The Outsourcing of Certainty
A note first. Fridays are going to be different. The daily forecasts stay what they are — what changed, what it means, what to do about it. Friday is for the question underneath: what this is doing to the people using it.
Sometime today you asked AI a question you could have answered yourself. Not for lack of the information. Because sitting with it for four minutes was less appealing than having it handled in four seconds.
Open your own history and look at what you have been typing. Write this email. Is this a good idea. What should I charge for this. What am I missing here. Read this and tell me what you think.
Nobody typed those into Google in 2015, because Google could not answer them. They are not searches. They are moments where you hand over your uncertainty and get it back a few seconds later in complete sentences.
We may be watching the beginning of something worth naming: the outsourcing of certainty. For the first time, people have an always-available intelligence willing to resolve ambiguity for them before they have had to resolve it themselves.
What Every Tool Took
The calculator took arithmetic. GPS took navigation, and people who use it develop a measurably weaker sense of direction over time. Google took memory — there is research going back to 2011 showing that once people know a fact is retrievable, they remember where to find it rather than the fact itself.
Each of those was a reasonable trade. Nobody is worse off for skipping long division.
This one takes something else. Not the answer. The uncomfortable few minutes between not knowing and working it out.
That space used to be unavoidable. You had a question, and the only route through was your own head, for as long as that took. It was unpleasant enough that nobody would defend it on its own terms. It was also the space where judgment got built.
The answer arrives. The part where you would have found it doesn’t.
And this particular thing has no external store. A fact you never memorized is still retrievable. A judgment you never made is not anywhere. It exists only as a capacity inside one person, built by use, and it fades the same way.
Tomorrow’s forecast, in your inbox.
One email. Five minutes. Written for owners, not engineers.
The Endoscopists
I would have left this as an interesting idea if not for one study.
Four centres in Poland introduced AI polyp detection for colonoscopy at the end of 2021. Researchers compared the three months before with the three months after, looking only at the procedures done without AI assistance. Nineteen endoscopists, each with more than two thousand colonoscopies behind them.
Their unassisted detection rate fell from 28.4% to 22.4%. Six percentage points in absolute terms, a twenty percent relative decline, in twelve weeks.
These were experienced specialists performing a skill they had spent careers building. Three months of help, and they were measurably worse at their own work when the help was removed. Nobody chose that. Nobody noticed while it was happening. A commentary published alongside it called this the first real-world clinical evidence of deskilling.
The broader research points the same way, with less precision. A study of 666 participants found a strong negative relationship between frequent AI use and self-assessed critical thinking, with cognitive offloading as the mediating factor, sharpest among the youngest participants. Other work finds people place excessive trust in AI answers and hold onto them even when they are wrong — particularly under time pressure, which is when nearly all of us are asking.
The mechanism is not mysterious. Judgment is pattern recognition earned by sitting inside uncertainty long enough for something to click. AI is remarkably good at removing that interval. That is exactly why it is useful, and exactly why it is worth watching.
Where I Might Be Wrong
The colonoscopy study is observational rather than randomised, and other researchers have published objections to it. The wider work on offloading and critical thinking is early — the International AI Safety Report says so directly. One correlation is not a life plan.
The more useful caution is that the effect is uneven. Several studies find that passive use produces skill decay while deliberate, structured use appears to improve critical thinking. The variable turns out to be whether you are thinking alongside it or in place of it, rather than how much of it you use.
Which is the good news. The answer is one small change in how you engage, costing about thirty seconds, rather than staying away from the thing entirely.
The Forecast
Every profession that runs on knowledge work splits in two over the next five years, and the line will not fall where people expect. It will not separate AI users from holdouts, or experts from beginners.
It separates people who use AI to replace thinking from people who use AI to extend it.
The first group becomes genuinely productive — more output, more domains, real competence — while their discrimination stays roughly where it was the day they started, because nothing asked them to develop it. The second group looks identical for a while, uses the same tools and often more of them, and gets steadily sharper, because every exchange is a repetition rather than a substitution.
For three or four years you will not be able to tell them apart. Then the difference becomes obvious. Everyone will have access to intelligence. What becomes scarce is the ability to decide what deserves that intelligence in the first place.
Moderate to high rather than high, because the human half of this resists measurement and the research is early. What I am confident about is the mechanism, and the endoscopists.
Thirty Seconds
The next time you open it, stop before you type.
Answer the question yourself first. Out loud, or in two badly written sentences. Whether you are right does not matter.
Then ask. Then compare.
Some days you will be close and notice you knew more than you thought. Some days you will be wrong in a way that teaches you something, and being wrong on paper before you see the answer is worth more than a hundred correct answers you read straight through.
AI can give you an answer before you have one of your own. Most of the time that is useful.
Sometimes, wait thirty seconds. What you are protecting is the person capable of forming one.
Sources: Budzyń K, Romańczyk M, Kitala D, et al. Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study. Lancet Gastroenterol Hepatol. 2025;10(10):896–903 · Gerlich, 2025, survey of 666 participants on AI use, cognitive offloading and critical thinking · Sparrow et al., 2011, on search engines and memory · International AI Safety Report, 2026