Essay

An AI walked 67 people to the right answer for a month. Their own judgment came out no sharper.

The one capability nobody claims AI will replace is human judgment. MIT Media Lab just ran the first real measurement of what happens to that judgment when you route it through an AI for a month, and the honest result is that the AI got people to the right answer and taught them nothing they could keep.

Here is what they did. Sixty-seven people spent a month sorting real news from fake. Each item, they called it cold on their own, then argued it out with a GPT-4o system that pulled evidence from Google and pressed them toward the right answer, up to nine rounds a headline. At intervals the researchers took the AI away and tested them on fresh items with no help. Three numbers came back, and the piece only works if you hold all three. With the AI in the loop, answers ran about 21 points more accurate, which is the tool doing the work, not a faculty sharpening. Their cold, start-of-session judgment, the cleanest read on the skill they actually own, drifted down about 7 points over the month and never cleared stat-sig. And on fresh items scored right after a session, accuracy sat 15 points below where they began, a drop the authors pin partly on leaning on the tool minutes earlier rather than on any permanent loss. About one-fifth of them, the group the paper calls Dependency Developers, slid over the month from doing the work to taking the AI’s call, one of them summing it up as “they did this for me, so I was fairly passive.”

Strip the scariest number and the finding that holds is just the paper’s own title: talking to the AI fixed the answer in the moment and built no lasting skill. That is the sharpest test yet of a bet I have made in public, that AI has collapsed the cost of book knowledge and left human judgment standing as the moat. It does not refute the bet. It refines it, and the refinement is the whole game. A moat behaves less like a wall you merely build once, and more like a muscle. Load it and it holds, stop loading it and it fades, and a month of getting walked to the answer was enough to engender the skill atrophy. “Judgment survives AI” was always the easy half. The hard half, the one nobody had a number for, is that it survives only if you keep spending it.

Now the objection. This is one narrow task, spotting fake headlines, 67 people on Prolific, four weeks, no control group, not yet replicated. “The human moat” is my analogy, not their result; real-versus-fake detection is the cleanest lab proxy we have for the judgment I keep going on about, but a proxy is all it is. Read the study as the first thermometer down the mine, not the autopsy. And the tidy fix everyone reaches for does not survive the data. Yes, the people who came out sharper were the ones whose exchanges had the AI asking them questions rather than handing over the verdict. But that is a correlation inside a single group, not two tools run head to head, and those same questioners started the month sharper, so it may mark judgment as much as make it. Worth clocking who benefits, too. The lab builds question-asking AI tutors, so “ask beats tell” is a handy result for them, and I build an AI product, so “keep the human sharp” is a handy sermon for me. Discount us both.

What is left after the discounting still stands. Nobody mourns the mental arithmetic a calculator took, because arithmetic was never the moat. Judgment is the one asset the whole thesis says you get to keep, so a tool that stops building it is spending down the only thing you own. Maybe the variable was never how much you delegate. Maybe it is how the tool makes you think on the way to the answer.

So the real question for anyone building or buying one of these is not “does it make people faster.” Nearly all of them do. It is whether the thing is training you or retiring you, and the first hard look we have found that even the good version, the one that talks you all the way to the right answer, left people with nothing of their own to show for it.

Sources: