Looking beneath the productivity gains
The first clue comes from a report you may not have seen. Anthropic went through thousands of Claude conversations looking for signs of healthy scepticism. Nine in ten users never checked a fact or questioned how an answer was generated. Fewer than one in five asked for reasoning or flagged missing context and those numbers dropped further when the output looked more polished. We don't argue with good-looking answers.
Microsoft researchers saw a similar pattern when they surveyed hundreds of knowledge workers. When a generative model handled the heavy lifting, people shifted from creating to supervising and tended to assume the output was correct. Participants who trusted their own judgement engaged more critically, while those who were impressed by the technology scrutinised less. The more faith we put in the machine, the more our guard drops.
Early on in my own use of AI, I experienced this complacency first‑hand. On one project I let a model draft a document and found myself nodding along at the shiny prose. At first glance it looked great. But it didn’t take much digging beneath the surface to realise it was mostly AI slop, coherent sentences masking shallow logic. I’ve also asked team‑mates about work they’d completed with AI only to get back blank stares. They couldn’t tell me how they’d arrived at the answer because the model did the thinking and they tuned out.
Other studies echo this. Michael Gerlich’s research links frequent AI use to lower critical-thinking scores; cognitive offloading, it turns out, can dull the very muscles you need to evaluate what the machine just gave you. Younger participants both relied on AI more and scored lower on critical thinking. The pattern holds across all age groups, just more pronounced at the younger end.
The International AI Safety Report 2026 has a name for it: automation bias. In one large experiment, people were less likely to correct an AI's errors when doing so took effort, or when they already had a positive attitude toward the technology. Evidence from healthcare and other fields points the same way: the more we trust the tool, the more likely we are to miss mistakes we'd otherwise catch. Across sectors, the story is the same, slick tools can make us complacent.
Why this matters beyond the research
These findings don’t mean AI is rotting our brains. They mean that as the technology becomes smoother, our instinct to question fades. The interface says “job done”, so we assume there’s nothing more to see. That has real consequences. In creative work, unchallenged AI output can narrow our thinking. In analytical work, unverified numbers can cascade into bad decisions. And in leadership, delegating judgement can erode ownership.
What alarms me isn't that a few people skip fact-checking. It's that organisations have almost no visibility into these behaviours. We measure outputs relentlessly: emails sent, presentations produced, tickets closed. But how often are people challenging AI-generated ideas, checking sources or questioning conclusions? Most organisations simply don't know. Without that visibility, quality drift can remain invisible until it starts showing up in outcomes.
How to keep our brains in the loop
The fixes are simpler than the problem. Think of it this way: if a new colleague handed you a piece of work, you'd question it, probe it, push back. Do the same with AI. Treat the first response as a draft, not an answer; iterate, ask why, refine. Be explicit about how you want the model to work with you; only around 30% of users ever say 'explain your reasoning' or 'tell me when you're uncertain' and that instruction alone changes the dynamic. Build a lightweight check into AI-assisted work: which facts need confirming, what context might be missing, do the numbers actually make sense? And start measuring what you currently can't see - how often people question AI output, ask for sources, or revise what it produces. Those are your early signals of a team that's still thinking.
The question we should all be asking
AI will only get more persuasive. Better copy, cleaner code, nicer slides. The risk isn't that the technology gets worse. It's that it gets so good we stop interrogating it. The opportunity is that AI frees us to spend more time on strategy, creativity and the uniquely human parts of work; if we stay engaged enough to use it well.
So here's a challenge: in your next AI-assisted task, consciously push back. Ask the model why. Check a fact. Add a nuance it missed. Don't treat the first response as a final answer, treat it as a draft. Then share that experience with your team.
Because the real question AI puts to us isn't about productivity. It's whether we still trust our own judgement enough to push back. That's what I'm working on, making the invisible parts of thinking visible, so organisations can move fast without quietly going stupid.
