How to Kill a Myth

Preliminary remark: I am using the abbreviation AI in this article as it is currently used in the press and the cited economic report to denote the vision of a technology that may solve all humanity's problems, currently tried to be achieved by a technology that I would rather call uninspired statistics on steroids, fueled by disregard for privacy and copyright laws. To denote the scientific field, I use the unabbreviated term Artificial Intelligence.

Breaking Point of a Bubble

The Annual Economic Report 2026 of the Bank for International Settlements leaves no doubt about the economic bubble created by AI:

[T]he optimism surrounding AI may not last, despite its promise of future productivity gains. The current surge in capital expenditure could prove unsustainable if supply bottlenecks restrain production. Intense competition for market leadership may fuel overinvestment further, as seen in previous innovation waves, increasing the risk of a sharp reversal if AI payoffs disappoint. [p.2]

The report gives ample evidence of the enormous debt that has flown into AI start-ups, namely OpenAI and Anthropic, directly or via hyperscalers, and the impossibility of generating gains, even if AI were to deliver technologically (Graph 11C on page 23). Any investment banker who is still pumping money into these companies or the AI hype industry should be held personally responsible when the bubble bursts and the general public will again have to pay for the greed of managers, backed by the inanity of politicians.

The economic bubble has been described in detail elsewhere (e.g. by Ed Zitron). And since it cannot be maintained indefinitely, all we can do is hope for it to end soon. This end will come with a bang that we cannot miss, as all of us (i.e. those who were not fortunate enough to secure their third yachts or purchase whole neighborhoods in Beverly Hills) will pay the bill.

But what about the other aspects of the hype, such as the narrative of a technological breakthrough or the social malfunctionings that the AI hype has uncovered or created? Will they disappear with the burst of the economic bubble?

History of the Tech Hype

Let us recap how we got into this mess. The technology hype came upon us in several waves, intensifying at each step, in conjunction with economic exuberance.

Wave 1: Cognitive Systems

When I started to work in the field, the term Artificial Intelligence was hardly used. Outside the scientific field, nobody cared about it as it had disappointed in previous hypes, and inside it many people were displeased with its fragmentation and lack of progress. In the early 2000's, several long-established scientists (such as Marvin Minsky, Pat Langley, and many others) started initiatives to get the field back to its original purpose of understanding intelligence—renaming it to cognition—in terms of systems, not single algorithms (after all, those would be the concern of Computer Science in general). At that time, new conference series were established, such as Advances in Cognitive Systems. Another one was called Artificial General Intelligence. This one turned out to have good public relations.

Into this mindset also falls the IBM Watson system. Their remarkable demonstration of winning a round of the quiz show Jeopardy against two human champions temporally overlapped with the second wave of the hype (it was in 2011). But their technology followed the initial trend in the field to take up the bits and pieces of the AI subfields and rejoin them into a complete cognitive system. Their Jeopardy success and their subsequent product fell out of any AI hype cycle. Instead of using older terms such as AI or expert system, they coined the term question answering. In contrast to currently popular Large Language Models, which are based on statistics alone, IBM used an uprecedented combination of all the technologies that were around. They made a genuine effort to build a system that would provide reliable answers. It turned out that the existing technologies, even if well-orchestrated, were inadequate to make it work.

In a way, IBM Watson befell the same fate as the whole scientific field of Artificial Intelligence. Their engineering efforts were overtaken by even more excessive computing power and uninspired statistics. I had little sympathy for IBM's aggressive marketing at the time. While clearly the best system around, it disappointed in real-world applications. But this was nothing to the marketing machinery we have been witnessing in the last years. IBM invested its own money and left the economy at large in peace. I also give them credit for their scientific contribution. They demonstrated beyond doubt that the algorithms we have are not adequate to simulate intelligence or get anything done beyond standard software engineering (and that's not really a surprise, after all, it runs on the same hardware). If we lived in a world where scientific methodology were applied and people were accepting empirical evidence, we could use these insights and rethink the whole endeavor of Artificial Intelligence. Alas, we do not live in such a world. We live in a world where everyone is waiting for some magical technological revolution to happen.

The cognitive systems effort stayed unnoticed by the general public until the DARPA Grand Challenge of 2005. This 212 km (132 mi) race of autonomous cars through the Mojave Desert brought sudden fame to the leader of the winning Stanford team, Sebastian Thrun. He convinced Google that self-driving cars were ready to be turned into consumer products (obviously omitting a few technical details about the race such as the route being defined by a dense sequence of GPS points, and that the technology had been around since at least a decade or two without any groundbreaking innovations). He was also the one to start the Google Glass project.

I remember feeling completely stunned at that time. Self-driving cars were suddenly taken as guaranteed to appear in the next few years. Everybody started to get ready for their smart glasses. As an expert in the field, I saw no technological progress beyond the normal paper-maximizing race of optimizing parameters in algorithms that had been around for decades. Other scientists were just as surprised as myself, but outside of private conversations, nobody took the trouble to calm down the press. After all, a new AI hype would secure research money. It felt as if the field just chose to pretend that suddently the technologies that had disappointed in the past magically had started to work.

Wave 2: Cheating with Computing Power

Those initiatives at Google eventually died or were spinned off into other companies to avoid any association with Google if their failures were openly recognized. But they never were, as they were covered up by the next wave of AI enthusiasm (around 2011): Deep Learning. Again the technology was not exactly new, but with Google having accumulated such a heap of image data, it could simulate a breakthrough in image classification (not understanding!). Up to then the computation of statistical models (aka Machine Learning, it has nothing to do with learning in a human sense) was a brittle exercise in which the amount of data had to fit the size of the model (i.e. the number of parameters that are optimized in the training process). The pretended technological breakthrough was basically a refined way of overfitting. If you have any possible perspective, type and lighting of a cat in your database, you don't have to generalize, you can just look up the closest-fitting picture and see if it had been labeled as a cat.

Their trick was to replace thorough engineering and inventiveness with brute force computing power. Not even this trick was new, just applied on a larger scale. The isolated subfields of Artificial Intelligence had pretended to make progress over the years, while just piggybacking on the increase of computing power. You could basically publish the same paper every year, solving artificially created problem instances faster or more accurately than the year before. Since Computer Science in general, and Artificial Intelligence in particular, has never embraced the idea of thorough empirical evaluation, nobody asked for a comparison of last year's algorithm to this year's on the same machine. And if in doubt you could always tweak some parameters to make the algorithm look different or solve some other problem instances than the year before (and there was plenty of material from benchmarks, such as the tasks of the annual Planning Competition or image classification databases).

Image classification impressed only for a while. Its applications are limited to image search on the web, recognizing plants via an app or labeling images for accessibility purposes or archiving, possibly some medical applications, which had, however, been around before. It was hard to turn that into a story of intelligence surpassing human abilities. So the deep learning wave was redirected to an old favorite of AI: game playing. The story goes something like this: In order to play chess, you must be intelligent. Therefore, if a computer were good at chess (possibly better than a human player), the computer would be intelligent. Since Go is considered an even more difficult game, and it took even longer to win against the human champion, a Go-playing program would have to be even more intelligent.

Let's first neutralize this nonsense. Alan Turing had proposed in 1950 [1] to use chess as a starting point for developing intelligent systems:

We may hope that machines will eventually compete with men in all purely intellectual fields. But which are the best ones to start with? Even this is a difficult decision. Many people think that a very abstract activity, like the playing of chess, would be best.

He never claimed that a chess-playing machine would be intelligent. It was merely a testbed, something specific to work on. What Turing considered a first step took almost half a century to achieve until in 1997 (again outside of any AI hype cycle), IBM's Deep Blue supercomputer beat the world champion. It is important to note that this was not due to a breakthrough in algorithms, but to computing power. Given that hardware, Turing himself could have made it work.

Just by simple observation, anybody could have understood that playing chess does not imply intelligence, otherwise the quest for magic AI would have ended in 1997. It hasn't, because game playing is a very simple task for a computer. Games are defined by strict rules that can easily be formalized as mathematical statements. A computer is a calculation engine, it works on math. So anything you can easily put into math is an easy task for a computer. A computer playing chess well is about as surprising as a washing machine cleaning clothes. This is what it is built for. A computer that cannot do math is not a computer. The only reason why chess had not been solved before, was that you need a lot of computation for this particular math exercise. The problem solved itself by better hardware plus the concentration of money.

But myths cannot be killed by reason or experience. So Google announced its AlphaGo program as the next big breaktrough when it won against the human world champion. The only problem was that Go was gone as a benchmark. And even Google couldn't hide the fact that a Go-playing machine is about as useful as an icecream-eating machine. If you enjoy playing Go (or chess), you don't need a machine to do it for you.

What was needed was a story that Go-playing could be turned into general intelligence. So the program that played Go was generalized to play other games. Same technology, different data, extended marketing. The research community jumped on board and every day you could read about the bright future of intelligent machines, now that they can play (and win) all sorts of games.

Eight years later, we can safely say that no game-playing program magically developed superhuman intelligence, or any intelligence whatsoever. And they don't have to, since we now expect text generators to turn intelligent.

Wave 3: Going Crazy

Language processing is another classical AI application. It is so intricate that it had split into a field of its own. You would not find papers about language processing in general conferences on Artificial Intelligence in the 1990's or 2000's (the same goes for image processing, by the way). It was more associated with linguistics, being heavily based on the work of Chomsky, who happened to have offered a mathematical model of language, though not very convincing to explain human communication. Even after more than 50 years the field had failed to realize that humans neither think nor converse in formal logic, and thus never delivered anything useful.

Continuing the proven approach of throwing statistics in enormous scale on data of even more enormous scale, burning immense computational resources and making up any shortcomings by human intervention and promises of future improvements, we come to the wave of Large Language Models. Statistical language models, as an instance of Markov Chains, had been around for a long time. But up to then, thinking in reasonable terms of data availablility and computation, you would only use groups of two or three words to predict the next. Now with the insolence of using the whole World Wide Web as a data source and not putting any restrictions on computing power (being financed by tech companies and investors that were desperately looking for the next big thing), it was possible to create long passages of text that were grammatically correct (even in inflected languages such as German) and somehow made sense.

The rest of the story is still fresh in the news. Text generation is supposed to replace lawyers, teachers, writers—just about any intellectual profession. Luckily, programming is basically the same as writing, instead of words, you chain commands. In fact, it must be easier, because you don't need that many words, only the few commands allowed in any programming language. So we fire the engineers first. And once we use text generation technology on robots, they will replace any kind of manual work.

Economy and Tech Lifting Off Together

I have stopped to work myself up when I read such nonsense. Partly because I am counting on the economic bubble to burst soon. After all, the economic situation only made the technology hype possible. Without investors and hyperscalers sitting on piles of money, running out of ideas of how to turn them into newsworthy innovations, the hype would have subsided after the first wave. The alleged advances had nothing to do with technological invention. The only novelty was that the Silicon Valey tech bros had lost any qualms regarding the world economy, environment and social norms (such as privacy or copyright).

When the economic bubble bursts, there will be no more money to fuel any new wave of the hype. There will also be no significant new data sources to drive any more statistics. Compared to the text corpus of the World Wide Web, even if you collected data from all sensors built into any machines available, you couldn't fake any visible success.

So we might hope that the hype would end there. Maybe we could even go back to solving problems that we actually have.

I wish I could believe it. But I conjecture that the story of AI being just around the corner is not going to subside with the economic breakdown. On the contrary. Once the public gets officially invited to rescue some more banks, lose their jobs and tighten their belts for the upcoming years, people will be looking for answers outside reality. They will turn to religion. And AI is a religion.

I once gave a presentation about robotics in eldercare. As usual, I presented the harsh reality, unveiling the actual (in)capacities of robots, clearly showing that robots are not going to make any significant contribution to eldercare anytime soon. Hearing this, one of the listeners got rather angry. It was not so much that she didn't believe me, but she had no Plan B: If we cannot solve our issues in healthcare by robotics, we are doomed.

We live in a complicated world with lots of big and small problems to solve, lots of decisions to make. It is understandable that people are looking for an easy way out. Technology is going to solve it all. Maybe not today, but the Tech Messiah will be here on time to solve all our problems.

We are polluting the planet by heavy use of technology. Let's add some more technology and hope that at some magical breaking point it will stop polluting and start fixing the issue. Our healthcare system (in Germany, probably also elsewhere) is a mess. Instead of taking the trouble to consider workable solutions, let's digitize everything (with every single project failing) and hope for inventions out of thin air to replace care work (without replacing humans, for sure!).

LLMs are adding even more pollution, make lonely people even more lonely, create misinformation as an inbuilt feature. But sure, at some point, they will magically turn into problem-solving machines rather than problem-creating ones. How?

The more I wonder why this mismatch is generally overlooked, the more I understand that I am asking the wrong question. People want to belief in AI or any other technology, exactly to avoid the trouble of answering such questions. The evidence is in front of everybody's nose. 20 years of unfulfilled promises and lots of new problems should at least raise some doubts about the ever more fantastical claims emerging from Silicon Valley.

When Hernán Cortés landed in America, the Aztec leaders (allegedly) considered him to be the god Quetzalcoatl. That empire had quite a few problems, they were waiting for salvation. So they were ready to accept this ruthless and malevolent paleface. While shaking our heads over history books, wondering how anyone could be so stupid, history repeats with AI. The world has been waiting for the Tech Messiah for about a century.

The Math Myth

I am neither a historian nor philosopher, but my impression is that some time around the start of the 20th century (or maybe even before), the success of mathematical formalisms in science (in fact, Physics, most other sciences just copy and pretend in order to look like real science) has led to a low regard for human capabilities. Anything you cannot express in math is regarded as subjective, irrational, worthless.

Industrialization may also have played a part. In factories optimization (as the mathematical version of doing your best) works pretty well. And you can measure all sorts of things, like the number of parts that a machine has produced in an hour. Or the number of parts a worker has assembled. Measures of productivity have not only been transfered from machines to humans inside the factory, but far beyond. From measuring produced parts, we now measure the number of applications read by HR, or the lines of code written by engineers, or the number of papers published by scientists. Outside of factories (and I'm not even sure about the inside) most metrics are complete nonsense. All they do is demotivate people by absurd incentives. If you are constantly punished for good work, just because it happens to be uncountable (which is the typical case), you will either fall into depression or give in to the metrics game, with your paycheck becoming your only reason to get out of bed.

Now that we measure everything, we can also optimize everything. Especially humans. We optimize our work output, our CVs, our work-life balance, our eating and sleeping habits. We are turning ourselves into machines. So why not just replace us underperforming humans by machines?

Ridiculous as this may sound, it is exactly the expectation that I found in an event about autonomous driving. The first speakers were talking about Law and Ethics of autonomous cars. It was clear that the they assumed an autonomous car to be like a human, just better—never drunk, always alert. I spent my talk on the technology side to make it clear that an autonomous car is more like Excel, just more complex and thus less reliable.

People forget, or simply do not think about, the limitations of math. They admire its precision and fall for the myth that math can never be wrong. Of course it can be wrong! First of all, any math is an extreme oversimplifiaction of reality. So if you set up the wrong formulae, your result will be useless. And even if you have formulae that model your real-life problem well, there is never any full certainty that any proof is correct. If 100 mathematicians check a proof, it can still contain logical errors that all of them have accepted, simply because they are humans and humans make assumptions. You can also use an automatic proof checker. But that proof checker has been programmed by humans and they can have put in any kind of logical mistakes. You can use 100 proof checkers, all being programmed by humans who can have made the same mistakes. And you can ask an LLM to generate a proof checker and it will take all those 100 faulty human-made programs and turn them into one with even more bugs.

While the proof issue may be purely philosophical, the modeling one is not. Try to express what you had for dinner yesterday in mathematical formulae. Morgenstern [1] tried to formalize the process of cracking an egg in formal logic. Nine pages in double column print. And it does not include everything:

To get the conclusion that if a cracked egg is opened into a paper bowl, the egg will leak through, we must add axioms on porosity.

To reason about hard-boiled eggs and coconuts, we must specify how a shell is attached to a shell-inside.

Apart from the fact that most of our world cannot be modeled in math, Gigerenzer dedicated his career to showing in a scientifically clean and convicing way that optimization, i.e. the mathematically clean and generally accepted best decision, turns out to be worse than simple gut feeling in many real-life situations. Optimization works well if the necessary information is present, i.e. the values of all variables are known or at least a statistical value distribution, which means you must have had exactly this problem many times before. In reality this is hardly ever the case. You can be lucky in car factories and this is why everybody admires and imitates car manufacturers. But they are not cleverer than everybody else, they are just operating in a simpler environment.

And this is where religion comes back full circle. Gigerenzer and his team have shown beyond doubt, with oh-so-highly-praised scientific methods (and they did not just follow some rotten rituals of some community, they did real science) that optimization is mostly worthless in real life. That humans outperform math. You would assume that at least inside other scientific fields these findings would be taken into accout and they would adapt their standards accordingly. Of course not! Computer science (Artificial Intelligence in particular), economics, psychology, neuroscience, sports sciences, any field I have some knowledge of or know people working in, are still worshipping optimization as the one and only path to wisdom.

Collective Burnout

Are we degenerating as a species that once praised itself for its intelligence? Or is it just those who are the loudest and most powerful in our society?

Our economy is run by people that don't participate in it and our tech companies are directed by people that don't experience the problems they allege to solve for their customers, as the modern executive is no longer a person with demands or responsibilities beyond their allegiance to shareholder value.
[...]
Think of the Business Idiot as a kind of con artist, except the con has become the standard way of doing business for an alarmingly large part of society.
Ed Zitron, The Era Of The Business Idiot

The Business Idiot is not only to be found in management positions. He is also around in politics and science, in any place where you can gain money, power or recognition. And let's be honest, most people want money, power and recognition. Therefore, it is not surprising that the disease is spreading.

I know several Business Idiots, and only few of them are idiots by nature. They are suffering from their job just as everyone around them is suffering from them. It feels like we are all in one big hamster wheel, chasing one another. The faster we run, the more our brains run out of oxygen. Constant panic mode keeps people pushing the blame from one to another. And instead of accepting the challenges of our world, many opt for conformism and recitation of the mantra of imminent salvation by technology.

As an individual it is getting harder and harder to break out of this madness. I have been working in Artificial Intelligence for more than 20 years. I have tried to make it work and failed. I have seen others fail. I have read about all those promising attempts of previous AI hypes. I know all the patterns how to fool others into attributing intelligence to a calculation engine. And still, with all the press releases and common belief around me, sometimes I start to wonder whether I might have missed something. Maybe human intelligence is based on text generation? Am I just too outmoded to understand the incredible benefits of text generation? Am I underestimating the power of scripts wrapped around the text generators to patch over the obvious shortcomings of statistics?

At least when it comes to using LLMS for software engineering I am sure of myself. I attended a conference this year with several presentators being convinced of the power of LLMs. All I saw were hopes for future success, a screwed attitude towards software development or the truthful admittance that coding with LLMs is just as time-consuming and costly as without.

The only believable application I have seen for LLMs is exactly what they have been built for: text generation. Business reports, proposals for science funding, meeting summaries, health service reporting. All of this can be done so much more efficiently with generated text. Why? The content doesn't matter. So much text in this world gets produced without ever being read. It is sad to see how so many people celebrate their new efficiency without ever questioning the need for those empty pieces of text (see also How AI can save science).

We are in a vicious circle of dumbing down our mental capacities, living in a crazy world, constantly overloading our brains. LLMs are an invitation to run even faster. They operate day and night to produce ever more work and distraction. And the more people are stressed, the less they are able to think. And that makes them even more subsceptible to the next round of propaganda.

Many people have already given up the fight, or never even started. They just accept the promise of the bright tech future, with AI/LLMs as the latest incarnation of the Promised One. They have been ignoring any evidence to the contrary. Why would the burst of the economic AI bubble make a difference? This is not about technology or economy or reason. It is about faith. And faith is sticky.