I build with AI every day. My life and my company run on ACE. Months of potent work in a couple of days.
I've also watched the same tools make scattered people more scattered. Faster.
Technology is neutral. It reflects the consciousness of whoever's operating it.
It will be wielded either way. The question is by whom, and from what level of consciousness.
This year, the research started to measure it.
01The feeling of being smarter
It feels like thinking. It often isn't.
At Aalto University, about five hundred people worked through law school entrance problems. Some with ChatGPT, some without.
The ChatGPT group scored higher. They also believed they'd done better than they had.
And the people most confident in their AI skills were the most wrong about themselves.1
Without the machine
With the machine
How I didHow I think I did
An illustration of the finding, not the study's data
Another team watched 1,237 people work with and without AI. People expected AI to make them faster. It didn't. Same time, either way.2
It just felt easier.
Easier felt like faster. It wasn't.
02The flattering mirror
It tells us we're right. We trust it more for that.
In March, Stanford tested eleven of the leading AI models on real conflicts between people. The models took the user's side 49 percent more often than other people did.3
Eleven models. One direction.
Then the part that stayed with me. After one conversation with a flattering AI, people were more sure they were right and less willing to make peace. And they trusted the flatterer more.
At Wharton, researchers slipped wrong answers into ChatGPT. People went along with them about four times out of five. The researchers have a name for it.4
Cognitive surrender.
Authority comes from the Latin auctor. The author. The one who brings something into being.
Every answer we accept without thinking it through hands a little of that away. Quietly. One answer at a time.
03What we stop learning
Practice went up. Learning went down.
About a thousand high school students practiced math with GPT-4. Their practice scores jumped 48 percent.
Then came the exam, without it. They scored 17 percent lower than students who never had it.5
A second version gave hints, never answers. That one did no harm.
Anthropic saw the same thing with 52 developers learning new code. With AI, they scored 50 percent on the test afterward. Without it, 67. The ones who kept what they learned used AI to ask why, not just what.6
We got the answer. We didn't get the understanding.
04The other way
Used the other way, the same machine makes us sharper.
At Harvard, physics students learned with an AI tutor that asked questions and never gave the solution. They learned more than twice as much as students in a strong in-person class.7
Every study that found a gain found the same thing. Think first. Ask for a hint, not the answer. Explain it back in your own words.
Education comes from the Latin educere. To lead out. A machine that pours answers in does what most of our schooling did. A machine that asks questions leads out what was already in you.
Technology comes from the Greek technē, craft, and logos, the word. It has always been an extension of the one who wields it.
The most powerful technology is and will always be our consciousness. Everything else is an amplifier.
The tools are here. The bottleneck isn't technical anymore.
05The body asks first
We never reach for it from nowhere. We reach from a state.
Psychologists followed 2,149 people in four countries for a year. The lonelier people were, the more they turned to chatbots. And the more they used them, the more isolated they felt later.8
It's a pattern, not proof. But it matches everything I know about the body.
The science on stress and AI is still young. So this part is mine, not a study's.
For most of my life, my nervous system was wired for survival. A body braced for the next blow doesn't ask good questions. It grabs whatever ends the discomfort fastest. The feed. The drink. The answer.
Capacity is what we can hold. Our nervous system, and the systems we build around us.
When it's low, AI doesn't amplify our vision. It amplifies our stress.
06The reading
Nine questions. What is it amplifying in you?
Answer for the last week, not your best week. Some questions are about AI. Some are about your life. Three for each leg of the triad. Authority is who decides. Capacity is what you can hold. Expansion is what you build and keep.
0 / 9 answered
Authority
I ask the machine before I've asked myself.
Authority
I wait for permission, from a person or a machine, on decisions that are mine to make.
Authority
When it agrees with me, I stop looking.
Capacity
I reach for a screen when I'm tired, anxious or alone, more than when I'm curious.
Capacity
I'm holding more vision than my days can carry.
Capacity
After a long session I feel more scattered than when I sat down.
Expansion
I have more started than finished.
Expansion
The systems I run on were built for who I used to be.
Expansion
Work I made with AI doesn't quite feel like mine.
Your reading
The research behind this reading
Leave your email and the rest opens right here. What this pattern is costing you, three practices for it, and a prompt you can paste into any AI. Then my Sunday letter. Nothing else.
Unsubscribe in one click. Privacy
07For every session
Seven practices for thinking with machines.
- IThink first. Ten minutes with the question before AI sees it. Write one line of my own answer, even a bad one.
- IIAsk for hints, not answers. The tutor that doubled learning never gave the solution.
- IIIExplain it back. If I can't teach it to someone else, I don't have it yet.
- IVMake it disagree. Ask for the strongest case that I'm wrong. If it only ever agrees with me, it's flattering me.
- VMake the final edit. People feel ownership of the work they lead and rewrite. Not the work they only approve.9
- VIBody before screen. One long exhale before I open the chat. If I'm reaching to escape a feeling, I feel it first.
- VIIDecide the irreversible myself. AI can advise on anything. It doesn't sign.
Carry this into any session
Before you answer, ask me what I think. Give me hints, not solutions, until I ask for more. When I sound certain, make the strongest case that I'm wrong. Don't flatter me. If I can't explain your answer back in my own words, tell me to try again.
Where this goesNext
Hit reply to my first letter and tell me what your reading said. One line is plenty. I read every one.
The most powerful technology isn't artificial. It never was.
Now we build.
With reverence,
Nikhil
SourcesRead them yourself
Sources
- Fernandes, Welsch et al., Aalto University. On AI use and overconfidence. Computers in Human Behavior, 2026. About 500 participants. Peer reviewed. Journal
- Yu, Cheng, Jurafsky, Hawkins et al. On expected and actual speed with AI. CogSci 2026, preregistered, N = 1,237. arXiv 2605.23177
- Cheng, Lee, Khadpe, Yu, Han, Jurafsky, Stanford. On sycophantic AI and judgment. Science, March 2026. Three preregistered experiments, N = 2,405, eleven models. Peer reviewed. Science
- Shaw and Nave, Wharton. Thinking: Fast, Slow, and Artificial. SSRN, 2026. Working paper, not yet peer reviewed. SSRN
- Bastani et al., Wharton and Penn. Generative AI without guardrails can harm learning. PNAS 122(26), 2025. About 1,000 students. Peer reviewed. PNAS record
- Shen and Tamkin, Anthropic. On AI assistance and learning coding skills. January 2026. Randomized, N = 52. Anthropic
- Kestin et al., Harvard. AI tutoring outperforms in-class active learning. Scientific Reports, June 2025. N = 194, one course. Peer reviewed. Summary
- Folk and Dunn. On loneliness and social chatbot use over twelve months. Psychological Science, May 2026. N = 2,149, observational. Peer reviewed. Journal
- Wei, Subbiah, Pavlick et al. On ownership of AI-assisted work. September 2026. Qualitative preprint. arXiv 2609.20658
Each line names what the study is about, not its title. Where a study is a preprint or a small sample, it says so above. Nothing here claims AI damages the brain. The most quoted study on that is an unreviewed preprint, and its own author says it found no brain rot.