Why AI is Systematically Incapable of Revolutionizing Science

AI capabilities have increased dramatically the past few years, both inside the fields of natural science and outside of them. Solving decades-old mathematical problems with AI has become so routine at this point it’s almost become a meme. Personally, I find modern AI tooling to be an amazing tool and use it constantly, for anything from movie recommendations to coding. I am astonished by the breadth and depth of its functionality.

However, something I find naive to the point of absurdity is the idea that AI is somehow going to revolutionize and solve scientific inquiry, biology and medicine and lead us to a disease-free golden age. This misses the key point that AI does not solve problems. Rather, people use AI tools to solve problems, and when an AI has a useful and credible insight, it requires some specific person to champion its ideas. And the people who do so are just as subject to science as a social and economic institution as has been anyone who’s proposed radical scientific ideas in the past.

In science we might classify discoveries into two camps - flag planters who invent or turn over entire subfields, and discoveries made within those subfields once established. This distinguishes someone like Newton, Einstein, or Donald Knuth from those who build on their ideas. It is clear that AI tools are incredibly good at building on existing subfields, particularly in mathematics, which does not necessarily require interfacing with the physical world. However, we have less evidence that they are good at coming up with genuinely revolutionary discoveries. And even if they were, the social process by which their new ideas would be integrated into current scientific knowledge is far from clear.

The history of correct scientists working outside the mainstream and being mocked for it is long and clear. Semmelweis was put in a mental institution for wanting doctors to wash their hands. Einstein was initially mocked for his idea of special relativity. Galileo was put under house arrest by the Catholic Church, which, as the foremost scientific insitution of its day, defended the then-current scientific orthodoxy that the sun revolved around the earth - just as our scientific institutions today defend our own orthodoxies. This is not entirely bad, as some stability is necessary for prolonged scientific inquiry. But it is a real historical fact about how science works and progresses.

To say that AI will revolutionize and solve science is a bit like saying any of these revolutionary thinkers would have been taken more seriously if only they had the right ChatGPT window supporting them. This makes no sense, because they were not (temporarily) defeated by lack of scientific acumen, but by their peers’ incredulity. And if we’ve learned anything from the past few years, it is that, outside of some difficult to dispute cases like finding a clear mathematical counter-example or proof, having AI support your views does not necessarily make them more credible.

But things are even worse than this. We tend to think of the medieval world as unsophisticated for believing the sun revolved around the earth. However, they had very detailed mathematical models based on this idea that for the most part worked at predicting the motion of the sun and planets. We say that the earth revolves around the sun not because it is impossible to construct a model supporting the medieval view (in fact, considering the relativity of reference frames in mechanics, we know it is possible), but because our view is supported by a simpler model. At its core, it is an aesthetic preference, one which prefers a world in which physical forces are simpler, easier to explain, and fewer in number.

Now, suppose Galileo back in the 1500s did have ChatGPT, perhaps running on his abacus. If the question of which heavenly body is at the center of our solar system is essentially a question of aesthetics, and AbacusGPT is trained on the aesthetics of a century or more of medieval astronomy and theology, why would it prefer the modern view that the sun is at the center? Perhaps it would be willing to admit the possibility, but it would have to be guided towards it by the prompter, and readers of the output of the modern-astronomy sun-pilled AbacusGPT chat window claiming the earth revolves around the sun would rightfully object that the model was merely guided to its conclusion by Galileo, and that you could get it to claim just about anything with the right kind of prompting.

For a more modern example, consider the medical history of tobacco usage. It seems obvious to us today that smoking cigarettes is bad for you. But throughout the 1900s, this fact was disputed. Many doctors believed the practice was harmless, with Camel even bragging in the 40s that more doctors smoked their brand than any other. Any intrepid researcher at this point claiming the link between cigarette smoking and lung cancer would be going up against an industrial behemoth. Many of them, such as Alton Oschner, were in fact mocked viciously by peers for their claims and countered by industry-funded science.

Now, if these health advocates had access to AI tooling back then, would they have been more successful? The tools themselves would have been trained partly on science funded by the tobacco industry, and at best would be likely to be ambivalent about the connection between smoking and cancer.

To claim that AI will take over and complete science is to claim that there are no revolutions left in science, because AI, if choosing between two competing models of a phenomena, without human guidance will simply default to whatever its training data tells it to be true, which in most cases will be the current consensus. It is to believe that we understand the universe and life at a deep and fundamental level, and all that is left is the bean counting of physical edge cases and molecular stamp collection.

Contrary to popular belief, modern science can be very heterogenous. An international team of scientists and engineers have been working on an alternative theory of the atom for about a decade. Plasma cosmology presents an alternative to how our solar system and ultimately universe is structured. Gilbert Ling’s work contains a revolutionary view of cell biology based on semiconduction which he promoted for nearly six decades before passing in 2019. Gerald Pollack has built on Ling’s work to propose a novel theory of how water works.

Now, I do not mean to argue here that all or any of the above scientists or scientific theories are correct. My aim is to point out that even if some of them or any similarly contrarian scientists were correct, the existence of AI tooling is unlikely to help their research become accepted. In every case, they are going up against peers who are more established and better funded, in a system of doing science that privileges those who are established and well-funded, and who often sit on peer review boards deciding whose works are published and who gets further funding. In every field people get comfortable with the way things are done, and most folks who’ve spent decades establishing themselves don’t really like to see the floor pulled from under their feet. What most people call knowledge is often a collection of accumulated habits. And while it may be argued that dissenting voices in science typically win out over the long run if they are correct, that may just be survivor bias speaking. It is entirely possible that there are revolutionary scientists in our history whose ideas never got the recognition they deserved, and are now buried in time.

This leads us back to the most absurd claim of the AI era, that AI is going to solve medicine and end disease. But this will only happen insofar as our medical industry and scientific institutions use AI to do so. And they are likely to only use AI to make advances that do not rock the boat too much, and which support their profit motives and entrenched experience. Some of those advances may be very useful. But we can see from history that many of the most useful insights such as heliocentrism or relativity come from thinking outside of the box. And just as before, anyone with a differing view from the scientific consensus will be only one opinion among many - no matter how much Claude tells them that they are a genius.