I caught myself doing it again last week. A complex piece of work landed, and before I’d even read it properly, I’d pasted it into my LLM of choice and asked for a summary. Not because I couldn’t understand it. Because it was faster. And in that moment, I wasn’t thinking. I was managing.
Advait Sarkar’s TED talk - “How to Stop AI from Killing Your Critical Thinking” - articulated something I’ve been feeling and reflecting on for a while. He describes how AI-assisted workflows are turning knowledge workers into what he calls “middle managers for our own thoughts.” We see fewer ideas, think about them less critically, and remember them less well. His research at Microsoft found that workers reported putting less effort into critical thinking when using AI - and that this effect was strongest in people who trusted AI the most and trusted themselves the least.
That finding stopped me. Because I recognise it in myself.
I’ve written before about automation bias - our well-documented tendency to accept the outputs of automated systems without adequate scrutiny. But what Sarkar describes goes further than that. Automation bias is about trusting incorrect outputs. This is about something more fundamental: the gradual atrophying of the thinking itself, regardless of whether the output is correct. You can get a perfectly accurate AI-generated summary and still lose something important - the interpretive work, the pattern recognition, the connections between ideas that only form when you do the cognitive labour yourself.
I notice this most acutely in my note-making. I use Obsidian extensively, and there’s a meaningful difference between notes I’ve written and curated myself and notes I’ve asked AI to generate from source material. The self-authored notes carry a mental model. I can recall the argument, the tensions within it, the links to other ideas. It’s the difference between a route you’ve walked and one you’ve only seen on a map. The information is the same. The knowledge isn’t.
This has direct implications for clinical practice, and it connects to concerns I’ve raised previously about ambient voice technology in healthcare. When a clinician uses an AVT tool to generate a consultation summary, the documentation may be accurate - but does that clinician carry the same depth of understanding of that patient interaction as one who constructed the record themselves? I’d argue they don’t. Clinical reasoning isn’t information retrieval. It’s pattern recognition built on accumulated interpretive experience. A summary can tell you what was said. It can’t replicate the cognitive work of having been present, attentive, and actively synthesising.
Some will counter that this doesn’t matter if the information can be surfaced reliably at each subsequent interaction. But surfacing facts is not the same as holding a mental model of a patient. The clinician who wrote their own notes doesn’t just have a record - they have a richer, more contextual understanding that shapes how they listen next time, what questions they ask, what they notice.
What gives me hope in Sarkar’s talk is that he doesn’t just diagnose the problem. His team at Microsoft Research in Cambridge have been prototyping what he calls “tools for thought” - AI systems built around the principle of productive resistance. Instead of generating your text, they offer lenses for reading, provocations that surface counterarguments, and challenges that keep you metacognitively engaged. The tool doesn’t think for you. It makes you think harder.
That’s a genuinely counterintuitive proposition in a market that sells AI on the promise of doing less. But it might be the more important one. Not AI that replaces your effort, but AI that makes your effort count for more.
I’d rather use a tool that argues with me than one that agrees with me faster.
Watch Advait Sarkar: How to Stop AI from Killing Your Critical Thinking (TED)
Originally published on LinkedIn on 21 June 2026.