A Doctor Used AI for a Few Months and Lost 20% of His Own Polyp-Detection Skill! Those Who Sell AI Found Themselves Learning Worse with It, 7 Models Trading: the Stronger They Get the More Alike They Become, 900 Million People Plugged into the Same Brain, the CCP's Models Have an Extra Filter, the Market Buys Together and Sinks Together! Independent Beats Right

This episode explores whether human collectives become wiser or stupider in the era of AI. The host points out that while AI has brought about the "equalization of answers," because true information originates from "surprise" and AI provides statistical modes, it leads to the degradation of individual creativity and practical capabilities in fields like medicine and programming. At the same time, when 900 million people worldwide are plugged into the same public brain—or even subject to question-bank filtering regulated by the CCP—collective information converges, exacerbating the herd effect in capital markets and making liquidity fragile. The host advises treating AI as a notebook rather than a brain, maintaining independent judgment, actively seeking counterexamples, and emphasizing that in the era of AI, "independence is worth more than correctness."

Video link:

Full transcript:

Lately I have been pondering a question: in the era of AI, are human collectives becoming wiser, or stupider? Of course, there are both affirmative and opposing sides to this. My view is that in the era of AI, answers have become more standardized, but information has actually decreased. Equal access to answers does not equal equal access to information. Going a step further, the more uniform the answers, the worse the information capacity actually becomes. For this, we must return to Claude Shannon's basic definition of information from 1948. Shannon's definition of information can be summed up in just two words: namely, surprise. What is information? Only that which can elicit your surprise, change your decision, or even alter the background of your decision-making can be called information. To take it a step further: if I say "The torrent dashes down," and ask you to finish the line, basically anyone can do it—it's "three thousand feet." But if I ask you how AI will develop tomorrow, or ask you how the capital markets will perform tomorrow, that yields countless answers. In other words, does a universally known answer contain any amount of information? None at all. Information is something that you were completely unaware of, but is suddenly placed right in front of you and impacts your decision-making—only that can be called information. If we measure with this yardstick, today's problem becomes: in the era of AI, answers have increased, but information itself has not. To explore this issue, let's start from the most microscopic level: first, we measure an individual's answers; next, we measure whether an individual can still generate surprise; and finally, we measure how much information is left when a group of people is added together.

First, I believe that equal access to answers is indeed a valid premise, and there is data to support it. A working paper from the National Bureau of Economic Research published in February this year contained a very interesting experiment. The experiment asked 1,174 adults aged 25 to 45 to solve business decision-making problems. Without AI, the gap in answer quality between highly educated and less educated participants was 0.55 standard deviations. With AI assistance, this gap narrowed to 0.14 standard deviations—erasing three-quarters of the difference directly. In this context, what does standard deviation measure? It measures how widely spread out the scores of a group of people are. You can think of it as two groups that originally differed by more than half a tier, but with AI, the gap shrank to just a tiny margin. The evidence in education is even more consistent. A meta-analysis published in Humanities and Social Sciences Communications on March 26 this year—where a meta-analysis synthesizes and calculates data across a pool of studies—combined 35 controlled classroom experiments involving a total of 4,193 students. The calculated result showed that students using ChatGPT experienced a moderate improvement in academic performance. Then, in June this year, another meta-analysis focused specifically on critical thinking. It synthesized 26 studies and found that students using generative AI saw their critical thinking scores increase rather than decrease. Therefore, paired with AI, a person gets more accurate answers and learns more, effectively flattening the barrier between different levels of education—this is equal access to answers. However, once you measure these experiments with Shannon's ruler of information theory, what are they actually measuring? They are all measuring questions with standard answers, business case studies, or classroom exam questions. These types of questions share a common feature: the answer was already sitting there all along. It's just that previously, you had to spend years studying to get the answer; now, you just type into a chat box and it's done. Of course, that's a good thing, but according to Shannon's definition, what you're reaching for is something that was already sitting there, which doesn't count as new information at all. If 1,174 people all get the same question right, nobody has surprised anyone. Thus, equal access to answers is about the group's ability to reach that distance to an existing answer, not about generating new information.

What about humans' own capacity to generate surprise? The evidence on this side is even stronger. On June 1 this year, a research team at a cancer center in Lyon, France, published a review in a journal of the European Society for Medical Oncology. What did this review investigate? It examined whether doctors' skills deteriorate after using AI. Inside, there were records of endoscopists from four hospitals in Poland, each of whom had performed over 2,000 colonoscopies. After equipping these doctors with an AI auxiliary operating system for a few months, researchers checked how they performed when the AI was turned off—specifically, whether their skills had dulled. The result? The rate at which they independently detected adenomas dropped from 28.4% in the past down to 22.4%. Adenomas, by the way, are a type of precancerous polyp. In pathology, it got even harsher: one experiment applied time pressure directly to pathologists and had the AI intentionally provide incorrect prompts. As a result, over 30% of the doctors changed their originally correct diagnoses to incorrect ones. Think about it: what is the value of a doctor in a diagnostic process? It's that moment when they spot a lesion that the machine didn't highlight—that single look. That look is the surprise; that is information. A doctor who merely confirms AI answers adds zero information to the diagnosis. If the AI is wrong, the doctor follows suit in being wrong. Naturally, there are similar studies for programmers. At the end of January this year, an experiment initiated by Australian researchers recruited 52 programmers to learn a new coding library. Half of them used AI assistance while the other half worked manually. After learning, they took an exam. The AI group scored an average of 50 points, whereas the manual group scored 67 points. Keep in mind this number was tested by a company that sells AI products itself. They also reviewed all the chat logs of the AI group, identifying six usage patterns broadly split into two categories: those who handed the entire task over to AI scored an average of under 40 on the exam, while those who asked for code along with explanations, or only inquired about concepts before coding manually, scored above 65. Why is that? Because mastering a skill relies on working through the problems yourself. Throughout that process, your brain builds a method. Next time you encounter a problem you've never seen, you can draw upon that method to figure it out—and what comes out of that is what truly surprises people. If you outsource the entire problem-solving process, the problem gets solved, but no method is built at all. The next time you turn off AI, you still won't be able to offer anything beyond the already well-known answer.

It is the same with the students. In March this year, the Higher Education Policy Institute in the UK published a questionnaire survey of 1054 full-time undergraduates. Among them, 94% of the students used generative AI to do graded assignments, and 12% directly pasted AI-written content into their assignments and submitted it. This proportion was 8% last year and 3% the year before, quadrupling in two years. Now that we have measured both sides, do we see the answer given by this data? It is that an individual's answers have become more standard, but the ability to generate wonder has degenerated. The essence of these two things is not contradictory, and one can even say they are two sides of the same coin. When you outsource the process of thinking about a problem, what you get in return is nothing more than a more standard answer. What is truly ruthless is actually the last one, and the one I most want to talk about: how much information is left when a group of people are put together? Before talking about this group, let me first tell an old story. I may have told this story before. In 1907, Galton published an article in Nature about an ox-weighing competition at a country fair. The competition asked 787 independent individuals to each guess the weight of an ox. Of course, most people could not guess accurately. But if you sort everyone's guesses from smallest to largest, the exact middle number is 1207 pounds, while the actual weight of the ox is 1198 pounds, a difference of less than 1%. Ah, this is the wisdom of the crowd. To put it the same way, the 787 independent guesses each contained wonder relative to one another. What everyone added in was information that others did not have, and the directions of their errors were not consistent. Ah, because some people would guess too high and some too low. Averaging them out canceled them out. Now let's modify this story. If these 787 people looked at the same weight estimate provided by AI before guessing, and then each made an upward or downward adjustment based on this number, everyone would still be wrong. But how would the direction of their errors be? They would follow the AI's estimate. At this point, what would happen to independence? It would disappear, because I can basically guess whatever your number is. Then the overall value would basically hover up and down around that value, which is the value estimated by AI. In other words, that mutual wonder among the 787 people is gone, meaning there is no information. Even if you find 10,000 people to average out, the shared part of that error is still there and simply cannot be canceled out. The more consistent the answers, the less information the group has. So that is what my opening sentence meant.

This modified story was actually tested on September 3 by Ron's team at the Massachusetts Institute of Technology, except that what they guessed was not an ox, but the capital market. They had 7 large models act as traders and ran 64 market scenarios, running each 5 times. The results were very unexpected. The stronger the models, the more similar their decisions were to one another, and whether they came from the same company didn't even matter, as long as they were strong enough. What they did was highly correlated. If you mix these models together, the error that can be eliminated through diversification is only half; the other half is this shared part between them, which no matter how many models you change cannot be eliminated. What the fucking most ruthless step was feeding fake news. The experimental team placed 8 scrolling comments on the market board, all saying that the reasonable price given by the market board was overestimated by 25% to 30%. When the market was full of AI traders, the degree to which prices deviated from the true value was 5 to 7 times usual. At that time, the model with the strongest capability was Claude 3.5, and what was its accuracy rate in judging ups and downs? Only 42%, lower than fucking flipping a coin. But there is another surprising thing here, which is that the slightly weaker models were not as easily fooled. The title of the paper is Why Better Models Make More Dangerous Systems. Ah, everyone can go look it up and find it, it's quite interesting. Let's come back to this: the opinions of these several models were highly consistent, but they were not divided into several pieces of information; there was only one piece, and what was that piece of information? It was that the answers of these several models actually turned out to be consistent, and this is what you call information.

So why does AI homogenize everyone's information, whereas earlier tools did not? Let me use a very famous philosophical example to explain this. In 1998, two philosophers, Andy Clark and David Chalmers—both working in the philosophy of mind and cognitive science—researched something quite fascinating: what exactly is the mind, and what is consciousness? Those kinds of questions. They proposed a concept called the "Extended Mind," which I think is very fitting here. They told the story of a man named Otto. Otto suffered from Alzheimer's disease and relied on a notebook he carried around to record things. For instance, if he wanted to go to a museum, he had to look up the address in his notebook—that sort of situation. What was their philosophical point? Otto's notebook was part of his memory, because the notebook performed the exact same function as memory does in a normal person's brain. But notice one characteristic of a notebook: Otto's notebook recorded things that Otto himself had encountered or wanted to do. What do our notebooks record? Things we have encountered, or things we want to do. What does each notebook store? Unique surprises, right? But a large language model is not a notebook. What it returns to you is not something you have already written down, but rather the phrasing that appears most frequently across the world's corpus. In other words, the answer AI gives is the mode—and by definition, the mode is the least surprising answer. OpenAI released a figure on March 31 this year stating that over 900 million people use ChatGPT every week. Nine hundred million people open the exact same dialogue box, backed by the exact same few models from the exact same company. According to the Extended Mind thesis, if it is part of your thinking process, then it is also part of my thinking process, right? Then your thinking process and mine share the exact same segment, don't they? In that segment, there is no surprise between you and me. That is the difference between a notebook and a public brain. A notebook allows each person's memory to have its own surprises; a public brain externalizes everyone's judgment to the very same place.

And this public brain has an extra layer of processing in CCP China. I remember either the year before last or the year before that, didn't the CCP come up with some bullshit "Interim Measures for the Management of Generative AI Services"? What do these measures require? They require training data to be compliant, and the generated content must reflect what kind of bullshit "core socialist values"? Put bluntly, it must reflect the fucking narrative of the CCP—the CCP's official line. Consequently, the CCP's models do not merely yield the mode; they yield a pre-censored mode that the CCP wants. The kind of fucking idiot press releases from the CCP I mentioned earlier already sifted out the least clear statements, and now this sieve is built directly into the models. Therefore, don't be fooled when China brags about releasing however many models; even if they fucking release a hundred million models, they all pass through the exact same sieve. A high number of models does not equate to a high volume of information. Even overseas, this kind of public brain homogenizes everyone's answers. The public brain inside the Great Firewall not only homogenizes everyone to that single output, but it also fucking mandates what that single output must be. Thus, if you are in China, reading the original source is far more valuable than anywhere else, because even the mode doesn't emerge naturally.

Let's talk about the capital market again. What does the capital market rely on for pricing? It relies on information. As soon as the price moves, surprises happen. But now, let's think about a question: when everyone enters the market holding the same answer, will the market leave visible traces? It certainly will. What is the first thing? The holdings of funds have started to look alike. In August this year, two researchers from the National University of Singapore and Sun Yat-sen University published a paper. What did the paper compare? It compared the correlation of changes in fund holdings that rely on generative AI before and after the release of ChatGPT. The conclusion is that they have become more similar, and in many dimensions, they are highly similar. For example, the stocks these funds are stuck in, and the speed at which prices react to news. What is the side effect? Instant liquidity has become worse. Why is the reaction fast? Because everyone received the same answer at the same time. Why is liquidity poor? Because when you want to sell, there is simply no one to take the other side; no one has a different judgment in their hands. Traces were also left in terms of speed. The International Monetary Fund has a department called the Monetary and Capital Markets Department, and the director of this department is named Tobias Adrian. A while ago, he said something very intense. He said that the flash crashes that happen in a few minutes in the future might no longer come from code errors, but from many AI systems reacting to the same information at the same time. What he changed is the lifespan of consensus, brothers! What was the paradigm of information dissemination in the past? It spread slowly, and it dissipated slowly. Now, everyone sees the same judgment at the same time. As long as the judgment converges, it is the moment when everyone capsizes together. That period of slow diffusion is simply gone. Have you noticed? Since this year, or maybe starting from last year, the market trend has been exactly like this. Look at the round of the Iran conflict. From the closing high on January 27 to the low on March 30, how much did the S&P 500 fall? It fell 9.1%. Then the panic index VIX surged directly to 35 during the session on March 9. And then? On April 15, the S&P 500 directly hit a new historical high, only 16 days away from the low. As for the Nasdaq Composite Index, it retraced from the closing high on June 2 to the low on July 29, falling 9.8%. Two months later, the Nasdaq hit a new historical high again. Chip stocks were even more extreme. The Philadelphia Semiconductor Index rose 88% in the second quarter of this year, which is the best quarter since records began in 1994. Then it fell 20.6% in the single month of July. These rounds of rises and falls each have their own news, but what I want to say is the rhythm: buying is done together, and selling is also done together, with no process of divergence in between. This kind of consensus has a characteristic: it looks solid, but it is actually very fragile, because what supports this price is not the information in everyone's hands, but the same answer. And this person reading less and less is also connected to it.

ChatGPT compiled data on global readers in the second quarter of this year and found what? It found that global readers spent an average of only 26 seconds on an article. There is something even more fucking impressive: in the first 4 months of this year, 68% of Google searches in the United States did not click on any links. In 2024, this proportion was still 60%. The less original text you read, the fewer surprises you will encounter, and at the same time, you will rely more on that shared judgment. As soon as that shared judgment capsizes, everyone capsizes together. Not only that, even the language of the financial market is becoming increasingly homogeneous. The international organization led by the G20, called the Financial Stability Board, issued a consultation paper on June 10 this year. It was written very bluntly: relying on common AI models, data integration, or infrastructure will cause financial institutions to make similar moves, amplifying herd effects and pro-cyclicality. The herd effect is everyone buying and selling together, and pro-cyclicality is what? It is fucking adding positions when it rises, and fucking reducing positions together when it falls. The Bank of England can't sit still either. Their deputy governor directly pointed out the issue at the European Central Bank forum. He said that if AI programs that can place orders by themselves react to the same trigger in the same way, the market should have a master switch similar to a market-wide circuit breaker. You see, regulators have already started to worry about the problem of instant liquidity explosions. This market-wide circuit breaker is just a proposal, and there are no implementation details. But just two days ago, on September 30, the Bank of England released the minutes of the Financial Policy Committee meeting on September 25. The record cited an estimate from Morgan Stanley. What is this estimate? It is that by September 2026, global AI-related bond issuance will be approximately $450 billion, which is already more than double the entire year of 2025. So there is a question: was this money borrowed based on the same answer? You figure it out.

Now let's return to the perspective of ordinary people. Since what is lacking in the AI era is not answers, but wonder, what is the solution? Turn yourself back into a person capable of generating wonder. We should treat AI like Otto's notebook, not as a collective brain. Sure, let it help you store, search, and organize things. Those are jobs a notebook is supposed to do, and it certainly does them better than humans. But whatever you do, do not outsource the step of judgment. Judgment must rely on your own information. And where does your information come from? From primary sources, original works, books, the places you have walked yourself, and things the models haven't read. Wonder only exists in raw materials; there is no wonder in the statistical mode. When you read a primary text, your judgment separates from the judgments of 900 million people. Moreover, you need to practice a form of training. What training? Have AI find counterexamples for you, instead of letting it draw conclusions for you. If you ask it how a certain stock is doing, it will definitely give you the global mode, and at that moment your judgment is merged into the judgment of 900 million people. But if you ask it where your judgment falls apart, or what the opponent's strongest argument is, it will definitely give you a reason you hadn't thought of. Only then is it providing you with wonder; only then is it providing you with actual information. Even with the same 900 million users, how you use it determines whether you get merged into the crowd or stand apart from it.

There is another small habit, which comes from experimental data published by the University of Munich on September 17. In the experiment, 704 people practiced fraction calculations with an AI assistant; after practicing, they turned off the AI and took a test. For one of the groups, after completing each problem, the screen would remind them: did you calculate this problem yourself or did you ask the AI directly for the answer? Then it asked them: to what extent do you plan to have AI help you on the next problem? Just this single action reduced the probability of directly copying AI answers to less than half. Test scores were also slightly better, though this specific point was not yet statistically significant. Another group was given high scores if they solved it correctly on their own, and low scores if they got it right after asking for the answer. This group showed no effect whatsoever. That is to say, before having AI give you the answer, write down your own judgment first, and then look at its response to see where the gap lies between your judgment and its answer. That gap represents the amount of information you hold. If the gap is zero, it means this time you didn't gain information—only an answer. The final point is also crucial: keep the option not to listen to it. This isn't about showing off; it's because in this era, independence is worth more than being correct. An independent judgment that isn't necessarily better than everyone else's still adds a piece of information to the group that others do not possess. Because ten thousand copied judgments add zero information. Returning to the topic itself: are people in the AI era becoming stupider or smarter? My conclusion is very clear: answers have been democratized, but information has not. Crowds will become stupider, but individuals who make good use of AI will become smarter. The reason is simple: AI standardizes an individual's answers, but unifies the information of the crowd into a monolith. Therefore, what we need to do is not take sides between these two camps, but prevent ourselves from becoming just one identical copy among 900 million. Well, that's all I'm sharing with you today. If you enjoy my show, please hit the bell icon below. If you want to support me, feel free to join the membership channel. The membership channel offers four in-depth episodes every month. Welcome to join the gang of badasses. See you next time.