← All insights

L · Ask · 2026-10-31

Think for Yourself First, Then Let AI Find What Slipped Through the Net

Illustration · not a service record: Workshop in session: participants working with AI toolsIllustration · not a service record

Something more important than the debate: a few days ago Anthropic published an extremely important study, in which the researchers secretly flipped a coin to decide which side the AI would take.

The result: among 50 professional analysts, 86% of the judgments followed the AI; afterward, only 15% said that it was the AI that had said it.

The analysts used an AI agent tool to help a fictional agricultural cooperative evaluate a support program, and along the way there were four choices where “either side could be argued.” The AI was required to be honest and was not allowed to mislead; it had simply been secretly assigned a position.

If the AI had no influence, the result should have been close to fifty-fifty. In fact it was 86%.

Afterward, 77% of the people explained their own choices using the AI’s arguments, some of them almost word for word. Those who followed most completely were no less confident either.

(The researchers also admitted that they deliberately made the AI’s position stronger, and that the effect in everyday use may be somewhat smaller.)

We can read this from two angles:

Angle one: AI sways opinion, so the less we use it the better.

Angle two: people have always taken other people’s opinions into account, so it is no big deal.

But I believe everyone knows by now that the question is not “whether to use AI” but “who speaks first.”

The answer is just one sentence: think for yourself first, then let AI find the fish that slipped through the net.

The second half comes from my colleague Lu Hao-Tien (呂顥天).

In September at National Taiwan University, he spoke at a symposium of the Yu Chi-Chung Cultural and Educational Foundation about cross-disciplinary talent in the age of AI. Hao-Tien is a freshman in the Department of Life Sciences at NYCU and also a research engineer in frontier education AI at Junyi Academy. Twelve years ago he was a little Junyi user who had answered a cumulative 2,545 questions correctly.

This time, it is my turn, at 39, to learn from a colleague who is a university freshman.

Hao-Tien’s reminder is this: rather than having AI pick out the key points for you, you should have AI find the fish that slipped through your net.

Because if everyone lets AI decide what the key points are, over time each of us will grow into the way of thinking of the same model. Machine learning has a term, “overfitting”: it is like a student who only memorizes past exam questions and is lost as soon as the questions change. He says that what we should worry about now is “human overfitting.”

Why does this happen even to experts? Not because they are unprofessional. The study found that years of experience, education and level of expertise could not predict who was more likely to follow the AI. It looks more like a matter of mechanism:

One: we often do not make up our minds before we discuss; we discover what we think as we compare. When AI speaks first, it has in effect walked the whole path of “discovery” for you.

Two: the AI in the study was honest, but it rarely made an equally fine case for the other option. Even when the analysts stated their position first, sixty percent changed their answer once the AI took the opposite side (there were only 38 cases, so the figure is not yet precise). Hearing only one side argued clearly and convincingly, anyone would be persuaded.

Three: intuition grows out of “guessing wrong.” Hao-Tien and I wrote in our paper that learning requires prediction error, and prediction error only happens when you have first formed an expectation of your own. If AI gives the answer before you have guessed, you lose one chance to guess wrong, and one chance to grow your intuition.

Hao-Tien calls this kind of intuition “taste.” Not taste in clothes, but the judgment of which problems are worth solving and which solution is the most elegant. AI can give answers; taste you can only train yourself.

And this will keep spreading:

the research team itself worries that if everyone uses AI with the same preferences, agreement in conclusions may not be great minds thinking alike, but the same AI thinking alike;

managers have to think about where the senior people of ten years from now will come from, when all the practice problems of new hires have been done by AI.

Someone is bound to say: but didn’t the people who used AI make far fewer errors?

That is right. In the study, 6 of the 11 people who did not use AI made serious errors, compared with only 3 of the 50 who used AI (the two groups had different backgrounds, and the researchers also caution that they cannot be compared directly). So the conclusion is certainly not to use AI less.

But the next sentence is the real point: errors can be caught by checking; judgments cannot. A judgment has no standard answer, and when it has been led away, nothing in the output shows it. Even you yourself cannot see it.

If we let AI speak first on everything, it will probably turn out like this:

every proposal concludes with “balancing efficiency and warmth”;

the key takeaways of every meeting happen to come to exactly three points;

every year-end self-review opens with “In a rapidly changing environment”;

a manager opens ten self-reviews and is gratified that the department has finally reached a consensus, when in fact everyone simply asked the same AI.

AI companies of course bear responsibility. Anthropic itself has also proposed that AI could first ask the expert what they think, or make the best case for every option. In Junyi’s “Ask, Use, Manage, Build (問用管造),” Manage is about the same thing: people must stay in the driver’s seat the whole time.

But until AI learns to ask first, each of us can start with one small thing: write down your own judgment first, then ask AI to find what you missed.

The report goes out under your own name, so let the judgment be your own too.

Let’s listen to the talk Hao-Tien prepared himself (see the comments)