What this covers

Gerd Gigerenzer, a psychologist from the Max Planck Institute, joins Russ Roberts on EconTalk to argue that the real danger from artificial intelligence lies not in machines becoming superintelligent but in human passivity and the motives of those who control AI systems. Gigerenzer challenges the assumption that algorithms are inherently smarter than people, pointing to concrete failures—IBM's Watson performed well at Jeopardy! but proved dangerous in cancer diagnosis, ultimately matching only a first-year medical student's capability. His core claim is that machines excel in stable, well-defined environments but fail under real-world uncertainty, where human judgment remains essential. The conversation traces how surveillance business models, rather than AI itself, manipulate users, and why the solution is not accepting algorithmic authority but making people smarter through critical thinking.

The discussion ranges across how deep neural networks work as correlation machines lacking common sense, causal reasoning, and the ability to navigate uncertainty. Gigerenzer demonstrates this through examples: doctors unable to interpret false positives, investors outperformed by simple division-of-assets rules, and crime-prediction systems that gain no advantage from complexity over basic variables. He presents the Privacy Paradox—people claim to value data privacy yet pay nothing to protect it—and examines why education has largely failed, with statistics taught as cookbook formulas rather than Statistical Thinking applied to real problems. The speakers also address how few people, mostly at a handful of tech companies, shape attention and emotion for billions, and how rising support for Social Credit Systems signals a troubling shift in values. Throughout, Gigerenzer emphasizes that teaching lateral reading, click discipline, and genuine critical thinking are the only antidotes to manipulation—though he acknowledges the difficult truth that people often prefer comfort to truth.

Sharpest takeaway

Gigerenzer argues that AI excels only in stable, well-defined worlds and fails under real-world uncertainty, so the real danger is not superintelligence but human passivity—and the solution is making people smarter and more critical rather than trusting algorithms.

  • Deep neural networks are powerful correlation machines but lack common sense, causal reasoning, and the ability to handle uncertainty
  • Surveillance business models and the Privacy Paradox mean people are manipulated by those behind AI, not AI itself
  • Education in statistical thinking and digital literacy is the antidote, but current schooling teaches mindless 'cookbook' methods

The argument · threads7 threads · 37 claims
0.85

Algorithms lack common sense and perform well only in stable domains; claims of superiority in uncertain human contexts are overstated.

6 pointscentrality 5/5
  • There is a general principle that if the world is stable—like a well-defined game—algorithms will most likely beat humans and perform much better, but if there is lots of uncertainty, as in cancer treatment or investment, then algorithmic claims are probably overstated and one must be very cautious.

    “If the world is stable, like a game, then algorithms will most likely beat us, perform much better. But, if it's lots of uncertainty as in cancer treatment or investment, then you need to be very cautious.”

  • Deep neural networks are statistical machines that excel at finding correlations—essentially a powerful version of non-linear multiple regression or discriminant analysis—and finding correlations is not the greatest ability of the human mind, which instead is strong in causal stories.

    “A deep neural network is about correlations and it's a powerful version of a non-linear multiple regression, or a discriminant analysis. Nobody has ever talked about multiple regressions as intelligence.”

  • Artificial intelligence as instantiated in deep neural networks lacks common sense—it cannot easily learn concepts the way a child can (a child grasps 'school bus' in one or two pointings, while a network needs 10,000 images), nor does it possess intuitive psychology or intuitive physics, making common sense a major unsolved challenge in AI.

    “A deep neural network, in order to learn, to distinguish pictures of, say school buses, from other objects on the street needs 10,000 pictures of school buses in order to learn that. If you have a four-year-old and point to a school bus, you may have to point another time, and then the kid has gotten it.”

  • Predicting whether someone will commit another crime is a high-uncertainty situation where algorithms are actually not good; recidivism predictions can be made by simple heuristics using just two or three variables—previous offenses, age, and maybe gender—as well as the secret COMPAS algorithm, demonstrating that complex algorithms offer no real advantage in such uncertain domains.

    “we know that recidivism predictions can be done by just simple heuristics with two or three variables as good as the secret COMPAS or other algorithms. The two or three variables are: Previous Offenses and Age. Maybe a third one, Gender.”

  • IBM's Watson, after its Jeopardy! success, was pivoted to healthcare not because it knew anything about healthcare but because of the money; the Watson Oncology cancer-treatment product produced some recommendations that were dangerously deadly, and IBM later clarified that Watson performed at the level of a first-year medical student.

    “Watson Oncology was the first thing for cancer treatment, only to find out that some of the recommendations were dangerously deadly. And then, IBM clarified that Watson is at the level of a first-year medical student.”

  • Both AI enthusiasts dreaming of uploaded brains and critics like Zuboff assume algorithms would be perfect, which is only true in stable worlds like astronomy; in reality algorithms predicting and controlling human behavior are not very good, yet we still submit to their recommendations—on YouTube some 75% of videos watched are no longer chosen by viewers but recommended or auto-played by an algorithm.

    “people today on YouTube, some 75% of all videos being seen are no longer chosen by the viewers. They are recommended by an algorithm or just played.”

0.80

Surveillance capitalism exploits user data through hidden mechanisms that people neither understand nor consent to.

5 pointscentrality 5/5
  • People are sleepwalking into surveillance: in the German studies, 85% are not aware that a smart TV may record every personal conversation held in front of it, whether in the living room or bedroom, even though this can be found buried in end-user notes that almost nobody reads.

    “most people are not aware that a smart TV may record every personal conversation people have in front of it, whether it's in the living room or in the bedroom. At least in the German data, 85% are not aware about that.”

  • Surveillance toys like Hello Barbie, which records a child's hopes, fears and anxieties and sends them to third parties for advertising analysis, risk a deeper consequence than lost trust: the child may not lose trust but instead come to believe that being secretly surveilled is simply how life is, so that algorithmic surveillance gradually changes our own values and produces a new generation no longer truly concerned about privacy.

    “the little girl may not lose trust. The little girl may think that being surveilled, even secretly, that's how life is. And so, here is another dimension that the potential of algorithms for surveillance changes our own values.”

  • Because Google gets about 80% and Facebook about 97% of revenue from advertising, the user is no longer the customer; like a free coffee house that bankrupts paying competitors while secretly recording everyone and letting salespeople (the real customers) push personalized products, the users' time and attention are the product being sold.

    “Google gets 80% of the revenue from advertisement. Facebook, 97%. And, that makes the customer--the user--no longer the customer.”

  • People underestimate how closely tech companies are interrelated with governments; the reassurance that 'Zuckerberg knows but the government doesn't' is false, as Snowden revealed how close the connection is in the United States, the United Kingdom has Karma Police, and similar arrangements exist in many countries—making surveillance data a potential tool for authoritarian control like the East German Stasi or China.

    “what people underestimate is how closely tech companies are interrelated with governments... Snowden has, a few years ago, shown how close the connection in the United States is. In the United Kingdom, there's Karma Police.”

  • Two kinds of personal data must be distinguished: harmless information like which books you buy (used to recommend other books) versus collecting all information including whether you are depressed, pregnant, or have had a heart failure or cancer, and using that to target you at the right moment with the right advertisement—the latter is what we do not need and should be removed.

    “The other thing is collect all the information about you that one came including whether you're depressed today, whether you are pregnant, whether you have had a heart failure or have cancer, and use that information to target you in the right moment with the right advertisement. So, that's the part that we do not need.”

0.66

People cannot evaluate information sources or understand data claims, leaving them vulnerable to manipulation and deception.

5 pointscentrality 4/5
  • The most recent international PISA study testing 15-year-olds across OECD countries, which included a component on digital understanding, found that 90% of these so-called digital natives do not know how to tell facts from fakes, revealing a major blind spot that governments fail to address by spending billions on hardware while spending almost nothing on making teachers and pupils smart.

    “90% of 15-year-olds, that [?] the digital natives, do not know how to tell facts from fakes. So, there is something to be done. Governments spend billions for tablets and whiteboards for technology in schools. They spend almost nothing on making teachers smart and pupils smart.”

  • The average doctor does not understand what sensitivity, positive predictive value, or a false positive mean, and as a result cannot critically read an article in their own field to evaluate questions like how good cancer screening is; even among lung cancer specialists, only a few understand these concepts while the majority do not.

    “the average doctor does not understand what a sensitivity, what a positive predictive value is, or a false positive. Ask your own doctor. And so, then they cannot read an article in their own field to evaluate: how good is cancer screening?”

  • Introducing statistics into the high school curriculum has largely failed because it is taught as cookbook calculation of means, medians and standard deviations that can be tested on exams, while the subtle, deep, common-sense ideas of how to think about an unpredictable world have no curriculum; the right approach is to teach statistical thinking—not probability theory—through eye-opening real examples like evaluating a positive COVID test using hit rates and false alarm rates.

    “It's mostly horrible. It's cookbook, teaching people how to calculate means and medians--things you can test on an exam: what's the standard deviation? The subtle, deep, common-sense ideas of how to think about the fact that the world is unpredictable, there's not a curriculum for that.”

  • A dating agency advertises that 'every 11 minutes a single falls in love,' which a simple calculation deflates: at six per hour and 144 per day that is roughly 50,000 per year, so for an agency with one million customers only about 5% find love annually—meaning a paying customer faces roughly a 50/50 chance only after ten years, a fact actual studies confirm, illustrating that people are impressed by numbers they never bother to think through.

    “every 11 minutes, a single falls in love... In an hour, it's about six who fall in love. On a day, it's 144. And in a year, it's roughly 50,000. If the online agency has 1 million customers, then in a year, 5%--50,000--fall in love.”

  • A Stanford education study found that 97% of undergraduates do not know how to evaluate a website's trustworthiness; the correct technique—called lateral reading—is to read a little, immediately go to 'About Us', leave the page, and find out who is behind it, whereas most young people still read a website from beginning to end as in pre-digital times and judge trust by whether it looks cool.

    “what one should do is read a little bit, and then immediately go into About Us, and then leave the page, and find out who is behind it. Ninety seven percent of these young people in Stanford don't know that. And that's called lateral reading.”

0.66

Teaching critical thinking and statistical reasoning, not just technology or test preparation, is essential to restore democratic autonomy.

5 pointscentrality 4/5
  • The situation is not destiny and can be changed by investing not only in hardware but in making young people smart—teaching lateral reading and 'click discipline' (not clicking the first entry, and why)—and by addressing the deep business-model reasons behind disinformation rather than repeatedly fining Facebook, because maintaining a democracy only works with people who critically think and want a life of self-determination and judgment.

    “if we invest not only in tablets and whiteboards, but in making young people smart, open their eyes--that can be done. And we know how it's being done: Teach them lateral reading; teach them click discipline--don't click at the first entry, and why.”

  • It is not a trade-off between teaching thinking and exam performance: Finland (and Japan, by very different routes) consistently rank at the top for 15-year-olds, yet the Finnish system is not geared to exams—demonstrating that schools focused on thinking rather than test preparation can still do well on the very measures exam-focused systems chase.

    “It's not a school that's geared--the Finnish system not geared to exams. On the opposite. But they do well on the exams. So, it's not a trade-off in a sense”

  • Mindless math gives a formula and exercises to plug parameters into; good math instead gives pupils a problem they have no solution for, letting them make errors and discuss ways to solve it, which is what thinking is and creates a positive error culture—because any intelligent system makes errors, while large corporations suffer from negative error cultures where nobody dares admit errors or avoids decisions to avoid errors, and we pay dearly for that.

    “good math would be to give pupils a problem they have no solution for. Give them a problem and let them make errors and find a way and discuss them among other to solve a problem. So, that is thinking. And, that also creates a different error culture that one understands making errors is a positive thing.”

  • Democracies have recently elected a number of presidents and leaders—such as Boris Johnson and Donald Trump—at least partly because of their entertainment value, suggesting that valuing understanding of the world and control over oneself is fading and must be steered against through education.

    “we have had in democracies now quite a number of presidents who seem to have been elected because of their entertainment value. So, Boris Johnson is just leaving at this moment. Donald Trump already left, but may come back.”

  • Already children understand intuitive physics, such as the fact that an object that disappears behind a screen is not gone, a form of common-sense knowledge that is very difficult to get into a neural network.

    “Already, children understand that an object that disappears behind a screen is not gone. How does a neural network know that? It's very difficult.”

0.56

Young people risk internalizing surveillance as normal, eroding privacy values and enabling authoritarian control systems.

3 pointscentrality 4/5
  • The Chinese Social Credit System, which assigns every citizen a single FICO-like score covering everything measurable including political and social behavior, shares the same spirit as an educational system geared to a final number—both reduce a person to a scalar that purportedly tells you instantly whom to trust, and Westerners' revulsion at the Chinese system should extend to the score-fixated logic they already accept.

    “The Chinese Social Credit System... is one way to go towards, basically, total control of the citizens in terms that every citizen has a Social Credit Score. It's like a FICO score, but now for everything that can be measured, including your political and social”

  • Support for a Social Credit System in Germany is increasing—from 10% in 2018 to 20% in 2022—and is notably higher among the young (28%) and strikingly highest among lifelong government civil servants (37%), who likely assume they are on the right side and will collect goodies, indicating a worrying drift in values even within democracies.

    “the number of people who think a Social Credit System would be a good idea in Germany is increasing. In 2018, it was 10%. And now, 2022, it is 20%... Among the young, it's 28%... 37% of them think it would be a good idea.”

  • China's Social Credit System creates a world where citizens are surveilled, predicted and controlled, with 'good guys' (defined by the government) getting goodies—such as being treated first in hospitals or earning points for visiting elderly parents—while those with low scores wait or are punished, with over 10,000 Chinese in recent years barred from buying plane tickets and points lost for searching terms like 'Dalai Lama.'

    “in China, in a hospital, you're treated first if you have a high score. Those with a lower score have to wait... there were 10,000 in the last few years of Chinese who were not allowed to purchase plane tickets because of low scores.”

0.56

Power over human behavior is concentrated in hands of wealthy tech executives whose motives are profit-driven, not public welfare.

3 pointscentrality 4/5
  • The real worry is not whether AI is omniscient or merely an assistant tool, but the people behind AI and their motives—today a few extremely rich, mostly young white males at a handful of tech companies shape the emotions, values, and time of almost everyone else.

    “it's not about AI by itself. It's about the people behind AI and their motives... a few tech companies, and mostly a few relatively young white males who are immensely rich, shape the emotions, the values, and also control the time of almost everyone else.”

  • If Facebook (Meta) were reimbursed for its entire per-user revenue so it no longer needed to sell data, it would cost roughly $2 per person per month to restore privacy and freedom—a trivial sum equivalent to a coffee—so the technical solution is affordable; the real difficulty is getting there given users' unwillingness to pay.

    “If you take the Facebook--now Meta Corporation--and would reimburse... Zuckerberg for his entire revenue, it would be about, per month, it would be about $2 per month. Per person. That's all.”

  • People exhibit a Privacy Paradox: surveys show their greatest concern about digital life is not knowing where their data goes and what is done with it, yet in three representative German surveys since 2018, 75% said they would pay nothing—not a single Euro—to keep their data, contradicting the economic prediction that they would pay for what they claim to value.

    “Seventy-five percent--75%--of Germans said 'Nothing.' Not a single Euro... if you have that situation--where people say, 'Oh, my greatest worry is about my data'; at the same time, 'No, I'm not paying anything for that,' then that's called the Privacy Paradox.”

0.43

Simple heuristics and statistical thinking often outperform complex mathematical models and big data approaches in real uncertain environments.

3 pointscentrality 3/5
  • Harry Markowitz won a Nobel Prize for an optimization model telling investors how to diversify across assets, yet when investing his own retirement money he did not use his optimization method but instead used the simple 1/N heuristic—dividing money equally across N assets—and studies show this heuristic often outperforms Markowitz optimization including modern Bayesian methods in the real, uncertain world of finance.

    “when he himself invested his own money, for the time after retirement, he used his Nobel Prize winning optimization method? No, he didn't. He used a simple heuristic. A heuristic is a rule of thumb, and that's called: invest equally. It's called one over N.”

  • Big Data analysis presumes the past will tell us what the future will be like, which works only when the environment is stable enough that past patterns persist; in most human environments they don't, so when the future is not like the past Big Data doesn't help—and notably the 1/N heuristic needs no past data at all, the opposite of Big Data.

    “particularly, if the future is not like the past, then Big Data doesn't help you. And for a finance with Markowitz optimization, you need lots of data to estimate all these parameters. The heuristic, 1 over N, needs no data on the past. It's the opposite of Big Data.”

  • Experiments by Californian economist Steve Tadelis on eBay found that a certain kind of brand keyword advertising actually produced a negative profit for eBay rather than the promised return (e.g. $12 for every $1 spent), suggesting companies should run their own experiments to verify whether personalized advertising investments actually pay—an interest aligned with creating a healthier internet.

    “There are these fascinating studies with eBay by a Californian economist, Steve Tadelis, who just ran experiments and found out that a certain kind of the keyword brand advertising actually was a negative profit for eBay and not the promised $12 for every $1”