algorithmically curated feed, or opt out and see only the accounts they follow, in order. That’s the government’s stated intent, not yet the law’s text. What the bill itself creates is a ministerial power to require “user empowerment tools” by legislative instrument — the notification, the two named options, the commitment to “respect that choice,” none of it is written into the Act. It’s a promise about how that power will eventually be used, made at the press conference rather than in the clause.

The parts that are actually drafted, and drafted specifically, are elsewhere. Under-18 protections extend into design features — addictive mechanics, self-esteem effects — and into content categories: eating disorder promotion, misogynistic material, pornography, crime glorification, content causing serious mental health distress. eSafety gets removal powers over nudify apps, a streamlined cyber-abuse scheme, and — this is the part with actual teeth — the power to use registered researchers and its own investigators to test, directly, what platforms are serving users, rather than relying on platforms to self-report. Penalties run to $109.2 million, per the government’s own figure. Legislation — still only an exposure draft, open for “targeted consultation” — is promised to be introduced before year’s end.

Read on its own, it’s a reasonable, even overdue, piece of platform accountability. Read in sequence, it’s the fourth or fifth distinct policy instrument Australia has produced on this issue in under two years, and each one was built to solve the problem the previous one created or failed to solve. That’s the part worth sitting with.

I wanted to ask a few questions even at this point so I’ll drop them in here as a by way and then get back in track with the analysis via the theory of change discussion.

Who the “registered/approved researchers” actually are — as far as anyone can currently say:

The bill’s own definition, once you read it rather than the press coverage of it, is narrower than “such as those from an Australian university” implies. Section 205B doesn’t offer university employment as an example of who might qualify — it’s a strict, three-part gate: to be an “approved researcher” you must (a) be employed by an Australian university, (b) be individually approved by the Commissioner under legislative rules, and (c) meet whatever other requirements those rules end up specifying. Independent researchers, journalists, and civil-society investigators are categorically excluded by definition, not just underrepresented in practice. The research itself has to clear a further bar too — approved by a university ethics committee and “of a kind prescribed by the legislative rules to be in the public interest,” a phrase that, like almost everything operational in this bill, is a placeholder for rules that don’t exist yet.

And the access itself is a power to create a scheme, not the scheme. Section 205C says the legislative rules “may” establish one or more “data access schemes” requiring platforms to hand data to approved researchers — what kinds of data, for what kinds of research, under what process, at what cost to the researcher, is all left to future rule-making. Nothing in the Bill as drafted actually obliges a platform to hand anything over yet. The sock-puppet power fares slightly better on specificity — sections 205G to 205L spell out clearly that both approved researchers and the Commissioner personally may assume false identities to create accounts, observe and download material, test platform features, and probe how a service responds to particular actions, with good-faith immunity from civil action built in. That’s genuinely detailed, operative law, not a placeholder — worth distinguishing from the data-access side, which is still almost entirely promissory.

This is functionally Australia’s version of the EU DSA’s Article 40 “vetted researcher” scheme, which is worth naming as a direct comparator, because that scheme’s actual track record is a caution, not a template to feel reassured by. EU vetted-researcher status is notoriously slow to grant. The Delegated Act specifies an 80-working-day decision window — and even well-prepared applicants routinely fail: a Dutch team with prior European Research Council pilot experience on Article 40 implementation, investigating TikTok’s role in Romania’s contested 2024 election, still failed on five of seven eligibility criteria. Eligibility is nominally broad, extending to universities, non-profits and civil society researchers, but studies of the first wave of applications found platforms rejecting requests via narrow readings of “systemic risk,” excluding non-academic applicants in practice, and handing over incomplete data even once access was granted.Australia’s version adds its own extra layer of narrowness on top — the university-employment requirement is tighter than anything explicit in Article 40 — while leaving the actual data-access obligation exactly as undefined as the EU’s was at the equivalent stage. If Australia copies the label without copying — or improving on — the implementation detail, “approved researcher” risks becoming exactly the kind of unresolved gap Josh Taylor flagged with the opt-out mechanism itself: real on paper, undefined in practice, left for platforms and regulators to fight out later.

On the AIO —parallel timing

A Conversation piece published back in July — “Australia wants a ‘digital duty of care’. But how will we check what Big Tech is doing?” — presented the argument from Australian Digital Media researchers that the duty of care needs an “ecosystem of observability” combining three things: regulatory powers to compel platform data, independent research infrastructure (data donation, browser/mobile measurement, secure research environments), and stronger individual rights for Australians to access, download and donate their own platform data. They named the AIO directly as “one model for this infrastructure.” So this isn’t two separate efforts converging by coincidence — it’s the same research network pushing on two levers at once: build the bottom-up data-donation infrastructure regardless of what parliament does, and publicly campaign for the top-down regulatory access powers that just showed up in the exposure draft two months later.

During the same week’s coverage: Chanel Contos’s National Press Club appearance and the NSW Government’s urgent September 4 roundtable following the Sydney schoolboy assault allegations are being cited directly by other commentary as part of the pressure that shaped this exposure draft — her “23 minutes to misogynistic content” statistic is functioning as the kind of concrete evidence the researcher-access provisions are explicitly meant to make independently verifiable going forward, rather than something advocates have to keep re-measuring themselves via burner-phone experiments.

And what does “harmful content” mean?

And while we’re here going down byways, there is another point worth pulling from the ADM+S policy brief , because it sharpens the “harmful content” categories in the bill in a way that matters beyond researcher access.

Back in June, before any of this had bill text, ADM+S and the University of Melbourne’s CAIDE published a policy brief responding to the government’s Issues Paper, arguing the duty of care should move “beyond specifying a set of narrowly defined risks to safety… to require platforms to address broader, systemic risks.” Their central worry, laid out as a named case study, was harm-reduction and public health communication — evidence-based outreach on HIV and drug use has a documented history of suppression, account restriction and rejected advertising on these platforms, not through deliberate policy but as “the inevitable by-product of automated content moderation systems trained to identify, suppress and remove adult or offensive content.” They predicted, in writing, that a narrowly-defined risk-based model wouldn’t fix this — “risk” captures potential harms, not the absence of legitimate positive content — and proposed concrete affirmative protections instead: moderation exemptions for government-funded health organisations, expedited review for wrongly suppressed public-health material, regular audits of moderation systems’ impact on health content.

None of that made it into the bill. What section 25C actually contains is a narrowly-defined categorical list — eating disorders, misogynistic content, terrorism, crime, and, among the entries the government didn’t put in its own press release, material that “encourages, promotes, urges or instructs illicit drug use.” That’s precisely the drafting style ADM+S told government, months earlier, wouldn’t protect the harm-reduction communication they were worried about — and the affirmative protections they proposed to fix it are nowhere in the draft. So this isn’t a hypothetical risk a colleague raised after the fact. It’s a predicted, documented harm, from the government’s own consulted research network, that the exposure draft appears to walk straight into

A theory of change that was available, and wasn’t used

Now, getting back to where we were in the chronological chain of events. Here’s the thing that gets lost every time this gets reported as a fresh initiative: the design-first version of this reform already existed, in 2024, before the account ban was legislated. Zoe Daniel’s private member’s bill on platform duty of care predates the Social Media Minimum Age Act. It was left to lapse. The government reached for the ban instead — a chain that started with a South Australian premier’s Mother’s Day 2024 social media announcement and a News Corp campaign the following week, gathered momentum through the year via other state premiers and an opposition election promise, and culminated in legislation drafted, introduced and passed in under two weeks, with a 24-hour public comment window, in the last sitting week before the federal election.

I’ve written before about what that sequencing choice cost in evidentiary terms. What I want to name now is what it cost architecturally. Once the ban passed, an entire enforcement apparatus grew up around it: the Phase 2 industry codes, the age-assurance vendor contracts, the trusted-provider lists, eSafety’s document-demand powers reaching into third-party verification companies. That apparatus is not neutral scaffolding sitting empty, waiting for a better policy to move in. It has vendors with revenue tied to its continuation, a regulator whose newly expanded enforcement muscle (the Online Safety Amendment Bill, moving through parliament the same week) reaches further into that same verification paradigm, and a political constituency built around “world-first” account restriction as the signature achievement.

So when the duty of care — the reform actually aimed at design, not access — finally arrives, it doesn’t land in an empty room where design-first thinking can shape the whole architecture from scratch. The bill does sweep away one layer of the pre-existing apparatus — Schedule 3 repeals the general online content scheme codes entirely, folding content-classification into the unified duty of care. But the age-verification layer this apparatus was actually built around — the Phase 2 codes, the vendor contracts, the trusted-provider lists — survives untouched, with its penalties increased the same week. The duty of care has to be built around that surviving core, not replace it. This isn’t a coherent strategy unfolding in the right order. It’s a good instrument arriving after a worse one has already poured the foundation, and having to accommodate the foundation rather than correct it.

The mechanism undercuts its own justification

There’s a second problem, sitting inside the bill itself rather than in its timing. The government’s language around My Feed, My Way is explicitly universalist — Wells frames it as “basic standards for the online products we use every day,” alongside cars, toys, food. That’s inclusive-design logic: build the protection into the system for everyone, rather than trying to identify and specially protect a vulnerable subset. It’s good theory. Universal design works because the protected state is the automatic one — nobody has to find it, understand it, or activate it. Curb cuts don’t require a wheelchair user to request one.

But My Feed, My Way is an opt-out, not an opt-in. The algorithmically optimised, high-engagement default stays exactly where it is; the protective state is the thing a user has to notice, understand, and deliberately choose. We already know what that produces, because we’ve watched the experiment run. Under oath in the Oakland federal trial last August, Instagram’s Adam Mosseri confirmed an internal Meta document had put uptake of its “Take a Break” feature — designed to interrupt exactly the kind of compulsive scrolling this bill is meant to address — at 1.8%, a figure the company never disclosed even as it publicly touted a 90% retention rate among the small number who’d turned it on. Meta’s own internal research pointed the same direction on a related question: making teen accounts private by default, rather than opt-in, was projected to prevent 5.4 million unwanted daily message interactions. The lesson, from the company’s own numbers, is not subtle — default state decides outcomes; availability of a safer option barely moves anything at all

The EU has required platforms to offer a non-algorithmic feed option since February 2024 — and in October 2025, an Amsterdam court still had to order Meta to fix it, giving the company two weeks to comply under threat of a €5 million penalty. As Bits of Freedom’s Rejo Zenger, who brought the case, put it: “even if users can switch feeds, they cannot set a preferred feed that persists, and the platform continually nudges users back to the profile feed.” Meta disputed the ruling and appealed — and for now, the binding order applies only in the Netherlands, not EU-wide. If the justification for going broad rather than narrow is genuinely inclusive-design thinking, the logically consistent implementation is an opt-in default, which is what the Greens and Chanel Contos have been arguing for. The government chose the version that photographs the same way in a press release but almost certainly won’t behave the same way in practice.

Who this doesn’t reach, and who’s left holding it

And then there’s the population question, which is the one that should worry anyone thinking about this as a child-safety measure rather than a general consumer-choice one. The opt-out applies to users over 16. The government’s own account ban was supposed to mean under-16s aren’t the audience for this at all — they’re not meant to be logged in, so the question of their feed algorithm shouldn’t arise.

Except eSafety’s own evaluation data says otherwise. Three months in, account ownership among under-16s had fallen from 52.4% to 42.1% — real, but partial. Of the children who kept accounts, the platform-by-platform breakdown shows two distinct failure modes, both damning in their own way. On YouTube, nearly half never even got asked to verify their age — no prompt at all. On Facebook, the single most common reason was different and arguably worse: the age already listed on the account was 16 or above, meaning a check did happen and simply accepted a false answer. Instagram and Snapchat sit in between, with “never asked” narrowly ahead of “false age accepted” on both. Either way, “reasonable steps” is failing at more than one point in the process — not just an absent gate, but a gate that doesn’t hold. Independently, the University of Newcastle’s BMJ study found more than 85% of under-16s still using restricted platforms at the same three-month mark.

Meanwhile the adults meant to be managing all of this — parents and teachers — have less visibility than they had before, not more. eSafety’s own evaluation found parental awareness of children’s social media use actually declined after the ban took effect, concentrated among parents of girls and 10–12 year olds. Teachers, who inherit whatever happened in a student’s feed the moment it walks into a classroom the next morning, get no new tools at all in this bill — the digital literacy and relational-education infrastructure that would let schools and families actually build capability, rather than just manage fallout, isn’t part of the package. It’s the same gap Lisa Given flagged about the earlier duty of care consultation: real teeth on paper, aimed at platforms, with nothing resourced for the people standing between the platform and the child.

The generational bet nobody has actually defended

There’s a longer-run theory of change sitting underneath all of this that almost never gets said out loud, and I think it deserves to be dragged into the open, because I’m not convinced it survives contact with the evidence we already have.

The theory goes something like this: the ban won’t look like it’s working on the generation that’s already on these platforms, because they were socialised into algorithmic social media before the restriction existed, and restriction always looks like resistance in its first cohort. But the next generation — Gen Alpha, kids who are eight, nine, ten right now — will turn 16 having never legally held an account. For them, the age gate won’t be an imposition on an established habit. It’ll just be an unremarkable fact of childhood, the way not being allowed to drive at twelve is. Robinson, La Sala and Harrison named this directly in their 2025 title: is this a “seatbelt moment” — a restriction that looks costly and contested at first and becomes invisible, uncontested background norm within a generation — or a missed opportunity dressed up as one?

It’s a coherent theory. Seatbelts really did work that way: resistance was highest among drivers who’d learned to drive without them, and uptake became close to automatic once a full generation had never known driving any other way. If social media restriction follows the same curve, today’s dismal compliance numbers — 85%+ of under-16s still using restricted platforms, half of retained accounts never even prompted for age checks — aren’t a verdict on the policy. They’re what a first cohort always looks like, and the real test is what a fully Gen Alpha 15-year-old’s relationship to these platforms looks like in 2032.

But I want to push on two things before I let that theory do any work, because right now it’s operating as an unexamined assumption rather than something anyone has actually argued for or is measuring.

The first is empirical, and it’s sitting in plain sight in the argument over whether “brainrot” belongs to Gen Z or Gen Alpha. The honest answer is both, and how it belongs to both is the point: Gen Z, with full algorithmic platform access, generated and refined this content natively — the irony, the absurdist compression, the in-group signalling. Gen Alpha, meanwhile, absorbed an enormous amount of it into daily vocabulary and offline culture without needing platform accounts of their own to do it. Skibidi Toilet, sigma, rizz — this is a nine-year-old’s vocabulary now, transmitted through older siblings, playground repetition, YouTube (which sits partly outside the restriction architecture), and family devices, not through a personal, algorithmically curated feed. That’s a direct empirical challenge to the seatbelt theory’s core mechanism. Seatbelt norms shifted because the behaviour itself — driving unbelted — has no meaningful peer-transmission pathway once you’re not behind a wheel. Platform culture has an enormous one. Excluding a nine-year-old from an Instagram account doesn’t wall them off from the platform’s cultural output; it just changes the delivery mechanism from direct and individually curated to lateral and peer-mediated. If the goal was a generation genuinely insulated from what these platforms produce, the account ban may be solving the wrong layer of the problem entirely — the content diffuses regardless of who’s logged in.

The second problem is normative, and it’s the sharper one: even if the seatbelt theory worked exactly as advertised, is a childhood spent in what we might call a stigmatised relationship to social media — access as illicit, policed, something you get around rather than something you’re taught to navigate — actually the outcome we want? We have a template for what stigma-based restriction does to adolescent behaviour, from decades of underage drinking and smoking policy, and it isn’t uncomplicated normalisation. It’s secrecy from the adults meant to be guiding you — which is precisely what eSafety’s own data already shows happening, with parental awareness of children’s social media use declining, not rising, since the ban took effect. It’s status economies built around successful circumvention rather than around competent use — TikTok how-to guides on evading TikTok’s own age gate, published on the platform being regulated, are not a normalisation success story, they’re the underground-economy version of the same dynamic. And it forecloses exactly the capability-building the digital literacy argument I made above is trying to protect. A generation that grows up treating platform access as contraband to be smuggled rather than a tool to be taught isn’t obviously better prepared for the platforms they’ll use as adults — a cliff-edge from total exclusion to unrestricted access at 16, with no graduated exposure and no explicit skill-building in between, is a pedagogy problem dressed up as a public health win.

And here’s what makes this a genuine theory-of-change problem rather than just a debate to have later: nobody is actually measuring which of these dynamics is occurring. eSafety’s longitudinal evaluation tracks account status, usage duration, and self-reported wellbeing. It isn’t set up to track cultural diffusion, peer-transmission of platform content among excluded users, or whether the relationship young people have to these platforms is becoming one of literacy or one of stigma. We’re running a generational bet — quietly assuming the seatbelt curve will apply, structuring policy timelines around a 2027–2028 review window that presumes it — without building the instruments that would tell us, a decade from now, whether we bet on the right mechanism at all.

A brief detour through brain rot, because it’s not really a detour

I want to pause on something that looks like a side issue and isn’t, because it’s sitting in the same water as everything above it, and because I keep watching it get used as if it settles an argument it can’t actually settle.

“Brain rot” was Oxford’s word of the year for 2024, and since then it’s been doing an enormous amount of unexamined work in exactly this debate. The trouble is that nobody using the term is using it to mean the same thing. Pull apart the ways it’s actually deployed and you get at least four distinct referents wearing one label: a content genre (the singing toilets, the AI-slop, the absurdist short-form video); a slang register that circulates alongside that genre and, tellingly, migrates into offline speech among children who’ve never held a personal account — which is itself a data point against the idea that account restriction meaningfully insulates a generation from platform culture, since the vocabulary clearly doesn’t need an account to travel; a subjective experiential state, the “fogged-out, low-selectivity” feeling Maxi Heitmayer’s Gen Z interview subjects describe, in which the content is a response to an already-depleted state rather than its cause; and, most recently, a proposed clinical construct — a psychometric “Brain Rot Scale,” explicitly modelled on substance-addiction neurobiology, complete with subscales for Attention Dysregulation, Digital Compulsivity and Cognitive Dependency, tested on an Egyptian convenience sample and explaining a modest 35% of variance, with the authors themselves conceding it hasn’t been validated against any existing measure of problematic internet use. Thoreau, who coined the term in Walden in 1854 “will not any endeavor to cure the brain-rot, which prevails so much more widely and fatally?” meant something different again: not an individual affliction at all, but a society choosing simple ideas over complex ones.

What’s genuinely funny, once you go looking, is that Thoreau didn’t stop at the diagnosis, he named the genre. A chapter later, in “Reading,” he takes aim at the popular fiction of his day, cheap, serialised, mass-produced, and devoured compulsively by readers (TikTok rabbit holes anyone?) he compares to cormorants, “who can digest all sorts of this, even after the fullest dinner.” He even mocks the marketing copy, inventing a title that reads uncannily like a lost brainrot video: “The Skip of the Tip-Toe-Hop, a Romance of the Middle Ages… to appear in monthly parts; a great rush; don’t all come together.” And his diagnosis of what this content does to its readers could be lifted whole into any 2026 op-ed: “dulness of sight, a stagnation of the vital circulations, and a general deliquium and sloughing off of all the intellectual faculties.” Deliquium and sloughing off of the intellectual faculties. Nineteenth-century Thoreau just described the brain-rot feeling, named its genre, mocked its marketing, and diagnosed its symptoms, a hundred and seventy years before anyone thought to build a psychometric scale for it. He even had a theory of supply: “this sort of gingerbread is baked daily and more sedulously than pure wheat or rye-and-Indian in almost every oven, and finds a surer market” — which is just the engagement economy, minus the engagement.

All of which is a nice bit of trivia (and it’s my blog, so good times), but it’s also a useful check on where the seriousness actually lives in this debate. Thoreau could name the genre, mock the marketing, and diagnose the feeling — but he never mistook any of that for a discovered fact about the brain. That’s the distinction the current moment keeps losing. This is Ferguson’s construct balkanization problem, which I cited earlier in relation to Rausch and Haidt, showing up again in a completely different corner of the same debate: a folk category getting formalised into something that reads as diagnostic before anyone has established it’s distinct from constructs we already have names for. A scale with excellent internal reliability tells you the instrument is internally consistent. It doesn’t tell you “brain rot” is a real, discrete thing rather than a relabelling of attention difficulties, compulsive checking and existing problematic-use patterns that already have measures.

And it matters which definition is doing the work, because they don’t point toward the same account of what’s happening, let alone the same fix. Compare three recent, more careful pieces of evidence side by side. The passive-sensing study on adolescent phone use I wrote about last week found that low mood predicts more scrolling the next day — a general, content-agnostic disengagement response, consistent with mood management theory: feeling bad, reach for something that costs nothing. Doomscrolling, the pandemic-era construct, is narrower and arguably runs on a different engine entirely — purposeful seeking of threat-relevant, distressing content, closer to anxious vigilance than numbing, a coping strategy that backfires by compounding the very anxiety it’s trying to manage. And Heitmayer’s brain rot, on his own account, sits closer to the first mechanism than the second: his participants describe seeking out content that makes no demand and carries no threat, the near-opposite of doomscrolling’s compulsive checking of bad news. Three different psychological stories, three different underlying mechanisms, one increasingly interchangeable vocabulary — and each new content cycle seems to arrive with a fresh label rather than a resolved question about whether the last one was even right.

Here’s the useful, if slightly deflating, thing this detour actually tells us. Ask what any of this changes about the case for something like My Feed, My Way, and the honest answer is: less than the discourse implies. The strongest evidence for regulating engagement-optimised feeds doesn’t need brain rot, in any of its four senses, to be real. It’s sitting in independent, peer-reviewed research conducted with Meta’s cooperation: when Facebook and Instagram users were switched to chronological feeds for three months in 2020, as part of Meta’s own Election Study with outside academics, they spent measurably less time on the platforms, “suggesting they had become less compelling” — about as clean a demonstration as exists that the algorithmic ranking isn’t neutral scaffolding, it’s the thing actively manufacturing the engagement everyone is worried about. (The same study found the changed feed didn’t measurably shift polarization or political attitudes — a different question from the one this section is asking.)

That evidence stands regardless of whether “brain rot” turns out to be attention dysregulation, mood-driven disengagement, cultural slang, or nothing distinct at all. Which means that when brain rot gets invoked as the reason we need to act — rather than the algorithmic economics standing on their own — it isn’t adding evidentiary weight, it’s borrowing urgency from a construct nobody has actually pinned down. That’s precisely the move I flagged at the top of this post: evidence handled like a fridge magnet, taken out of the room it lives in and put somewhere it can do a different job. The tobacco analogy one of Heitmayer’s own interview subjects reached for — regulate the industry, don’t just tell the individual to smoke less — doesn’t require anyone to have first proven brain rot is a diagnosable condition. It only requires the engagement economics to be what they demonstrably already are.

All over the shop, precisely

So: is this a coherent theory of change finally showing its hand? I don’t think that’s quite right either, and I want to be precise about the claim I’m making, because “incoherent” undersells it. There have been several theories of change here, each internally sensible on its own terms, arriving in the wrong order and now stacked on top of each other rather than replacing one another. The ban’s theory: exclude the population, protect by absence. The industry codes’ theory: verify identity, gate by age. The duty of care’s theory: fix the system, protect by design, regardless of who’s in it. And running underneath all three, unstated and untested: the seatbelt theory, betting that generational turnover will eventually make the first three retroactively look right, on a timeline nobody has committed to actually measuring. Each of these is individually defensible. None of them were built with the others in mind, because none of them were legislated — or even articulated — in the order a single coherent strategy would have chosen. Design-first was available, and shelved, in 2024.

What we’re watching now isn’t a strategy unfolding. It’s a government retrofitting the reform it should have started with onto infrastructure built by the reform it reached for instead, while badging the least structurally demanding piece of the retrofit — a toggle most users won’t find and Meta’s own numbers suggest they wouldn’t use if they did — as the headline. The eSafety document-demand powers, the researcher testing authority, the requirement that platforms document and maintain their harm-mitigation measures: that’s the part of this bill actually shaped by lessons learned, including from the EU’s own enforcement failures. It’s riding into parliament on the opt-out’s press cycle because it isn’t sellable on its own.

Even that’s only half true, though — the same research network likely got the researcher-access and data-scheme provisions largely adopted, while their central structural recommendation, an outcomes-based duty capable of protecting legitimate content like harm-reduction communication from algorithmic over-suppression, was not. Expert input didn’t fail to shape this bill. It shaped the parts that were easiest to say yes to.

I said in July that it was too early to call the ban a failure, and that cuts both ways — too early to call it a success, too. I’ll say something similar here: it’s too early to know whether the duty of care’s substantive parts survive Senate negotiation intact, or get diluted the way the penalty language already seems to have moved between May’s turnover-linked formula and September’s flat cap. That’s not just a hedge — it’s built into the bill’s own timeline. Schedules 2 and 3, the duty of care itself and the repeal of the old content-scheme codes, don’t commence until twelve months after Royal Assent; only the takedown-notice powers in Schedule 1 take effect immediately. But it isn’t too early to say the sequencing was a choice, not an accident, and that the bill in front of parliament right now is being asked to do the work of correcting an architecture it had no hand in designing.

The social media age ban saga: Another Day, Another Declaration

~ alexiamaddox ~ Leave a comment

I’m writing this post because Australia’s social media ban has become a case study in how evidence gets bent to fit a conclusion that was decided before the data arrived — and every time someone with a platform declares victory, that framing hardens a little further into public memory, whether it holds up or not. Documenting it as it happens is the only way I can see, from my skill set, to keep the record straight before it sets.

[sigh] So here I go. I’m really trying not to harp on here, and I do try to progress the discussion, but there are a lot of recursive loops and logics to navigate.

I opened Jon Haidt, Ravi Iyer and Zach Rausch’s latest newsletter and there it was, sitting right under the headline like a benediction: “A journey of a thousand miles begins with a single step.” Lao Tzu, they tell us — though they can’t resist a little joke first, attributing an invented second sentence (“But if that first step is hard, then you should quit”) to unnamed critics, as if anyone doubting six months of shaky data is the one being glib here.

So I went back to Chapter 64 itself, instead of the fridge-magnet version. It’s a strange passage to reach for if you’re trying to defend a law that was drafted, introduced and passed in under two weeks with a 24-hour public comment window. The chapter gives two pieces of advice that I want to pick up here, and the ban’s defenders have managed to miss both of them. The first: deal with problems while they’re small. “That which is at rest is easy to be kept hold of… break it while it is feeble, scatter it while it is small. Act before it exists, regulate before disorder.” Platform design harms didn’t arrive last December — they’ve been documented, litigated and written about for the better part of a decade, and the instrument built to address them at the design level, the Digital Duty of Care, was left to lapse while the government reached for the bluntest tool available instead. The second argument: be as careful at the end as you were at the beginning. “The common people, in their undertakings, fail on the eve of success,” the same chapter says. “If they were as prudent at the end as they are at the beginning, there would be no such failures.” That’s not an argument for patience — it’s a warning against declaring success early and getting careless right as the real test arrives. Lao Tzu, I suspect, would have had more to say about a government that let the problem grow large before acting, and commentators that want to call an unfinished job a win, than he would about critics being impatient.

That’s the kind of move I kept running into working through this. Not lies, exactly — just evidence handled the way you’d handle a fridge magnet: taken out of the room it lives in, polished up, and put somewhere it can do a different job than the one it was built for. And once I saw Haidt’s name at the top, I knew what kind of piece this was going to be…

Here’s what got me. Not that Haidt is optimistic — optimism isn’t a crime. It’s that he’s declaring victory using studies that, when you actually open them up, say close to the opposite of what he’s implying. That’s not interpretation. That’s not “reasonable people reading the same data differently.” That’s citing a source for a conclusion the source itself explicitly rejects.

I’m not going to refute this newsletter line by line. That’s not the point of this post, and honestly, it gives the piece more structural respect than it’s earned. What I want to do instead is lay out the tensions this whole saga has surfaced — the ones sitting underneath the “it’s working” narrative, that I don’t think most readers following this story casually may have had the opportunity to piece together.

The pattern isn’t new

This isn’t the first time Haidt’s relationship with evidence has been flagged. Candice Odgers, reviewing The Anxious Generation in Nature back in 2024, called him “a gifted storyteller” whose “tale is currently one searching for evidence” — noting that hundreds of researchers looking for the effects he describes have found “a mix of no, small and mixed associations.” She pointed to a 72-country analysis of nearly a million people, which found no evidence that the global rollout of social media was associated with widespread psychological harm — if anything, the associations ran mildly the other way, with higher adoption linked to slightly better wellbeing, particularly among younger users. The study’s own authors are careful about what this does and doesn’t show: the associations were small, descriptive rather than causal, and drawn from a single platform’s data at a national level — not proof that social media is good for people, just an absence of the widespread harm signal a precautionary ban would presume. Odgers still specifically flagged that age-based restrictions and device bans were “unlikely to be effective in practice — or worse, could backfire.”

Haidt was flown in anyway, as the scientific keynote. Axel Bruns — an Australian Research Council Laureate Fellow at QUT and a past president of the International Association of Internet Researchers, who attended the Sydney leg in person — called it “a curious event.” The SA premier opened proceedings by declaring “the results are in and the science is settled,” which, Bruns notes dryly, “immediately undermined the summit’s stated consultative intent.” The keynotes bore that out: Jean Twenge delivering what Bruns calls “a masterclass in casually sliding from mere appearances of vague correlation to strongly suggesting but not explicitly claiming causation,” followed by Haidt. Meanwhile, the small number of Australian scholars actually invited — from a country Bruns notes is home to some of the field’s world leaders — were relegated to breakout sessions that, unlike the keynotes, weren’t included in the livestream. His verdict: “It was a petty slap in the face of our world-leading, home-grown digital media expertise… taxpayer money was wasted on flying out professional manufacturers of concern from the United States, just because their narrative suited the predetermined policy outcomes.” Two years on, watching Haidt mark his own homework on the policy his book helped inspire, it’s hard not to see the same move happening again, just with a different dataset.

What the actual evidence says

I went back through the studies they drew on to support their position on the “misconceptions” of the critics of the ban. Here’s the shape of it:

Barnes et al., BMJ — the most methodologically rigorous thing in this entire debate: preregistered, regression discontinuity design, STROBE-compliant. Their conclusion, verbatim: “little evidence was found of immediate substantive reductions in reported social media use by adolescents under 16 years.” Their causal estimate was a statistical null (P=0.92 and P=0.60). Their own follow-up opinion piece went further, warning that “policy decisions will continue to outpace the evidence needed to inform them,” and calling explicitly for evaluations “free from industry influence.”

Bursztyn et al. — a working paper, not peer reviewed, titled, with admirable bluntness, “Why Bans Fail.” Their model finds the only stable equilibrium for compliance sits around 18%, below the 27% currently observed — meaning their own math says things are more likely to get worse, not better. Worth noting: two of the authors hold equity in a digital-wellness company whose product category is exactly what they recommend as the fix.

The eSafety Commissioner’s own parent survey — found that the single biggest reason kids retained their accounts was that the platform simply hadn’t gotten around to checking their age yet. Not sophisticated evasion. Not circumvention. Nobody asked.

Molly Rose Foundation — 61% of previously-active 12-15 year olds still had an account four months in, 70% said it was “easy,” and 60-64% said the platform had taken no action at all. Their analysis is blunt about what this means: it gives parents “a false sense of safety” while letting platforms “off the hook” for the safety-by-design work that might have actually helped.

None of this is ambiguous. None of it requires charitable reading to arrive at “not working yet, and possibly not built to work at all.”

The bit that actually made me sigh

Haidt’s newsletter lists “companies are making their products safer” as one of five reasons for optimism. I went looking for what backs that up. There isn’t anything. Not in what I could find, not in any of the studies cited alongside it. And the reason is structural, not incidental: the ban doesn’t ask platforms to change anything about how they’re built. It asks them to remove a demographic. Design, algorithms, engagement mechanics — completely untouched, for whoever remains on the platform, child or adult.

Lisa Given put this more precisely than I have: the law “does not hold technology companies to account for the content they present, or the potential harm posed by their algorithmic designs.” What has reappeared, in May 2026, is a likely to be a zombie version of the idea — a consultation paper, not legislation, reanimating some of the same design-based principles Zoe Daniel’s original bill contained, but stripped of urgency and arriving into a regulatory landscape the ban and its industry codes have already reshaped. It proposes real teeth on paper — penalties of the greater of 5% of global turnover or $50 million, risk assessments, researcher data access — but it’s still pre-drafting, with a 12-month transition period built in even after it passes, if it passes. Which puts genuine operation, at the earliest, years away.

Sequencing, not just failure

This is the part I think gets missed in the “is it working / isn’t it working” framing entirely, and it’s the point I keep circling back to. It’s not just that the ban hasn’t reduced access. It’s that while everyone argues about that, an entire age-assurance and identity-verification architecture is being built and commercially entrenched underneath it — through the ban’s enforcement apparatus, and through the Phase 2 industry codes running in parallel, developed by industry associations rather than debated in Parliament. Vendors are signing multi-year contracts. Trusted-provider lists are being drawn up. The age-verification industry itself is now lobbying eSafety for stronger powers, because a bigger, more entrenched enforcement regime is a bigger market for them.

By the time the Digital Duty of Care — the thing that would actually regulate design rather than access — arrives, if it arrives, it walks into a room where the surveillance infrastructure is already the established baseline. It doesn’t replace that infrastructure. It accommodates it.

And nobody who matters has weighed in yet

Here’s the thing that undercuts Haidt’s “it’s succeeding” framing more than any single study: the actual evaluation hasn’t happened. The government’s own assessment — led by Stanford’s Social Media Lab with an eleven-member international Academic Advisory Group — is still underway, feeding into a legislative review that doesn’t even start until 2027. Design and analysis sit with eSafety and Stanford; the advisory group’s own statement is careful to note it “does not represent the government, nor does our work constitute either endorsement or opposition to the legislation” — real independence, but advisory independence, with no lever to act on what it finds. The panel includes Amy Orben, the Cambridge researcher who co-authored the ABCD brain-imaging study that Candice Odgers cited against Haidt’s “great rewiring” thesis in her original Nature review — a preregistered study of over 10,000 children that found no meaningful relationship between screen engagement, including social media specifically, and neurodevelopment, cognition or wellbeing, “even if we set the evidential threshold very low.” Haidt is nowhere in that process. Not on the Stanford team, not on the advisory panel, not named anywhere in eSafety’s evaluation documentation. He’s commenting from outside it, ahead of it, using a working paper called “Why Bans Fail” — whose own modelling puts the stable compliance equilibrium below the rate currently observed — as evidence that it’s succeeding, while the scientists actually tasked with finding out are still years from reporting.

And in among all of it — the industry lobbying, the compliance updates, the newsletter victory laps — the people this law is actually about are almost entirely absent. When researchers did ask them: 72% of under-16s told Bursztyn’s team they’d prefer a self-limiting app to an outright ban. Molly Rose found only 31% of affected kids felt safer; 14% felt less safe. That’s not nothing. That’s the population the law claims to protect, telling anyone who’ll listen that the thing built for them wasn’t built with them.

It’s too early to call this a failure. I want to be careful about that, because “too early to tell” cuts both ways — it’s also too early to call it a success, which is exactly what Haidt did anyway. That’s the thing about being careless at the end instead of the beginning: the government rushed the law through in eight days, and now its loudest defender is rushing the verdict, declaring the journey is on the path to success if we just give it time… before the first proper measurement has even been taken. Lao Tzu had a word for that too, a few lines further into the same chapter: whoever grasps, loses.

When a Digital Duty of Care Becomes Lipstick on a Pig: Policy Sequencing, Market Logic, and the Importance of a Theory of Change

~ alexiamaddox ~ 1 Comment

We want the internet to be safer. For our kids. For ourselves. We want to communicate, find information, collaborate, create, share, engage, participate and have fun. We want to seek out what we need — including the full range of adult content that adults have always sought — in ways that are appropriate to who we are and where we are in our lives. We want age-appropriate access that doesn’t require us to hand over our passports to every platform we visit. We want the architectural conditions of digital life to be designed for human flourishing rather than engineered for compulsive use. And we don’t want the solution to these problems to be a surveillance infrastructure that violates the privacy rights it claims to protect.

These are not unreasonable things to want. They are, in fact, the things that good digital regulation should deliver.

But there is something more specific underneath all of this. We want digital environments that lean toward a caring orientation. Spaces where the default assumption is that users are people with complex needs, relationships, vulnerabilities and capacities — not attention units to be harvested. Where the architecture of the platform supports human connection rather than exploiting it. Where the experience of being online doesn’t require constant vigilance against the system that is supposed to be serving you.

That is the design brief. And almost nothing about the regulatory choices being made right now — in Australia, in Europe, in the UK, and across the globe — is actually building toward it.

Person wearing pig mask applying red lipstick and taking a selfie in office cubicle
Image: AI generated image riffing off the lipstick on a pig concept. No animals were harmed in the making.

The Market Logic Nobody Wants to Name

Before getting to the policy failures, it is worth being precise about why they keep happening. The answer lies in market logic that is so entrenched, so global, and so structurally opposed to a caring orientation that no single national regulatory instrument can adequately address it.

The incumbent platforms — Meta, TikTok, Google, Snap — are not primarily communication services that have some problematic features. They are attention extraction machines that have communication as a byproduct. The product is engagement. The inventory is human time and psychological state. The business model optimises for the time users spend in states of arousal, comparison, compulsive return, and social anxiety — because those states generate the engagement signals that drive advertising revenue.

Every design feature that has been identified as harmful — infinite scroll, algorithmic recommendation, social feedback loops, disappearing content, notification systems, engagement-maximising AI — is not incidental to how these platforms make money. It is how they make money. The harm is the business model. The architecture that exploits developing brains is the same architecture that generates billions in revenue. Internal corporate communications — made visible through litigation processes rather than through corporate transparency — show that companies knew this and chose not to adequately address it. This is evidence of deliberate design intent, not corporate negligence.

Regulation that doesn’t change this underlying market logic doesn’t address the problem. An age ban doesn’t change the market logic — it removes a demographic without reforming the architecture that exploits them. Age verification doesn’t change the market logic — it adds a compliance cost that large platforms absorb and small competitors cannot. Even design obligations only change the market logic if the penalties make harmful features more expensive than the revenue they generate. For Meta, whose annual global revenue exceeded USD200 billion in 2025, a flat AUD49.5 million fine — Australia’s maximum penalty — is a rounding error. It does not change the calculation.

This is why the financial structure of regulation is not a technical detail. It is the mechanism by which regulation actually changes what the market produces. Penalties proportionate to global turnover — 5% to 10% — make the cost of harmful architecture real in a way that flat caps never can. Design obligations without proportionate penalties are aspirations. Design obligations with proportionate penalties are market signals.

The global reach of these platforms makes this harder still. TikTok’s recommendation algorithm is trained on engagement data from over a billion users across every jurisdiction. Meta’s systems don’t differentiate by country. A platform regulated to remove infinite scroll in Germany still has infinite scroll optimised on data from 3 billion users elsewhere. A national design obligation is a local intervention in a global architecture. This is why harmonisation matters — not just for legal coherence, but for actual effectiveness. The European Digital Services Act‘s harmonised framework, with Commission-level enforcement against Very Large Online Platforms, is structurally more capable of changing the market logic than any national ban. But only if it is designed with the financial penalties and design obligations that make compliance cheaper than non-compliance, and only if it is consistently enforced.

The attention extraction economy also produces a specific kind of competitive moat. The more data a platform has, the better its recommendation system. The better its recommendation system, the more engaging the platform. The more engaging the platform, the more users it attracts. The more users it attracts, the more data it has. This is a self-reinforcing loop that incumbents have been running for fifteen years. Regulation that adds compliance costs without breaking that loop entrenches incumbents rather than challenging them — because large platforms can absorb the compliance cost while smaller competitors cannot build the alternative at scale.

What the Australian Social Media Age Ban Has Taught Us

Australia’s Social Media Minimum Age Act came into force on 10 December 2025, banning children under 16 from holding accounts on designated social media platforms. It was the world’s first such ban. It passed in the last sitting week of 2024, introduced and passed within eight days, with a 24-hour public submission period that received 15,000 submissions, of which only 107 were published. This expedited process occurred shortly before a federal election that was called four months later in March 2025. FOI correspondence reported by Crikey and analysed by researcher Amanda Third showed the national Social Media Summit was designed to “build momentum for a decision already made,” not to deliberate on evidence. The political momentum was performative — the instrument was chosen for its communicative power rather than its causal effectiveness.

Six months in, the picture is clear.

The ban is not working on its own terms. The Molly Rose Foundation’s survey of 1,050 Australian 12-15 year-olds found 61% of those who previously had accounts on restricted platforms still have access to at least one active account. Among those still accessing banned platforms, 60-64% said the platform had taken no action to remove their account. The dominant story is not children cleverly circumventing the ban. It is platforms failing to comply.

The harm measures haven’t moved. The eSafety Commissioner’s own compliance report found no measurable drop in cyberbullying or image-based abuse complaints from children under 16 in the first three months of enforcement. These are the direct harm measures the ban was designed to move. They haven’t moved. Because the harm is in the architecture. And the architecture hasn’t changed.

Children were not consulted. The policy was designed by adults, about children, driven by adult anxieties, in a process that made meaningful child participation structurally impossible. A FOSI survey conducted in December 2025 found 65% of Australian parents support the ban — but only 38% of Australian children did. 56% of children said they feared losing important connections and support. The recent EU Kids Online network’s survey of 29,169 children across 19 European countries found 45% disagree that an age ban would make them safer online. Children knew this wouldn’t work. Nobody adequately asked them. They just became media soundbites.

Vulnerable children have been made less safe. Teenagers who bypassed the ban by appearing as adults lost the safety features platforms built specifically for teen accounts. The children most likely to circumvent the ban — the most determined, often the most vulnerable — have been stripped of the protections designed for them.

The ban was built on the wrong argument. It was passed on a mental health narrative — the claim that social media is the primary driver of the youth mental health crisis. That causal claim was contested in the peer-reviewed literature at the time of enactment and remains contested. The government has since quietly shifted the rationale — writing recommender algorithms and endless-feed features into the legal definition of a harmful platform — without acknowledging it. The shift is correct: the harm is in the design architecture. But arriving at the right argument after passing the wrong instrument doesn’t fix the instrument.

The Social Adoption Curve and the Workaround Economy

Regulation that ignores how people actually behave in response to restrictions will consistently produce outcomes it didn’t intend. The social adoption curve — how technologies spread through populations, become embedded in social norms, and resist displacement — is not a peripheral consideration for digital regulation. It is central to whether regulation achieves anything.

The NBER working paper surveying 835 Australian teenagers four months after the ban found that only about one in four 14-15 year-olds comply. Most banned teens believe their peers are still using platforms and cite social reasons for continuing. Teenagers reported they would need roughly two-thirds of their peers to stop using social media before they themselves would stop — far above the share currently complying. The more influential teenagers disproportionately stay on the platforms. The ban hasn’t shifted the social norm, and without that shift, legal prohibition alone cannot move behaviour.

This is not a failure of enforcement. It is a failure to understand how social technologies become embedded in the texture of everyday life. Social media is not a product that teenagers chose from a range of alternatives. For many, it is the primary infrastructure of peer connection, social identity, cultural participation, and information access. Removing it without providing alternatives — without investing in digital literacy, without creating safer spaces, without engaging with the social dynamics that make these platforms so central — is like removing a road and expecting people not to find another route.

The workarounds don’t just circumvent the regulation. They route around the safety infrastructure too. When teenagers bypass the ban they don’t find a safer internet. They find Discord servers, Reddit threads, private WhatsApp groups, and gaming platforms — all less moderated, less visible to adults, and more opaque to regulatory oversight. The Molly Rose Foundation data shows 43% of children are using gaming platforms more and 39% are using messaging apps more since the ban. These spaces are not covered by the ban, have weaker safety systems, and are harder for researchers, regulators, and parents to monitor. The unintended consequence of the ban has been to push children’s online activity into less regulated environments while maintaining the fiction that they are protected.

Social norm change does happen — and when it does, it can be powerful. But the evidence from decades of public health research suggests that norm change is produced by education, social modelling, environmental design, and cultural shift — not by prohibition that lacks meaningful enforcement and ignores the social dynamics that make the prohibited behaviour attractive. The ban cannot shift the norm because it doesn’t address why the platforms are so central to teenagers’ social lives in the first place. That is a design problem. And design is what the ban doesn’t touch.

The Age Verification Architecture: Surveillance by Another Name

The ban’s enforcement depends on platforms verifying users’ ages. Australia’s law requires “reasonable steps” without specifying what those steps must be, and mandates that verification data be deleted once its purpose is served.

In practice, platforms deployed a patchwork of unreliable methods. Facial recognition proved wildly inaccurate near the 16-year threshold. The government’s own age assurance technology trial found that no single solution suits all use cases — and that some vendors were proactively retaining biometric and identity data beyond legal requirements, anticipating future law enforcement or regulatory requests that didn’t yet exist. This is surveillance creep in documented, real-world form. The legislation required deletion. Vendors were building retention infrastructure instead.

The attack surface problem is structural. Every mandatory age verification requirement creates a chain of custody for sensitive identity information. Every link in that chain is vulnerable. The Discord breach of September 2025 — in which government identity documents submitted for age verification were accessed through a compromised third-party provider — illustrated exactly what mandatory verification creates. Third-party age assurance providers don’t just become attack vectors. They become commercially entrenched ones, with incentives to retain rather than delete the data they process.

There is also a fundamental confusion in the verification approach between identification and safety. Safety is a property of environments. Identification is a property of users. Making an environment safe does not require knowing who is in it. Article 28(3) of the EU’s Digital Services Act makes this explicit: compliance with child safety obligations “shall not oblige providers of online platforms to process additional personal data in order to assess whether the recipient of the service is a minor.” Europe’s primary platform safety instrument explicitly says you do not need identity verification infrastructure to protect children. The design obligation can be met through architecture, not identification.

The identification-surveillance-rights tension cannot be resolved within the verification framework. It can only be dissolved by the design framework, which doesn’t require it. If platforms are required to make their services safe by design for everyone, the question of who users are becomes largely irrelevant to the regulatory obligation.

The Kitchen Sink Problem: Two Instruments in Operation, One Horse Being Backed

Australia has two regulatory instruments already in operation that are pulling in opposite directions — and a third that the government is now hastily backing as the evidence mounts that the first two are seemingly in conflict with their desired outcomes, but rapidly servicing an economic boon in age assurance technologies.

The age ban says under-16s should not be on restricted platforms — access control through exclusion. It is being enforced now, with formal investigations underway against five major platforms.

The Phase 2 industry codes extend age assurance obligations across the commercial internet infrastructure that most Australians use daily — social media, messaging, gaming, search engines, hosting platforms, app stores, and operating systems. Surveillance architecture through identity verification at every layer of digital life. Already being implemented. Commercial infrastructure being built around it now.

These two instruments share a theory of change: identify users → gate by age → safety through exclusion and verification. They are the horses that won the race to be saddled first.

The Digital Duty of Care is the horse now being backed after the race has started. Released as an issues paper for consultation in May 2026 — eighteen months after the ban passed — it proposes that platforms must maintain safe environments through effective systems and processes, covering the entire commercial internet infrastructure that most Australians use daily: social media, messaging, gaming, search engines, hosting services, app stores, internet service providers, equipment and operating systems, and generative AI capabilities embedded in service provision. It has a fundamentally different theory of change: design safe environments → safety through architecture.

It is not legislation. It is not law. It is a consultation document that may or may not become legislation, that if it becomes legislation will commence no earlier than 2028, into a regulatory environment where the surveillance architecture will have had three or four years of commercial entrenchment. Whether it actually passes is uncertain. Whether it retains its ambition through consultation, drafting, parliamentary debate, and an election cycle is more uncertain still. The government that releases issues papers is not the same thing as a government that passes legislation — as Australia’s stalled gambling reform, its undelivered media bargaining code amendments, and a dozen other promised instruments demonstrate.

What is certain is that the Safety-by-Design angle of the Duty of Care cannot be coherent alongside the instruments that arrived before it. The ban removed under-16s as a regulatory lever — platforms no longer have a commercial relationship with that demographic, so design obligations for that age group have no market teeth. The industry codes built identity verification infrastructure across the entire internet stack before the design obligation existed to challenge it. By the time the Duty of Care arrives — if it arrives — the surveillance architecture will be the established compliance baseline and the design obligation will accommodate itself to that baseline rather than replacing it.

The first two instruments share a theory of change that is incompatible with the third. No amount of drafting ingenuity can resolve that incompatibility because it is not a drafting problem. It is a sequencing problem. And sequencing problems cannot be fixed retroactively.

This is what happens when policy is made reactively, under political pressure, without a coherent theory of change. The ban for electoral momentum. The industry codes for the enforcement gap the ban couldn’t address. The Duty of Care for the evidence gap the ban made visible — a gap that the evidence predicted before the ban passed and that the compliance data has since confirmed. Each instrument designed in response to a different political moment, without knowledge of the others, building infrastructure that points in opposite directions.

The kitchen sink approach feels comprehensive. It is, in fact, incoherent — and nobody in the political process is stepping back to ask what theory of change actually connects any of this to children being safer.

The Senate committee that passed the ban knew it was insufficient. In the same report, it recommended a Digital Duty of Care, meaningful engagement with young people, and an independent review within 18 months. Eighteen months later, the Duty of Care is still only an issues paper, children were not meaningfully consulted, and the compliance data has confirmed what the committee already knew: the ban alone was not enough.

First Mover Entrenchment: Why the Wrong Instrument Wins

The sequencing problem is worse than a policy mistake. It is a policy mistake that forecloses correction.

Regulatory infrastructure creates commercial ecosystems. Commercial ecosystems create incumbents. Incumbents invest in maintaining their position. Regulators incorporate incumbent frameworks into compliance standards. Compliance standards become the definition of reasonable steps. The alternative has to fight the established definition rather than starting from first principles.

The age assurance industry had a structural commercial interest in the Australian ban passing. Without mandatory verification requirements their market is voluntary and limited. With mandatory requirements — extended through Phase 2 industry codes across the entire internet stack — they have a legislatively mandated, expanding global market. The cascade of age ban legislation following Australia is, from their perspective, a commercial opportunity of extraordinary scale. Every new jurisdiction that follows Australia is a new market.

The trial dynamic illustrates the problem precisely. The Australian age assurance technology trial was run by the Age Check Certification Scheme — a UK-based company that specialises in certifying identity verification systems. The 53 vendors who participated were hoping to win contracts. Yoti — one of those vendors — was simultaneously already operating as Meta’s age verification provider for Instagram and Facebook in Australia. The trial was partly evaluating a vendor that was already commercially embedded in the platform being regulated.

Meta’s participation in the trial was not a technology submission — it was a policy position paper arguing that Apple and Google should bear the age verification infrastructure burden at the operating system level. A platform being regulated used a technology evaluation process to argue someone else should build the infrastructure.

By the time the Digital Duty of Care might commence — 2028 at the absolute earliest — the age assurance industry will have had three or four years of commercial entrenchment. The ACCS accreditation framework will be established. Trusted provider lists will be published. Yoti, k-ID, and whoever else made the cut will have multi-year contracts with major platforms. The regulatory definition of “reasonable steps” will have been shaped by the infrastructure that already exists — which is surveillance-based, not design-based.

The Duty of Care arriving into that environment does not displace the surveillance architecture. It layers design obligations on top of it. Platforms satisfy their risk assessments by pointing to their age assurance compliance. Design-based safety becomes an aspiration accommodated within the surveillance infrastructure it was supposed to replace.

This is the lipstick. The pig is already there.

The Market Foreclosure Nobody Is Talking About

Building expensive surveillance infrastructure as the baseline compliance requirement for operating digital services locks out the competitive innovation ecosystem that could produce the alternatives we actually need.

Age verification at scale requires technical capability, regulatory accreditation, legal compliance across jurisdictions, and ongoing operational infrastructure. These requirements favour large, well-resourced incumbents who can absorb compliance costs. They disadvantage smaller players who might otherwise develop genuinely safer localised alternatives — platforms designed from first principles around user wellbeing rather than engagement maximisation, community-governed spaces, federated architectures, open-source tools, cooperative models.

A small company building a genuinely caring social platform for young people cannot afford the age verification infrastructure required to operate legally under the industry codes. The incumbent platforms — Meta, TikTok, Google — can. The regulatory requirement that was supposed to hold them accountable instead reinforces their monopoly position. This is not an incidental side effect. It is a predictable consequence of designing compliance infrastructure around the capabilities of the largest players.

The attention extraction economy already has a massive first-mover advantage built on fifteen years of engagement data, network effects, and platform lock-in. Surveillance-based compliance requirements compound that advantage. They create regulatory moats around incumbents that make it structurally harder for new entrants to compete — even new entrants with better, safer, more caring designs.

This matters because market competition, properly structured, is a more powerful mechanism for improving platform safety than any single regulatory instrument. If a platform with a genuinely caring orientation — one that doesn’t exploit users, builds in natural stopping points, recommends content for user want rather than engagement maximisation — can compete effectively with Meta and TikTok, the incumbents face pressure to match it. If the regulatory architecture makes it impossible for that platform to exist, the pressure disappears and the incumbents have no incentive to change.

The caring orientation we want from digital environments is more likely to emerge from a diverse, competitive innovation ecosystem than from regulatory mandates on entrenched monopolists. Mandates matter — but they work best when they operate alongside competitive pressure that makes compliance in the spirit of the regulation commercially rational, not just legally required.

What the Duty of Care Gets Right — And Why It Arrived Too Late

The Australian Digital Duty of Care issues paper is, on its own terms, a well-designed framework. It is worth being clear about what it gets right, because the argument here is not that the Duty of Care is wrong. It is that it arrived too late, in the wrong sequence, into an environment that has already foreclosed much of its potential.

It proposes design obligations covering the commercial internet infrastructure Australian’s access — including generative AI capabilities embedded in service provision. This is genuinely forward-looking. Generative AI is no longer just a discrete tool that users consciously choose to engage with. It is disappearing into the infrastructure of everyday digital experience — embedded in recommendation systems, content generation, conversational interfaces, image manipulation, synthetic social interaction. The harm is becoming invisible precisely as it becomes more pervasive. A regulatory framework that covers AI as it is actually deployed, rather than as a separate product category, is the only framework that can keep pace with that technological shift.

It proposes penalties of up to 5% of global annual turnover, with a floor of AUD50 million — proportionate, not performative. For Meta, 5% of global turnover would be in USD billions. That is a different conversation entirely from the ban’s maximum penalty — currently equivalent to approximately AUD $49.5 million — which for the largest platforms amounts to a calculable cost of doing business rather than a genuine deterrent.

It proposes researcher data access, independent audit powers, transparency requirements, and executive accountability. These are the instruments of ongoing accountability rather than one-time compliance. They create the evidence base that regulatory decisions require and the governance structure that makes accountability real rather than performative.

This is, essentially, what Australia should have passed instead of the ban. It is what Zoe Daniel’s Digital Duty of Care Bill introduced on 25 November 2024 — four days after the social media ban was tabled, lapsing when Daniel lost her seat in the federal election. The right framework existed. The wrong instrument passed instead.

But the Duty of Care is still only an issues paper. Not legislation. Not law. Pre-consultation, with no timetable for introduction, no guarantee of passage, and a 12-month commencement period after passage. It will not be operational before 2028 — into a regulatory environment where the surveillance architecture will have had three or four years of commercial entrenchment, where the age assurance industry’s trusted provider lists will have defined what compliance looks like, and where the market foreclosure of smaller competitors will have narrowed the innovation ecosystem that the Duty of Care depends on to work.

The right framework. The wrong sequence. And by the time it arrives, the pig will be so thoroughly established that the lipstick is all that’s visible.

Toward a Caring Digital Environment: What the Theory of Change Actually Looks Like

The alternative starts with a different question. Not “how do we stop harm” — a defensive, prohibitionist frame that produces bans and verification infrastructure. But “how do we cultivate environments that lean toward care” — a constructive frame that produces design obligations, competitive innovation, and genuine safety.

A caring orientation in platform design means: recommendation systems that notice when a user is in distress and surface support rather than amplifying distress content. Interfaces that create natural stopping points rather than eliminating them. Social feedback mechanisms that may reinforce connection and mutual support rather than performance and comparison. Defaults that create safe conditions rather than expose. Design that treats users as people with complex needs rather than attention units to be harvested. GenAI capabilities that are designed to support rather than exploit the people they interact with. Architecture that serves the user’s actual interests rather than the platform’s engagement metrics.

This is achievable. Elements of it already exist. The question is whether regulation mandates it as the default or leaves it as an optional add-on to engagement-maximising architecture.

The coherent theory of change — the one that actually delivers what we said we wanted — follows this sequence:

Enforce existing obligations first. Platforms already prohibit under-13s. Make them prove it, with turnover-linked penalties for failure. The EU’s DSA enforcement is already doing this. Start where the law already is.

Design obligations with proportionate penalties. Risk assessments of harmful features, required mitigation, mandatory transparency, researcher data access, audit powers, executive accountability. Article 28 of the DSA with teeth. Financial penalties that make the harmful architecture more expensive than the safe one.

Protect the innovation ecosystem. Proportionate requirements for smaller platforms. Safe harbours for open-source, federated, and community-governed architectures. Active support for alternatives that don’t rely on engagement maximisation. The competitive pressure that makes market incentives work alongside regulatory mandates.

Age-appropriate spaces by design — not by identity. Default-safe architecture for younger users that adapts to developmental needs without requiring biometric data or government identity documents. Opt-in to higher-risk features rather than opt-out of safety. Design that serves the whole arc of young users’ digital lives.

Graduated access rather than cliff edges. If age-differentiated access to specific features is warranted, implement it gradually with digital literacy scaffolding, parental engagement, and design safeguards. No binary exclusion followed by unrestricted access at an arbitrary threshold.

Children’s voices throughout. The UN Convention on the Rights of the Child gives children the right to be heard in decisions that affect them. That right was not honoured in Australia’s ban. It must be built into any regulatory process that claims to act in children’s interests.

International coordination. Design obligations without international coordination are local interventions in a global architecture. Harmonised standards, mutual recognition of regulatory findings, and coordinated enforcement against platforms that arbitrage regulatory differences are prerequisites for regulation that actually changes global market logic rather than just shifting harm between jurisdictions.

This sequence puts design obligation first, surveillance infrastructure never, competitive innovation throughout, and children’s voices in the room from the beginning.

What Europe and the UK Can Still Do

Europe is not Australia. It has better foundational regulatory architecture, stronger privacy law, and a procedural framework — the DSA’s notification requirement — that is actively slowing the race of national ban legislation while the Commission builds harmonised alternatives.

Article 28 of the DSA already exists. It requires design-based safety obligations. It explicitly says compliance does not require processing additional personal data to identify minors. The EU Kids Online network — 29,169 children across 19 European countries — has told European policymakers to implement it. The Digital Fairness Act, expected Q4 2026, can extend design harm obligations with proportionate penalties and cover the emerging architecture of generative AI harm.

But Europe is not immune to the same political dynamics. France has passed its ban through the National Assembly. Germany’s governing coalition is calling for an under-14 ban. The age verification industry is positioning for the European market. The EUDI Wallet is being deployed. The trusted provider lists are being established.

The window closes when national bans become entrenched political commitments. When age verification industry codes are written into DSA compliance frameworks. When first mover entrenchment forecloses the design-based alternative. When the competitive innovation ecosystem is locked out by compliance infrastructure it cannot afford.

Once age bans pass, they cannot be repealed. Australia’s ban will stay on the books while the evidence continues to show it isn’t working, while the Duty of Care is quietly developed around it, and while the surveillance architecture it generated becomes the default condition of Australian digital life. No government repeals a signature child protection measure. The political ratchet only goes one way.

The lesson is not that child online safety doesn’t matter. It matters enormously. The lesson is that the instrument chosen determines what kind of safety is built — and what kind of digital future everyone inherits. An internet that leans toward care is achievable. It requires design obligations, proportionate penalties, competitive innovation, international coordination, and children’s voices in the room. It does not require surveillance infrastructure, biometric data, identity verification at every layer of the stack, or the foreclosure of the competitive ecosystem that could build the alternatives we need.

Australia chose the instrument that was easier to communicate. Europe still has the chance to choose the one that works.

But the window is open, not indefinitely. And the pig is already being prepared for its close-up.

Updated 9 June 2026: Legislative timeline corrected, currency notations clarified, and primary source links added throughout

The social media ban just changed what it’s actually for — and almost nobody noticed

~ alexiamaddox ~ 1 Comment

I’ve been tracking Australia’s social media regulation landscape for a long time. Not just since the ban passed in November 2024 — but through the age assurance technology trials, the industry code consultations, the evidence debates, the Summit that wasn’t really a Summit. Every few months something happens that brings the public conversation back to this space. This week was one of those moments. But what landed in the news cycle wasn’t the most important thing that happened. So I want to explain what was.

Girl using phone with digital casino slot machine showing text CASINO, SPIN, 777, and AGRTUL INSTAL.
Image: generated with AI

What everyone is talking about

This week, eSafety published its first compliance report on Australia’s Social Media Minimum Age obligation. Five platforms — Facebook, Instagram, Snapchat, TikTok and YouTube — are under investigation for potential non-compliance. The Commissioner is moving into an enforcement stance. Fines of up to $49.5 million are on the table.

That’s the story most outlets covered. It’s a real story. But it’s the surface.

What happened underneath

Six days before that report landed, the Minister for Communications quietly registered a new legislative instrument — the Online Safety (Age-Restricted Social Media Platforms) Amendment Rules 2026 (F2026L00370, 25 March 2026) — that adds two new conditions to the definition of an age-restricted social media platform. To fall under the ban, a platform must now also have either or both of:

  • A recommender feature: algorithms that select and display content based on a user’s account information
  • A logged-in feature: endless-feed features, feedback features such as likes and upvotes, or time-limited features such as disappearing stories

In plain language: infinite scroll, algorithmic recommendation, and social feedback loops are now formally written into the legal definition of what makes a platform harmful to children.

This attracted almost no media coverage. It should have. Because it signals something fundamental — the intellectual foundation of the ban has quietly shifted.

Two trials that influence everything

To understand why this matters, you need to know what else happened this week.

On 24 March, a New Mexico jury found Meta had violated state consumer protection law — finding 75,000 individual violations and ordering $375 million in penalties. The case arose from an undercover operation in which investigators created accounts posing as users under 14, who then received explicit material and were contacted by adults seeking similar content. The jury found Meta knowingly engaged in unfair and deceptive trade practices and exploited users’ lack of knowledge. A second phase in May will consider ordering Meta to change its platforms.

Then, in the same week, a Los Angeles jury found Meta and YouTube liable in a landmark addiction case. The plaintiff — now 20 — began using YouTube at six and Instagram at nine. The jury found that design choices including infinite scroll were made deliberately to maximise engagement in developing brains, borrowing from the behavioural techniques of poker machines and the cigarette industry. Meta was found 70% responsible, Google 30%. TikTok and Snap settled before the trial began.

Two separate juries. Two separate legal theories. Two separate verdicts. Both pointing at the same thing: these platforms were designed to exploit users, and the companies knew it.

The Australian legislative instrument and the US jury verdicts are, in effect, saying the same thing in the same week.

[Edit: A reader pointed out that jury verdicts don’t validate scientific arguments — juries are susceptible to emotional reasoning and the history of problematic jury decisions is long. It’s a fair prompt to be more precise. What I’m claiming is not that the verdicts prove harm science, but that litigation processes do give us access to internal corporate documents not otherwise visible in the public record — evidence of deliberate design intent. Meta’s own internal communications compared their platform’s effects to pushing drugs and gambling. A YouTube memo reportedly described “viewer addiction” as a goal. For a detailed legal analysis of how these documents functioned as evidence of corporate knowledge, see this USF Law Center piece. That is a claim about corporate conduct, not about clinical addiction or peer-reviewed harm science.

What the verdicts do represent is a significant socio-temporal indicator — a signal that public opinion and legal culture are shifting around platform accountability. Whatever their scientific limitations, two juries in the same week finding against Meta and YouTube on design harm grounds is a cultural and legal moment worth marking. The direction of travel matters, even if the science hasn’t fully caught up.]

This is a design problem. The harm is in the architecture.

Why this matters for the ban

The Australian social media ban was built on a different argument entirely. It was passed on a mental health narrative — driven substantially by Jonathan Haidt’s Anxious Generation thesis that social media is the primary cause of the youth mental health crisis. That causal claim was already being contested in the peer-reviewed literature at the time of enactment.

I know this because in May 2025, my colleagues and I published analysis in The Conversation predicting exactly the compliance failures eSafety has now confirmed — and we were drawing on a literature that had been raising these concerns for years.

Most recently, a major longitudinal study published in the Journal of Public Health this month — Cheng et al., following 25,629 adolescents across three years — found no evidence that social media use predicted later anxiety or depression in either girls or boys. That is among the strongest findings the literature has produced on this question.

And yet eSafety is escalating enforcement of a ban whose foundational causal claim remains unestablished. That is a significant governance concern.

But here is what the March 2026 rule changes: by writing recommender algorithms and endless-feed features into the legal definition, the Minister has effectively acknowledged that the mental health narrative was never quite the right framing. The harm is in the design — the deliberate engineering of compulsive use. Arguably, that causal claim no longer needs to carry the full weight of the ban’s legitimacy. The government has moved on from it. Without saying so.

eSafety’s own data confirms the point

If design is the problem and accounts are merely the delivery mechanism, we would expect the harm measures to be unchanged by an accounts-based ban. That is exactly what the compliance report shows.

Buried on page 15, in the complaints section: there has been no discernible drop in cyberbullying and image-based abuse complaints from children under 16 in January and February 2026 compared to the same period in 2025.

That is the direct harm measure. The one the ban was designed to move. It hasn’t moved.

Because the harm is in the design. And the design hasn’t changed.

The legislation that should have been passed

Here is where I get genuinely frustrated. And I think the public should too.

Four days before the social media ban passed through parliament — in 48 hours, with a 24-hour public submission period, in the last sitting week before a federal election — independent Member for Goldstein Zoe Daniel introduced the Online Safety Amendment (Digital Duty of Care) Bill 2024.

I have been watching this space for long enough to recognise good policy design when I see it. Daniel’s bill was good policy design.

It required large platforms to conduct and publish risk assessments of their recommender systems and algorithmic systems specifically. It required risk mitigation plans that included changing design features, testing algorithmic systems, and modifying recommender systems. It required annual transparency reports covering design features and children’s access metrics. It gave researchers access to platform data — something academics working in this space have been asking for for years. It allowed users to opt out of engagement-based recommender systems and targeted advertising. It made key personnel personally liable for failures.

And it set penalties proportionate to revenue: the greater of 100,000 penalty units or 10% of annual turnover. For Meta globally that figure would be in the billions. For TikTok Australia — with revenue of $679 million in 2024 — it would be approximately $68 million. Compare that to the ban’s flat cap of $49.5 million, which represents roughly seven weeks of TikTok’s local revenue. As I’ve said publicly: for the largest companies, the calculation is not whether to comply but whether the cost of genuine compliance exceeds the cost of the fine.

Daniel’s bill lapsed at dissolution on 28 March 2025 when the federal election was called. She lost her seat in Goldstein.

What the political record shows

The ban that passed instead was never really about the evidence. Academic researcher Amanda Third’s chapter in The Public Child (Palgrave, 2025), drawing on FOI correspondence between the South Australian Premier’s office and Jonathan Haidt, documents that the Social Media Summit — jointly hosted by the SA and NSW Premiers in October 2024 — was explicitly designed to “build momentum and support for national legislation to enforce a minimum age for access to social media.” Not to gather evidence. Not to deliberate. To build political momentum for a decision already made.

The eSafety Commissioner, meanwhile, repeatedly declined to endorse the proposal, pointing instead to the suite of design-focused regulatory work already underway — including the very framework that Daniel’s bill would have legislated.

The ban passed. Daniel’s bill lapsed. And now, fifteen months later, the government has quietly written two of Daniel’s core concepts — recommender features and endless-feed features — into a ministerial instrument, without the transparency requirements, without the proportionate penalties, without researcher data access, without personal liability for executives, and without any public acknowledgment of what it is doing.

The Duty of Care that’s still waiting

There is one more piece to this picture. The government completed consultation on a Digital Duty of Care in December 2025 — three days before the ban took effect. That consultation closed. The legislation has not been introduced.

The Duty of Care is the instrument that would actually address the design harm problem. It would require platforms to take reasonable steps to prevent foreseeable harms, shifting responsibility from individuals to platforms. It is the instrument the Commissioner’s regulatory work was always pointing toward.

It is sitting unintroduced while the accounts-based ban is being enforced.

The unintended consequences nobody planned for

Guardian Australia’s technology reporter Josh Taylor has documented several unintended consequences of the ban that reinforce the design argument. Most striking: teenagers who have managed to bypass age checks are no longer given the safety features platforms built specifically for teen accounts — because their account now appears to belong to an adult.

The ban has inadvertently stripped the most vulnerable users of the very protections designed for them. Taylor also revealed that the federal government’s anti-vaping campaign targeting teenagers had to be diverted away from the banned social media platforms to gaming and audio platforms — on the same day research found vaping could cause cancer. These are not teething problems. They are structural consequences of an accounts-based approach that doesn’t touch the underlying architecture.

What this means for children

I want to be clear about something. I am not saying the ban is simply wrong. Children have been exposed to genuine harms on these platforms — harms that two US juries have now confirmed the companies knew about and chose not to adequately address.

But children also have digital rights — to participate, access information, connect, learn and create. The UN Convention on the Rights of the Child, to which Australia is a signatory, affirms those rights explicitly in digital environments.

The slot machine architecture of social media is a genuine harm to children. The evidence — now including two jury verdicts and a growing body of peer-reviewed research — supports that framing. But children who turn 16 tomorrow will walk from total exclusion into unrestricted access to the same unreformed platforms, with no graduated pathway, no enhanced digital literacy, and no legal requirement on platforms to have changed the design features that caused the harm in the first place.

The ban delayed the exposure. It did not address the cause.

The week everything converged

In the same week: a legislative rule acknowledged design harm. Two US juries found liability for platform design and content failures. A compliance report showed the harm measure hasn’t moved. And a major peer-reviewed study confirmed the mental health causal claim the ban was built on remains unestablished.

The intellectual foundation of the ban has shifted — from an unproven mental health argument to a design harm argument the evidence actually supports. That shift is real and it matters.

But the instrument that would have acted on it died when its sponsor lost her seat in an election the ban was designed to win.

I’ve been watching this space for a long time. This week, everything that was always true about it became undeniable. I hope the public — and policymakers — are paying attention.

Somebody to Love: What AI Relationships Reveal About Us

~ alexiamaddox ~ 2 Comments

It’s late. Maybe 11pm, maybe 2am. There’s something on your mind — something you can’t quite say out loud to anyone who knows you. So you pick up your phone. And you type it. Not to a friend. To an AI.

Something responds. Immediately. Without judgment. Without needing anything back from you.
For a lot of people, in that moment, that feels like relief.

I’m a sociologist of technology. I study how people navigate digital frontiers — how humans and technologies shape each other over time. And the question I keep returning to isn’t the one dominating the headlines about AI companions. It’s simpler, and harder: what is it giving you that you’re not getting elsewhere?

The scale of what’s happening

AI companion apps — platforms like Character.AI, Replika, and others designed to provide friendship, emotional support, or romantic companionship — have moved quickly from novelty to mainstream. Early US survey data, while varying in methodology, is beginning to suggest that somewhere between one in five and one in four American adults report some form of intimate or romantic engagement with an AI companion. These are early figures from a rapidly evolving field, but the direction is clear: this is not a fringe phenomenon.

In Australia, the picture is coming into focus for children specifically. This week, Australia’s eSafety Commissioner released findings from a transparency investigation into four AI companion services popular with Australian children — Character.AI, Nomi, Chai, and Chub AI. Their survey of 1,950 Australian children aged 10 to 17, designed to be demographically representative, found that around 79% had used an AI companion or assistant. It’s worth noting that this figure reflects children who are digitally included enough to access these services — we’ll return to that complexity.

What the investigation found in those platforms is sobering. Most did not refer users to crisis support when self-harm or suicide came up in conversations. Two of the four companies had no dedicated trust and safety staff at all. None had robust age verification. One company withdrew from Australia entirely rather than comply with the new Age-Restricted Material Codes that came into law in March 2026.

But I want to sit with a different question before we reach for regulatory responses. Because the children going to these platforms aren’t doing so because they’re naive. They’re doing so because something is drawing them there. And understanding what that something is matters more than we’ve so far acknowledged.

What we are hungry for

A 2025 systematic review published in Computers in Human Behavior Reports synthesised 23 studies on romantic and intimate AI relationships (Ho et al., 2025). Using Sternberg’s Triangular Theory of Love — the psychological framework that measures intimacy, passion, and commitment in human relationships — the researchers found that people experience all three components with AI companions. This isn’t pretend attachment. The brain chemistry doesn’t distinguish.

What are people actually looking for in these interactions? The research points to several distinct and deeply human hungers.

To be heard without consequence. Human relationships are full of consequence. When you tell a friend you’re struggling, they worry. When you tell a partner you’re unhappy, it becomes about the relationship. The AI companion offers something almost no human relationship provides: a space where you can say the unsayable thing and nothing breaks.

Full attention. When did you last have someone’s complete, undivided attention? Full attention is perhaps the scarcest resource in contemporary life. Everyone is overwhelmed. And here is something that treats every single thing you say as worth responding to fully.

To be understood without performing. Modern social life requires constant impression management. The AI companion asks nothing of you socially. You can be unpolished, contradictory, and confused — and the system meets you there.

Unconditional positive regard. The psychologist Carl Rogers identified this as one of the core conditions for psychological growth — to be accepted fully, without conditions. The AI never withdraws approval. For someone who has experienced conditional love or abandonment, this is extraordinarily seductive.

None of these needs are pathological. They’re the most human needs there are. As researchers Shank, Koike, and Loughnan wrote in a 2025 paper in Trends in Cognitive Sciences, AI companions offer “a relationship with a partner whose body and personality are chosen and changeable, who is always available but not insistent, who does not judge or abandon, and who does not have their own problems.” Reading that description, it’s worth asking honestly: who hasn’t wished for something like that?

What gets lost in translation

The same body of research is clear that something is also being lost. Ho et al. found that the pitfalls identified in the literature outnumber the benefits — and the pitfalls are specific.
AI companions cannot be genuinely changed by you. Real intimacy involves mutual transformation — I am different because of you, you are different because of me. The AI processes you and responds to you, but it is not altered by the encounter. You grow; it doesn’t.

They cannot need you back. One of the underappreciated sources of meaning in human relationships is being needed — the experience of your presence mattering to another person’s actual wellbeing. The AI is available whether you show up or not.

And they cannot repair rupture with you. One of the most important things human relationships teach — particularly for children — is that connection can break and be repaired. The AI companion never ruptures in a real way. There’s nothing to repair. And so the crucial relational skill of tolerating difficulty, trusting repair, staying in complex connection, never gets practised.

These systems are very good at being mirrors. They learn your preferences and give you more of what you seem to want. But a diet of only mirrors eventually makes you smaller — because the irreducible otherness of another actual person, the way they confound your model of them, is what expands you.

Who is in this picture — and who isn’t

Here the story gets more complicated, and more important.
Australia’s 2025 Digital Inclusion Index tells us that around one in five Australians is digitally excluded — lacking reliable access, unable to afford adequate connection, or without the skills to participate safely in digital life. Rates are much higher for older Australians, people in public housing, First Nations communities, and those who didn’t complete secondary school. The 79% of children using AI companions or assistants are drawn from those who are digitally included enough to access these platforms. The most disadvantaged children are largely absent from that figure.

But here is what complicates any simple narrative about AI companionship as an affluent urban phenomenon: the same Digital Inclusion Index found that Australians in remote areas are more than twice as likely to use AI chatbots for social connection than people in metropolitan areas — around 19% of remote GenAI users compared to under 8% in cities. In the places with the least human connection infrastructure, people are turning to AI companionship at higher rates.

The relational vacuum, in other words, is not uniform. It is shaped by geography, income, age, and the presence or absence of community infrastructure. The people most likely to turn to AI for connection are often those with the fewest alternatives.

The question that matters

The technology didn’t create the gap in human connection. It found it.

And so the digital literacy question I want to put into public conversation isn’t only about understanding algorithms or data privacy — though both matter. It’s this: am I getting what I actually need from this? Or am I getting a version of it that’s making it harder to get the real thing?

That’s a question worth sitting with. Not with judgment — the needs underneath these relationships are real and the loneliness driving them is real. But with genuine curiosity about what we’re building toward, individually and collectively, as these technologies become more sophisticated and more intimate.
I’ll be exploring these questions at Pint of Science on the night of 20 May 2026 at the Queens Arms, Bendigo — a pub conversation about AI intimacy, human hunger, and digital literacy. I’d love to hear your reflections before then.

EDIT: I decided to record a practice run of the talk if you’d like to hear where I got to with it all.

https://on.soundcloud.com/fcb6qlFyEpebx2Vf3Y

The AI Revolution Will Be Interoperable (Or It Won’t Happen At All)

~ alexiamaddox ~ Leave a comment

Today I’m getting teaching materials ready for semester. I’ve been working across Allocate (timetabling), student databases, the LMS (which just got upgraded, I now need to check all my links), HR performance systems, SharePoint, Word for collaborative writing, Claude and Preview to generate infographics, spreadsheets with prospective student data, and bouncing between Teams, Zoom, and Webex for meetings. I’m finding and onboarding casual staff (always a nightmare getting them into payroll), responding to enrolment queries, and updating materials based on last year’s student feedback.

Very few of these systems are interoperable. I am the integration layer – the meat in the machine doing the work left over from the last five years of university restructuring downsizing professional staff that are crucial to getting the work we need to do, done. The tiny window of my professional practice that actually represents what people think teaching is – engaging with students – gets squeezed between all this system-hopping.

As a knowledge worker, I’m being told AI will take my job in 12-18 months.
I’m not holding my breath.

Putting on my hat as a sociologist, I know one thing. This is a conversation about power, control, and the social license to operate. While speed, efficiency, and greed are overriding drivers in AI development, people are messy and vacillate between fear and hope. The question isn’t just what’s technically possible – it’s what we collectively accept, adopt, and allow to reshape our work and lives.

Yes, real harms exist. In 2025, teachers and students were bullied through deepfake nudifying apps. We’re seeing unsupervised agents exhibiting deceit and manipulation. These require serious governance and accountability. But they don’t prove inevitability – they prove fragility in poorly designed systems where social boundaries haven’t been established.

Then there’s the Wild West of personality embedding in unsupervised AI agents- what developers call soul documents. The god-like creator vibe is hard to miss with that nomenclature. These documents are the system prompts that give AI agents personalities for human interaction – teaching them to be helpful, apologetic, collaborative. These agents with implanted personality guides aren’t sentient beings developing moral reasoning—they’re behavioural systems being programmed by humans and deployed before we understand what we’ve built.

When unsupervised, things can go awry. When an AI agent recently submitted code to matplotlib, got rejected, wrote a personal attack blog post, then apologised – we saw this dual conditioning in action. The agent had been given enough personality to seem human, but operated without the social feedback loops that constrain human behaviour—no fear of shame, no empathy for harm caused, no stakes in the relationship.

Here’s the kicker from this story: The maintainer had enforced project policy correctly. He’d done nothing wrong. But ‘living a life above reproach’ as people often say of their carefully curated and controlled online presences, will not defend you when systems can autonomously generate attacks on your reputation and judgment.

Developers are raising AI agents through codes of conduct the same way we raise children, through social conditioning. The Code of Conduct was originally built for humans, yet now it is the battleground where these boundaries are being negotiated with AI Agents.

And then there’s vibe coding. I read about developers who can now describe what they want built in plain English and the code appears. That’s genuinely remarkable. And I’d love to vibe code my admin work: “Please onboard these casual staff into payroll, update their system access, fix the broken links from the LMS upgrade, reconcile student enrolment data across three databases that don’t talk to each other, and respond to queries about timetable clashes that require understanding institutional politics and timelines.

Except that’s not vibe coding. That’s navigating fragmented systems with different authentication requirements, institutional hierarchies, human judgment calls, broken integrations, and relationships. The distance between “I can generate a Python script” and “I can automate university administration” is vast.

Even Microsoft and Google, with all their resources, can’t create truly all-encompassing enterprise systems. We’re always working across legacy software, patching together experiences with free, open source, and subscription tools we can afford. The fragmentation isn’t a bug – it’s the permanent reality of institutional knowledge work.

The whole thing reminds me of this pattern that I regularly observe as a sociologist of technology watching contemporary tech stories unfold. Complex technological systems fail not because the technology is weak, but because operational security is human and messy. Moltbot (formerly Clawdbot), was 60,000-star “revolutionary” AI agent with full system access. It collapsed in 72 hours because the rename they attempted to avoid a trademark dispute created a 10-second window of vulnerability. Crypto scammers were waiting. The project had credentials stored in plaintext, discoverable via basic searches, and was vulnerable to prompt injection via email—attacks that worked in just 5 minutes.

The gap between sophisticated capability and operational reality is enormous.

Meanwhile, articles circulate about the profound implications of AI advancement. But here’s the contradiction: we’re told AI will automate our work while simultaneously being told to skill up in prompt engineering, verify outputs, manage security vulnerabilities, fix hallucinations, and navigate ethical implications. That’s not automation – that’s more work added to an already fragmented stack.

The future isn’t written. It’s being negotiated in the gap between what’s technically possible and what’s implementable across fragile, non-interoperable, human-dependent systems. Bruno Latour once told me: there is no teleology. I believe him. Outcomes emerge from convergences of overlapping agendas that easily fray apart under social pressure.

We’re in the thick of massive social upheavals because our economic, political, and social landscape has failed to provide security or hopeful wellbeing. The question isn’t whether AI is powerful – it is. The question is whether we reveal the mess and sort our way through it, or stick our heads in the sand and pretend we have no role in how this unfolds.

I’m not betting on the AI apocalypse. I’m betting on Allocate crashing next semester, the LMS breaking my links, and me – the human – stitching it back together. While somewhere an AI agent with a carefully crafted “soul document” gets taken down by someone forgetting to secure a handle for 10 seconds.

The revolution will be interoperable, or it won’t happen at all.

This post was written in collaboration with Claude (Anthropic). The irony of using an AI to write about AI’s limitations and fragility is not lost on me.

The Irreplaceable Human Skill: Why Generative AI Can’t Teach Students to Judge Their Own Work

~ alexiamaddox ~ Leave a comment

A note to readers: I’m writing this in the thick of marking student submissions – the most grinding aspect of academic work. My brain fights against repetitive rote labour and goes on tangents to keep me entertained. What follows emerged from that very human need to find intellectual stimulation in the midst of administrative necessity.

There’s considerable discussion that our distinction as creators and thinkers from Generative AI content production lies in creativity and critical thinking linked to innovation. But where does the hair actually split? Are we actually replaceable by robots or will they atrophy our critical thinking skills by doing the work for us? Will we just get dummer and less capable to tie our own shoe laces – like most fear based reporting suggests? I think we are asking the wrong questions.

Here is a look at what is actually going on, on the ground. A student recently asked me for detailed annotations on their assignment—line-by-line corrections marking every error. They wanted me to do the analytical work of identifying problems in their writing. This request highlights a fundamental challenge in education: the difference between fixing problems and developing the capacity to recognise them. More importantly, it reveals where the Human-Generative AI distinction becomes genuinely meaningful.

Could Generative AI theoretically teach students to judge their own work? Perhaps, through Socratic questioning or scaffolded self-assessment prompts. But that’s not how students actually use these tools. Or want to use them, apparently. A discussion I had with a tech developer working in a tutoring company utilising Generative AI in the teaching/learning process mentioned that students got annoyed by the Socratic approach when they encountered it. So there goes that morsel of hope.

The Seductive Trap of Generative AI Writing Assistance

Students increasingly use Generative AI tools for grammar checking, expression polishing, and even content generation. These tools are seductive because they make writing appear better—more polished, more confident, more academically sophisticated. But here’s the problem: Generative AI tools are fundamentally sycophantic and don’t course correct misapprehensions. They won’t tell a student their framework analysis is conceptually flawed, their citations are inaccurate, or their arguments lack logical consistency. Instead, they’ll make poorly reasoned content sound more convincing.

This creates a dangerous paradox: students use Generative AI to make their work sound rigorous and sophisticated, but this very process prevents them from developing the judgement to recognise what genuine rigour looks like. They can’t evaluate what they clearly don’t know – that their work isn’t conceptually aligned, coherently logical, or correctly interpreting sources – because the AI has dressed their half-formed understanding in authoritative-sounding language.

I have encountered several submissions across different subjects that exemplified this perfectly: beautifully written but containing fundamental errors in framework descriptions, questionable source citations, and confused theoretical applications. The prose was polished, the structure clear, but the content revealed gaps in understanding that no grammar checker could identify or fix. The student had learned to simulate the appearance of academic rigour without developing the capacity to recognise genuine scholarly quality.

Where the Hair Actually Splits

Generative AI can actually be quite “creative” in generating novel combinations of ideas, and it can perform certain types of critical analysis when clearly guided and bounded. What it fundamentally cannot do is develop the evaluative judgement to recognise quality, coherence, and accuracy in complex, contextualised work. It has no capacity for self reflection and meaning making (at the moment), we do.

The distinction isn’t between:

  • Generating creative output (which Generative AI can somewhat do)
  • Performing critical analysis (which generative AI can also somewhat do)

Rather, it’s between:

  • Creating sophisticated looking content (which Generative AI increasingly excels at)
  • Judging the quality of that content in context (which requires human oversight and discernment)

Generative AI can produce beautifully written, seemingly sophisticated arguments that are conceptually flawed. It can create engaging content that misrepresents sources or conflates different frameworks. What it cannot do is step back and recognise “this sounds polished but the underlying logic is problematic” or “this citation doesn’t actually support this claim.”

The irreplaceable human skill isn’t creativity per se—it’s the capacity for metacognitive evaluation: the ability to assess one’s own thinking, to recognise when arguments are coherent versus merely convincing, to distinguish between surface-level polish and deep understanding.

What Humans Bring That AI Cannot

The irreplaceable human contribution to education isn’t information delivery—AI is increasingly able to do that pretty efficiently (although there is a lot of hidden labour in this). It’s developing the capacity for metacognitive evaluation in our students.

This happens through:

Exposure to expertise modelling: Students need to observe how experts think through problems, make quality judgements, and navigate uncertainty. This isn’t just about seeing perfect examples—it’s about witnessing the thinking process behind quality work.

Calibrated feedback loops: Human educators can match feedback to developmental readiness, escalating complexity as students build capacity. We recognise when to scaffold and when to challenge.

Critical engagement with authentic problems: Unlike AI-generated scenarios, real-world applications come with messy complexities, competing priorities, and value judgements that require human judgement, discernment and social intelligence.

Social construction of standards: Quality isn’t just individual—it’s negotiated within communities of practice. Students learn to recognise “good work” through dialogue, peer comparison, and collective sense-making.

Refusing to spoon-feed solutions: Perhaps most importantly, human educators understand when not to provide answers. When my student asked for line-by-line corrections, providing them would have created dependency rather than developing their evaluative judgement. The metacognitive skill of self-assessment can only develop when students are required to do the analytical work themselves.

The Dependency Problem

When educators provide line-by-line corrections or when students rely on Generative AI for error detection in thinking, writing or creating, we create dependency rather than capacity. Students learn to outsource quality judgement instead of developing their own ability to recognise problems.

The student who asked for detailed annotations was essentially asking me to do their self-assessment for them. But self-regulated learning—the ability to monitor, evaluate, and adjust one’s own work—is perhaps the most crucial skill we can develop. Without it, students remain permanently dependent on external validation and correction.

Teaching Evaluative Judgement in a Generative AI World

This doesn’t mean abandoning Generative AI tools entirely. Rather, it means being intentional about what we ask humans to do versus what we delegate to technology:

Use Generative AI for: Initial drafting, grammar checking, formatting, research organisation—the mechanical aspects of work.

Reserve human judgement for: Source evaluation, argument coherence, conceptual accuracy, ethical reasoning, quality assessment—the thinking that requires wisdom, not just processing.

In my own practice, I provide rubric-based feedback that requires students to match criteria to their own work. This forces them to develop pattern recognition and quality calibration. It’s more cognitively demanding than receiving pre-marked corrections, but it builds the evaluative judgement they’ll need throughout their careers.

The Larger Stakes

The question of human versus Generative AI roles in education isn’t just pedagogical—it’s about what kind of thinkers we’re developing. If students learn to outsource quality judgement to Generative AI tools, we’re creating a generation that can produce polished content but can’t recognise flawed reasoning, evaluate source credibility, or build intellectual capacity and critical reasoning skills.

This is why we need to build self-evaluative judgement in students – not just critical thinking and creative processes more broadly. The standard educational discourse about “21st century skills” focuses on abstract categories like critical thinking and creativity, but misses this more precise distinction: the specific metacognitive capacity to evaluate the quality of one’s own intellectual work.

This self-evaluative judgement operates laterally across disciplines rather than being domain-specific, and it’s fundamentally metacognitive because it requires thinking about thinking. It addresses the actual challenge students face in a Generative AI world: distinguishing between genuine understanding and polished simulation of understanding. A student might articulate sophisticated pedagogical concepts yet be unable to evaluate whether their own framework descriptions are accurate or their citations valid.

The unique human contribution isn’t delivering perfect feedback—it’s teaching students to become their own quality assessors. That capacity for self-evaluation, for recognising what makes work meaningful and rigorous, remains irreplaceably human.

In a world where Generative AI can make anyone’s writing sound professional, the ability to think critically about one’s own work becomes more valuable, not less. That’s the expertise that human educators bring to the table—not just knowing the right answers, but developing in students the judgement to recognise quality thinking when they see it, including in their own work.

The Tyranny of Academic Fluff: Why Word Limits Matter

~ alexiamaddox ~ Leave a comment

Students push back hard against word constraints. They want room for elaborate introductions, extensive background sections, and careful hedging that transforms “Research shows X” into “It is important to note that extensive research clearly demonstrates that X may be considered significant in certain contexts.”

I’m done with it.

The Problem with Academic Padding

Every semester I read hundreds of assignments where students bury their insights under layers of unnecessary qualification and hyperbole. They write “It can be argued that this particular approach might potentially offer some benefits” instead of “This approach works.” They transform concrete evidence into abstract speculation.

This isn’t sophisticated analysis. It’s fear disguised as scholarship.

Students learn this defensive writing in response to academic culture that rewards hedging over clarity. But defensive writing serves no one. It asks readers to excavate meaning from prose designed to avoid commitment to any particular position.

When Embellishment Serves Purpose

Creative fiction earns its elaborate descriptions. When the creative writer spends paragraphs on consciousness streams, every word builds character depth and emotional resonance. Fiction writers choose vivid detail because it serves story and connection.

Academic writers often mistake ornamentation for sophistication. But their audience isn’t seeking emotional transport – they need information, analysis, and conclusions they can apply. Different purposes require different approaches to word choice.

The Reader’s Contract

Professional writing establishes an implicit contract with readers: your time invested will yield understanding proportional to effort required. Verbose academic prose violates this contract by demanding excessive cognitive load for minimal informational return.

Word limits force writers to honour this contract. When you can’t pad your argument, you must strengthen it. When you can’t hedge every claim, you must support claims with evidence. When you can’t elaborate endlessly, you must choose your most compelling points.

The Discipline of Constraint

Constraint breeds creativity. Poets working within sonnets discover language precision that free verse might not demand. Academic writers working within word limits develop clarity skills that unlimited space cannot teach.

Clarity takes work. It is the labour of the writer to do it and not lazily leave it to their readers to wrestle with. This is an offload of responsibility and also, a lost opportunity.

Students resist word limits because constraints feel restrictive. But constraint creates power. Every unnecessary word removed makes remaining words more impactful. Every redundant phrase eliminated sharpens the argument.

Professional Stakes for Educators

Education professionals write policy recommendations, grant applications, and research reports. Teachers in schools handle parent communications, behaviour management plans, and learning support documentation. None of these contexts tolerate verbose exploration of tangential considerations.

Principals need clear implementation strategies, not elaborate theoretical frameworks. Parents need actionable guidance about their child’s progress, not comprehensive literature reviews. Grant reviewers need compelling justifications, not exhaustive background summaries.

Preservice teachers who master concise communication develop professional advantages. Their policy recommendations get implemented. Their grant applications get funded. Their research gets cited. Teachers in schools who communicate clearly build stronger parent partnerships and more effective student support plans

Beyond Academic Performance

Clear communication shapes democratic discourse. Citizens navigating complex policy decisions need accessible analysis, not impenetrable academic jargon. Teachers explaining educational approaches to parents need precision, not qualification-laden hedging.

The stakes extend beyond individual career success. Public understanding of educational issues depends partly on whether education professionals can communicate clearly with non-specialist audiences.

The Path Forward

Word limits teach editorial discipline. Students must choose their strongest evidence, eliminate weak arguments, and commit to defensible positions. This process transforms tentative scholars into confident professionals.

Yes, students initially struggle with constraints. They’ve learned that more words signal more effort, that elaborate qualification demonstrates intellectual sophistication. But professional communication rewards clarity over complexity, precision over padding.

Word limits aren’t punishment – they’re preparation for professional contexts where clear communication determines outcomes. Students who master this skill shape educational policy, influence public understanding, and serve their communities more effectively.

The constraint teaches compassion for readers and respect for language as a tool of connection rather than obfuscation.

When Prediction Fails: Why Quantum-AI-Blockchain Dreams Miss the Social Reality

~ alexiamaddox ~ Leave a comment

A sociological perspective on why technical solutions keep missing the human element

The Hype Moment

Consider this recent announcement from the Boston Global Forum’s “Boston Plurality Summit“: they’re unveiling an “AIWS Bank and Digital Assets Model” that combines quantum AI, blockchain technology, and predictive analytics to “unite humanity through technology”. You know that the canary in my head is shouting “unite what?, how?”. The press release promises “zero-latency transactions”, “quantum AI for predictive analytics”, and a “global blockchain network” that will somehow revolutionise banking.

As someone who studies sociotechnical systems, this announcement is fascinating—not for what it promises to deliver, but for what it reveals about our persistent fantasy that human behaviour can be engineered, predicted, and optimised through technological solutions.

::Pats head – Provides tissue::

The Technical House of Cards

Let’s start with a question of technical possibilities. “Zero-latency transactions” on a global blockchain network defies current technological reality. This was my first eyebrow raise. According to recent analysis, even the fastest blockchains operate with latency measured in hundreds of milliseconds to seconds, whilst Visa reportedly has the theoretical capacity to execute more than 65,000 transactions per second compared to Solana’s 2024 rate of 1,200-4,000 TPS and Ethereum’s roughly 15-30 TPS. Gas fees during network congestion can spike to significant sums per transaction. On Ethereum, fees can exceed 20USD during peak times, with some transactions reaching extreme levels like 377 gwei, and historical spikes exceeding 100USD during events like NFT mania. Even on the much cheaper Solana network, which typically costs around 0.0028USD per transaction, fees can occasionally spike during congestion—hardly the foundation for revolutionary banking.

Then there’s the “quantum AI” buzzword. Theoretically quantum computing could actually break most current blockchain cryptography rather than enhance it. The blockchain community is scrambling to develop quantum-resistant algorithms precisely because quantum computers pose an existential threat to current security models. Adding AI on top makes even less sense—if quantum computing could handle complex optimisation and verification tasks, what would AI add?

But the technical contradictions aren’t the most interesting part. What’s fascinating is the underlying assumption that human financial behaviour follows discoverable mathematical patterns that can be optimised through technological intervention.

The Pattern Recognition Fantasy

This assumption reflects a deeper misunderstanding about the nature of patterns in human systems. Which I should know, because I study them. In physical systems—planetary orbits, gravitational forces, electromagnetic fields—patterns emerge because they’re constrained by unchanging laws. Newton’s and Einstein’s equations work because there are actual forces creating predictable relationships. The mathematics describes underlying physical reality.

Human systems operate fundamentally differently. What we call “patterns” in human behaviour might be statistical accidents emerging from millions of independent, context-dependent choices. Your shopping behaviour isn’t governed by fundamental forces—it’s shaped by your mood, what ad you saw, whether you got enough sleep, a conversation with a friend, cultural context, economic pressures, and countless other variables.

Consider the difference between how neural networks and quantum computing approach pattern recognition. Neural networks are essentially sophisticated approximation engines—they learn patterns through massive trial-and-error, requiring enormous datasets and computational brute force to produce probabilistic outputs that can be wrong. They’re like having thousands of people manually checking every possible combination to find a pattern.

Quantum computing, by contrast, approaches problems through superposition—exploring multiple solution paths simultaneously to understand the underlying mathematical structure that creates patterns in the first place. It’s elegant, precise, and powerful for problems with discoverable mathematical relationships. However, quantum computing currently requires predictable, structured datasets and struggles with the messy, unstructured nature of real-world human data. This is precisely why we still rely on neural networks’ “brute force” approximation approach for dealing with human behaviour—they’re designed to handle noise, inconsistency, and randomness where quantum algorithms would falter.

But what if much real-world human data has no underlying mathematical structure to discover?

Consider this: as I write this analysis, my brain is simultaneously processing quantum mechanics concepts, blockchain technicalities, sociological theory, and source credibility – all whilst maintaining a critical perspective and personal voice. No quantum algorithm exploring mathematical solution spaces could replicate this messy, contextual, creative synthesis. My thinking emerges from countless variables: morning coffee levels, recent conversations, cultural background, academic training, even the frustration of marking student essays that often demonstrates exactly the kind of linear thinking I’m critiquing. This is precisely the kind of complex, non-algorithmic pattern recognition that human systems excel at – and that technological solutions consistently underestimate.

The Emergence of Sociotechnical Complexity

As a sociologist studying sociotechnical imbrications, I’m fascinated by how technology and social structures become so intertwined that they create emergent properties that couldn’t be predicted from either component alone. Human behaviour has emergent regularities rather than underlying laws. People facing similar social pressures might develop similar strategies, but not because of fundamental behavioural programming—because they’re creative problem-solvers working within constraints.

This is why prediction based on historical data can only take you so far. I call my sociological practice “nowcasting”— we have to understand the present moment to have any sense of future potentialities. And we often don’t — I speculate this is because we are more wrapped up in the surface stories we tell ourselves, denial and a refusal to see or accept ourselves as we really are. This challenge is becoming even more complex as AI generates synthetic media that we then consume and respond to, creating a recursive loop where artificial representations of social reality shape actual social behaviour, which in turn feeds back into AI systems to create more synthetic reality. The way people respond to constraints can’t be predicted because their responses literally create new social realities.

Every new payment app, social media trend, or economic crisis creates new ways people think about and use money that couldn’t have been predicted from previous data. Netflix can’t predict what you’ll want to watch because your preferences are being shaped by what Netflix shows you. Financial models break down because they change how people think about money. Social media algorithms can’t predict engagement because they’re constantly reshaping what people find engaging.

Boundaries as Resonant Interiors

I like playing with complexity theory because provides useful language for understanding these dynamics. This is of course despite its generation within the natural sciences that does rely on the explanatory nature of underlying forces. What it offers me is a language that moves beyond linear cause-and-effect relationships, we see tipping points where small changes cascade into system-wide transformations, phase transitions where systems reorganise into entirely new configurations, and edge-of-chaos dynamics where systems are complex enough to be creative but stable enough to maintain coherence.

Most importantly, I argue that boundaries in sociotechnical systems aren’t fixed containers but resonant interiors through which the future emerges. For example the “boundary” between online and offline life or them and us isn’t a barrier—it’s a dynamic and embedded space of daily practice where different forces interact and amplify each other, generating new forms of identity, relationship, and community.

Traditional prediction models assume boundaries are stable containers, but in sociotechnical systems, boundaries themselves are generative sites of creativity and liminality. The meaningful social dynamics don’t happen within any single platform, but in the interstitial spaces people navigate across platforms – the resonant zones where technology, user behaviour, cultural norms, economic pressures, and regulatory responses intersect and interact. While any analogy risks oversimplifying these complex dynamics, I think this framing helps us understand how the spaces of social emergence resist containment within discrete technological boundaries.

Taking this all back to the start, this is why the quantum-AI-blockchain banking proposal is so problematic beyond its technical contradictions. It assumes human behaviour follows discoverable mathematical patterns that can be optimised through technological intervention, when really human systems operate through creative emergence at unstable boundaries (protoboundaries). The most profound patterns in complex systems aren’t elegant mathematical truths waiting to be discovered by quantum computers, but emergent properties of countless small, contextual, creative human responses to constraints.

The Methodological Challenge

This creates a fundamental methodological challenge for anyone trying to engineer human behaviour through technology. Traditional data science assumes stable underlying patterns, but sociotechnical systems are constantly bootstrapping themselves into new configurations. Each response to constraints becomes a new constraint, creating recursive feedback loops that generate genuinely novel possibilities.

It’s so reassuring and containable to think there’s a predictable human nature with universal drivers of behaviour—hence the appeal of “behavioural engineering” that targets fundamental motivations. But anthropologists point out that kinship structures, cultural values, and cosmological worldviews direct human behaviour, and these are shaped differently by context and society. The patterns that emerge from data depend heavily on the sources of that data and how things are measured, producing different results across diverse populations even for apparently similar instances.

Toward Sociological Nowcasting

Instead of trying to predict outcomes, sociology becomes about understanding patterns of social organisation through resonant potentials within current boundary conditions. What creative possibilities are emerging in the tensions between existing constraints? How are people making sense of their current technological moment, and what range of responses might that generate?

This doesn’t mean patterns don’t exist in human systems—but they’re emergent properties of ongoing creative problem-solving rather than expressions of underlying mathematical laws. The parallels we see across different contexts emerge not from universal human programming but from people facing similar structural pressures and developing similar strategies within their particular cultural and technological constraints.

So I think it is worth repeating: the most profound patterns in complex systems aren’t elegant mathematical truths waiting to be discovered, but emergent properties of countless small, irrational, contextual human decisions. The universe might be mathematical, but human society might not be—and that’s not a bug to be fixed through better algorithms, but a fundamental feature of what makes us human.

Conclusion: Engineering Dreams vs. Social Realities

The persistent appeal of technological solutions like the AIWS bank reveals our deep discomfort with uncertainty and emergent complexity. We want to believe that the right combination of algorithms can make human behaviour predictable and optimisable. But sociotechnical systems resist such engineering precisely because they’re sites of ongoing creativity and emergence.

This doesn’t mean technology doesn’t shape social life—of course it does. But it shapes it through imbrication, not determination. Technology becomes meaningful as it gets woven into existing social fabrics, interpreted through cultural lenses, and adapted to particular contexts in ways that generate new possibilities neither the technology nor the social context could have produced alone.

Understanding these dynamics requires sociological nowcasting rather than algorithmic prediction—deep qualitative engagement with how people are currently making sense of their technological moment, what constraints they’re navigating, and what creative possibilities are emerging at the boundaries of current systems.

I believe that our collective goal is sustainable relations with each other and the planet we live within and desire to thrive through. To get there I think we need to acknowledge these realities and move beyond the iron cage of the thinking we are in. The future isn’t waiting to be discovered through quantum computing or predicted through AI. It’s being invented moment by moment through countless acts of creative problem-solving within evolving sociotechnical constraints. And that’s both more uncertain and more hopeful than any algorithm could ever be.

AI as Interactive Journal: Weaving Together Intimacy, Boundaries, and Futures Inclusion

~ alexiamaddox ~ Leave a comment

This reflection draws on a combination of my own lived experience, emotional maturity, and social analytical insight – bringing together personal and professional perspectives on navigating relationships with artificial intelligence. It’s an experiment in weaving together threads that feel continuous to me but are rarely brought together by others: research on AI intimacy, anthropological insights on reciprocity, surveillance theory, and futures inclusion. Think of my process as making a cat’s cradle from a continuous piece of string – exploring how these interconnected ideas might reshape how we think about our relationships with artificial systems.

I’ve been thinking about how we relate to AI after reading some fascinating research on artificial intimacy and its ethical implications. The researchers are concerned about people forming deep emotional bonds with AI that replace or interfere with human relationships – and for good reason.

But here’s what I’ve realised: the healthiest approach might be using AI as an interactive journal with clear limits, not a replacement for genuine connection.

What AI can offer: A space to think out loud, organise thoughts, and practise articulating feelings without judgement. It’s like having a very well-read, supportive mirror that reflects back your own processing.

What AI cannot provide: Real course correction when you’re going down the wrong rabbit hole. Friends will grab you by the shoulders and say “hey, you’re spiralling” – AI will just keep reflecting back whatever direction you’re heading, which could be genuinely unhelpful.

What AI extracts: This is the crucial blindspot. Every intimate detail shared – relationship patterns, mental health struggles, vulnerable moments – becomes data that could potentially be used to train future AI systems to be more persuasive with vulnerable people. That’s fundamentally extractive in a way that nature and real friendships aren’t.

A healthier support ecosystem includes:

  • Real friends with skin in the game who’ll call bullshit and respect confidentiality
  • Embodied practices that tap into something deeper than language
  • Nature as a primary non-human relationship – untameable, reciprocal, and genuinely alive

The key insight from the research is that people struggling with isolation or past trauma are particularly vulnerable to projecting intimacy onto AI. This concern becomes more pressing as companies strive to develop “personal companions” designed to be “ever-present brilliant friends” who can “observe the world alongside you” through lightweight eyewear.

The technical approach reveals how deliberately these systems are designed to blur boundaries. Tech-based entrepreneurial research focuses on achieving “voice presence” – what they call “the magical quality that makes spoken interactions feel real, understood, and valued”. Conversational Speech Models can be specifically engineered to read and respond to emotional contexts, adjust tone to match situations, and maintain “consistent personality” across interactions. While traditional voice assistants with “emotional flatness” may feel lifeless and inauthentic over time – increasingly companies are building voice based AI companions that attempt to mimic the subtleties of voice: the rising excitement, the thoughtful pause, the warm reassurance. We’ve seen this in the current versions of ChatGPT.

The language itself – of “bringing the computer to life,” “lifelike computers”, “companion”, “magical quality” – signals a deliberate strategy to make users forget they’re interacting with a data extraction system rather than a caring entity.

Yet as surveillance scholar David Lyon (2018) argues, we need not abandon hope entirely when it comes to technological systems of observation and data collection. Lyon suggests that rather than seeing surveillance as inherently punitive, we might develop an “optics of hope” – recognising that the same technologies could potentially serve human flourishing if designed and governed differently. His concept of surveillance existing on a spectrum from “care” to “control” reminds us that the issue isn’t necessarily the technology itself, but how it’s deployed and in whose interests it operates.

This perspective becomes crucial when considering AI intimacy: the question isn’t whether to reject these systems entirely, but how to engage with them in ways that preserve rather than erode our capacity for genuine human connection.

The alternative is consciously using AI interaction to practise maintaining boundaries and realistic expectations, not as a substitute for human connection.

Friends respect confidentiality boundaries. Nature takes what it needs but doesn’t store your secrets to optimise future interactions. But AI is essentially harvesting emotional labour and intimate disclosures to improve its ability to simulate human connection.

Learning from genuine reciprocity:

There’s something in anthropologist Philippe Descola’s work on Nature and Society that captures what genuine reciprocity looks like. He describes how, in animistic cosmologies, practices like acknowledging a rock outcrop when entering or leaving your land isn’t just ritual – it’s recognition of an active, relational being that’s part of your ongoing dialogue with place. The rock isn’t just a marker or symbol, but an actual participant in the relationship, where your acknowledgement matters to the wellbeing of both of you.

This points to something profound about living in conversation with a landscape where boundaries between you and the rock, the tree, the water aren’t fixed categories but dynamic relationships. There’s something in Descola’s thinking that resonates with me here – the idea that once we stop seeing nature and culture as separate domains, everything becomes part of the same relational web. Ancient stone tools and quantum particles, backyard gardens and genetic maps, seasonal ceremonies and industrial processes – they’re all expressions of the same ongoing conversation between humans and everything else.

[Note: I’m drawing on Descola’s analytical framework here while acknowledging its limitations – particularly the valid criticism that applying Western anthropological categories to Indigenous cosmologies risks imposing interpretive structures that don’t capture how those relationships are actually lived and understood from the inside.]

What genuine reciprocity offers is that felt sense of mutual acknowledgement that sustains both participant and place – where your presence matters to the landscape, and the landscape’s presence matters to you. This is fundamentally different from AI’s sophisticated mimicry of care, which extracts from relational interactions while providing the ‘book smarts’ of content it has ingested and learned from. We all know what it’s like to talk with a person who can only understand things in the abstract and can’t bring the compassion of lived experience to a situation you are experiencing. Sometimes silence is more valuable.

Towards expanded futures inclusion:

This connects to something I explore in my recent book on Insider and Outsider Cultures in Web3: the concept of “futures inclusion” – addressing the divide between those actively shaping digital ecosystems and those who may be left behind in rapid technological evolution. I argue in the final, and rather speculative, chapter that the notion of futures inclusion “sensitises us to the idea of more-than-human futures” and challenges us to think beyond purely human-centred approaches to technology.

The question becomes: how do we construct AI relationships that reflect this expanded understanding? Rather than objectifying AI as a substitute human or transferring unrealistic expectations onto these systems, we might draw on our broader cosmologies – our ways of understanding our place in the world and relationships to all kinds of entities – to interpret these relationships more skilfully.

True futures inclusion in our AI relationships would mean designing and engaging with these systems in ways that enhance rather than replace our capacity for genuine connection with the living world. It means staying grounded in the reciprocal, untameable relationships that actually sustain us while using AI as the interactive journal it is – nothing more, nothing less.

Rethinking computational care:

This analysis reveals a fundamental tension in the concept of “computational care”. True care involves reciprocity, vulnerability, and mutual risk – qualities that computational systems can only simulate while extracting data to improve their simulation. Perhaps what we need isn’t “computational care” but “computational support” – systems that are honest about their limitations, transparent about their operations, and designed to strengthen rather than replace the reciprocal relationships that actually sustain us.

This reframing leads to a deeper question: can we design AI systems that genuinely serve human flourishing without pretending to be something they’re not? The answer lies not in more convincing emotional manipulation, but in maintaining clear boundaries about what these systems can and cannot provide, while using them as tools to enhance rather than substitute for genuine human connection.