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.