New models every week, new use cases, endless what-ifs. Most of us are swimming in it and still feel behind, and I've been there too. In that overwhelm, the line between what AI should do and what we should do blurs.

That's how AI brain rot creeps in: not from using AI a lot, but from handing over the parts that make us us. The first idea, the judgment call, our own flavor. Then we feel dumber and reach for AI even more. The fix isn't using AI less. It's deciding which parts you keep.

The 60-second self-check

Answer yes or no:

  • I open AI before I've formed my own view
  • I accept the first decent answer without checking it
  • I couldn't explain my last AI-assisted piece of work without looking at it
  • My ideas have started to sound like everyone else's
  • I've stopped doing something I used to be good at, and haven't decided that on purpose

Two or more yeses? Start with the shifts below.

The seven shifts

1. From "AI drafts first" to "you think first." In an MIT study, people who wrote on their own and then switched to AI showed higher recall and brain activation. People who switched the other way, from AI to writing alone, showed signs of under-engagement. The order matters. The move: write two messy lines of your own take before you prompt.

PROMPT / COPY-PASTE
My take first: [2 lines]. Now show me what I'm missing, where I'm wrong, and what a sharp critic would say.

2. From "the average is safe" to "the average is the enemy." In a Science Advances study, AI-assisted stories were rated better but were more alike. Same model, same ideas. The move: ask for the obvious ideas first, then ban them.

PROMPT / COPY-PASTE
List the 10 ideas anyone using AI would come up with for this. Now give me ideas that avoid all 10.

3. From "picking is lazy" to "choosing is the skill." Taste is selection. One polished answer means AI made the choice for you. The move: ask for three genuinely different options, pick one, and write one line on why. That line is the muscle.

4. From "AI as answer machine" to "AI as sparring partner." The move: make it question you, grade your reasoning, or make the problem harder.

PROMPT / COPY-PASTE
Don't solve this. Ask me questions one at a time until I solve it myself. If I'm stuck twice, give a hint, not the answer.

5. From "trust the confident answer" to "your confidence is the safeguard." In a Microsoft and Carnegie Mellon survey of 319 knowledge workers, more confidence in AI went with less critical thinking, and more confidence in their own skills went with more. The move: before accepting an output, name one thing you'd check if a junior colleague had handed it to you. Then check it.

6. From "saved means learned" to "recalled means learned." In the MIT study, people who wrote with AI struggled to quote their own essays. The move: after AI-assisted work, explain it back in three sentences without looking.

7. From "all friction is waste" to "some friction is the training." The move: for the one skill you're known for, do the hard part yourself.

What the research actually says

None of this says AI makes you worse. How you use it decides what you keep.

Your map: delegate, collaborate, own

BucketWhat goes hereHow you use AI
DelegateWork you don't need to get better atHand it off fully
CollaborateWork where AI makes you faster, but you still judgeShifts 1, 3 and 5
OwnThe skills you're known for, or want to beShifts 4, 6 and 7. AI trains you, it doesn't replace you
PROMPT / COPY-PASTE
Here's my recurring work: [list]. Sort it into delegate, collaborate and own, based on what I'm trying to get better at: [goals]. Challenge any placement that looks like I'm outsourcing something I'll regret.

Start with one

  • Take the self-check
  • Fill in your map
  • Pick one shift for this week (shift 1 is the easiest)
  • In a week, redo the self-check

Go deeper: Stop Saving Everything: An AI Filter That Keeps What's Worth Remembering · Follow-Up Prompts That Fix Weak AI Answers · How to Give AI Taste

Research details checked against the linked papers on September 18, 2026.