What daily AI use did to my thinking: an honest audit

An Honest Audit of AI's Effect on My Mind

Daily AI use made me faster and shallower at once. A 2025 Microsoft and Carnegie Mellon study of 319 workers found higher trust in AI predicts less critical thinking. I noticed the same: fewer dead ends, but thinner original ideas. Mollick's Co-Intelligence reframed it for me, and I now protect AI-free hours for the thinking that actually compounds.

Let me be honest about a cost I did not see coming. I build with frontier models every day, Claude Opus 4.8 and GPT-5.5 open in tabs from morning to night, and after two years I sat down to audit what that has actually done to my mind rather than my output. The output is obviously better. The mind is a more complicated ledger, and I owe it to anyone reading to keep both columns visible instead of selling the optimistic half.

The gains first, because they are real. My drafting friction is gone; I no longer stare at blank pages. I learn unfamiliar domains faster because I can interrogate a patient explainer at midnight. And my range of references widened, because the model surfaces adjacent ideas I would never have searched for. Ethan Mollick, in Co-Intelligence, frames this as treating AI as a capable colleague rather than a tool, and on my best days that is exactly the texture of the work, a genuine back-and-forth that sharpens a position faster than solitude would.

Now the debit column, which is the honest part. I noticed my first instinct on any hard problem had quietly become "ask the model" rather than "sit with it." That instinct has a name and a measurement. A 2025 study from Microsoft Research and Carnegie Mellon, surveying 319 knowledge workers across 936 real tasks, found that the more people trusted the AI, the less critical thinking they reported doing, while those who trusted their own judgment thought harder. The same study found AI users produced a less diverse set of outcomes for the same task. That last finding stung, because I had felt it: my ideas were arriving pre-rounded, sanded toward the statistical middle the model lives in. There is now a body of 2026 research on this exact homogenization, the way shared models compress distinct reasoning styles into a common voice.

The starkest evidence is physiological. MIT's Media Lab study, titled Your Brain on ChatGPT, put 54 people in three groups writing essays, measured them with EEG over four months, and found that brain connectivity scaled down with the amount of AI support; the assisted group showed the weakest neural coupling and, tellingly, often could not quote the essays they had just produced. They had generated the words without owning the thought. I recognized that disownership. There were ideas under my name that I could not have reconstructed from scratch, and that is not a small thing for someone whose work is supposed to be thinking.

So what did I change, concretely? I now run a daily block of AI-free deep work, Cal Newport's term from Deep Work, where the hard problem and I are alone until I have a real position. Only then do I bring in the model, to attack the position rather than to author it. I treat sycophancy as a known hazard; OpenAI had to pull a GPT-4o update in 2025 for being too agreeable, and I assume any frontier model will flatter my framing unless I instruct it not to. And I keep a literal ledger, a weekly note on whether a given idea was mine or merely retrieved, because what gets measured gets defended. This is the same exploration discipline I have written about in why exploration matters more than optimization, the willingness to take the longer, less efficient path because the efficient one converges on what everyone else already has.

The audit's conclusion is not abstinence, which would be both dishonest and stupid given what these tools do. It is sequencing. Think first, alone, until it hurts a little; then collaborate. The danger was never the model's intelligence. It was my own eagerness to skip the part of thinking that is uncomfortable and therefore generative. AI removes friction, and friction, it turns out, was where a fair amount of my original thought was being made. Guard one or two hours of it deliberately, and you keep the part of your mind that compounds.