AI Is Most Useful When You Don’t Need an Answer
Something has changed in the way my brain works since I started using AI every day.
The number of ideas I can process has exploded. Things that once took days of research happen during breakfast. I can move from an idea to a website, an offer, an argument or a business model at a speed that would have looked ridiculous a few years ago.
I love it.
I am also no longer convinced all of it is good.
Ten Answers Where There Used to Be One
When every question produces ten intelligent answers, every idea produces five viable directions, and every direction can be challenged by another model, you end up with more intelligence than you know what to do with.
Which raises a question I keep circling. When I used to have one good idea, I acted on it. Now I can generate ten in the same hour. What happens next?
What happens is that some of those hours turn into rabbit holes. Which idea is best. Which model to trust. Whether the machine sees something I do not. Whether I am being stubborn by trusting myself or foolish by letting it talk me out of something I already knew.
The distance between having an idea and building it has collapsed. The distance between seeing a possibility and committing to it has not, and some weeks it feels longer than it used to.
You can spend an hour getting smarter and finish the hour less certain what you think.
Tomorrow’s forecast, in your inbox.
One email. Five minutes. Written for owners, not engineers.
The Kenyan Entrepreneurs
Researchers from Harvard Business School and Berkeley ran a field experiment with 640 small business owners in Kenya. Half were given access to a GPT-4 business assistant over WhatsApp. The other half received a written business guide.
On average, nothing happened. No measurable effect on revenue or profit across the whole group.
Underneath the average, two opposite results. Entrepreneurs who were already performing well improved by just over 15%. Entrepreneurs who were struggling did nearly 10% worse.
The part that matters is why. The researchers checked whether the two groups were asking different questions or receiving different advice. They were not. What separated them was which pieces of advice they chose to implement.
Lower performers were more likely to act on generic suggestions — lower your prices, advertise more — even when those were poorly suited to their situation. Higher performers picked the recommendations that fit.
Same machine. Same quality of advice. The difference was judgment about which advice deserved to be acted on.
That result probably matters more than anything anyone will teach you about prompting.
The Harder Version
Handing over a task I could do myself usually feels efficient, and mostly it is. Occasionally there is a sense that I have handed over something I needed to exercise.
That feeling gets much sharper with decisions that matter. Should I pursue this. Kill this. Change direction. Trust this instinct. Spend the money. Walk away.
AI can build an excellent case for almost anything. If you dislike the answer, another prompt produces another interpretation. Another model gives a second opinion. Then a third.
And because every round contains legitimate intelligence, it never feels like avoidance. It feels like research.
Useful information has become functionally infinite. Human clarity has not.
What the Brain Scans Suggest
An MIT Media Lab team had participants write essays using an LLM, a search engine, or nothing, while recording brain activity. Fifty-four people across the first three sessions. It is a preprint rather than peer-reviewed work, and the sample is small, so it deserves to be read as a signal rather than a verdict.
The LLM group showed the weakest neural connectivity of the three. They reported the lowest sense of ownership over what they had produced. And 83% of them could not supply a single correct quotation from an essay they had finished minutes earlier. Almost everyone in the other two groups could.
Then in a final session they swapped the groups.
People who had written unaided and were then given the AI showed higher neural connectivity than the AI group had shown in any of its own sessions. The authors read that as rewriting with AI, after writing without it, engaging more extensive brain activity than using AI from the start.
Only 18 people completed that session, which makes it the thinnest finding in an already thin study. It is also the one pointing somewhere useful. Order of operations may matter more than whether the machine is in the room.
When It Works and When It Does Not
The conversations that leave me sharper have one thing in common. I am riffing.
No agenda. No answer I need to reach. I throw something out, it pushes back, I reject part of it, it catches a contradiction, I follow that somewhere else. Occasionally we arrive at an idea I could not have produced alone and it could not have produced without me.
Those conversations leave me with more of my own thinking rather than less.
The ones that go wrong are the ones where I need the machine to settle something. That is when another round of analysis starts looking irresistible, because the next response might finally deliver certainty. There is always another response available.
An endless supply of useful information turns out to be perfect fuel for a rabbit hole.
Clarity does not work that way. It arrives when it arrives, and no amount of additional analysis manufactures it. Past a certain point, continuing to prompt is a way of postponing the irreducibly human act of choosing without being sure.
Which gives me the rule I have been looking for.
Use AI freely to explore. Be far more careful when you need it to decide.
The Forecast
Human discernment becomes a premium skill.
As these systems get more capable and more persuasive, the advantage moves from people who are best at extracting answers toward people who are best at deciding which answers deserve to influence them.
The skill worth developing next is knowing when to stop prompting and make the call yourself.
The people who become exceptional with AI will probably look different from the people who use it most. They will keep a conversation open without surrendering where it ends. They will let the machine challenge a belief without assuming that eloquence is the same as being right. They will use unlimited intelligence to widen the field, then accept that a person still has to choose.
High confidence, because the mechanism is already visible. The Kenyan entrepreneurs received similar advice and got opposite results based partly on what they chose to act on. As AI makes good advice abundant, selecting what fits your situation becomes increasingly consequential.
Sources: Otis N, Clarke R, Delecourt S, Holtz D, Koning R. The Uneven Impact of Generative Artificial Intelligence on Entrepreneurial Performance: Evidence from a Field Experiment in Kenya. Management Science · Kosmyna N, et al. Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. MIT Media Lab, 2025 (preprint)