Has AI brought progress in chemistry?

Everyone talks about AI these days. What can it do? How will it make our lives better? How intelligent is it – more so than us? And, is it conscious? Honestly, sometimes it feels like we are entering a Terminator movie. Whether this is because people purposely want to make us feel like this or whether we actually are, I cannot tell.

As things stand, it would simply be ignorant to claim that AI has not brought progress of some sort. The case of AlphaFold is the most celebrated example of progress through AI: the longstanding problem of protein folding is said to have been finally solved thanks to the development of this model.

However, specifying how AI has ameliorated our lives requires careful unpacking. After all, there are different kinds of progress that can be achieved. Take some examples from chemistry. Did Mendeleev’s periodic table constitute the same kind of progress as —say— the discovery that urea can be made of inorganic material; that of oxygen; the invention of the spectroscope; or, the development of AlphaFold? These are different kinds of achievements in the sense that they have brought about different sorts of improvements to humanity.

In the literature on this question, this complexity is generally captured by distinguishing between three kinds of progress: technological progress, social progress, and cognitive progress. Technological progress seems to be straightforward. Anything that humans have come up with, including easier or more efficient ways of doing things as well as entirely new inventions that allow us to do something we couldn’t do before, constitute examples of technological progress. For example, the fast travel enabled by cars compared to carriages constitutes technological progress. Likewise, digital thermometers are safer and faster than mercury thermometers. 

Social progress is said to be achieved when a discovery or invention improves our quality of life in some respect, either at an individual or social level. For example, the discovery of new dyes and of new techniques of making them at an industrial level brought about economic growth through the rise of the dye industry in Europe and the USA during the 19th century. Then, advances in energy storage and batteries that allowed for the production of electric cars can be said to have brought social progress in the sense of diminishing the use of fossil fuels in the automotive industry and thus minimizing pollution. The synthesis of new drugs continuously brings social progress in that they allow us to treat previously incurable diseases. And, the list goes on.*

But what about cognitive progress? This is a trickier concept to pin down and philosophers have extensively thought about different the ways it can be achieved. Larry Laudan defined it as being “nothing more nor less than progress with respect to the intellectual aspirations of science” (1977: 7). From this very definition, it becomes evident that the main source of cognitive progress is scientific knowledge, including of course chemical knowledge.

So, how has AI contributed to this form of progress? To answer this, it is perhaps useful to approach it in a piecemeal manner. From the perspective of chemistry, one can ask: How has the use of AI in chemical practice contributed to cognitive progress? Has it advanced the intellectual aspirations pursued through chemistry?

This is a question I’m currently investigating through my research project PExAI: Predicting and Explaining with AI. In fact, in my recently published paper, I argue that it is not evident that AI’s use in chemistry constitutes cognitive progress just yet. While its development surely constitutes technological progress, it is not as clear that it has led to advances in the way we understand and explain chemical phenomena.

To support this, I present the four main accounts of cognitive progress that have been developed in philosophy of science, and investigate them from the perspective of chemistry. Very briefly, these accounts are:

1.     The truth likeness account: cognitive progress is achieved by discovering new truths about the world.

2.     The epistemic account: cognitive progress is admitted not by the mere discovery of true facts. It also has to be the case that we have produced empirically well-established justifications for those facts.

3.     The problem-solving account: progress is achieved through the solution of puzzles. These need not concern solely theoretical puzzles or conundrums, but can include ordinary puzzles that occupy scientists in, say, a lab.

4.     The noetic account: cognitive progress is only admitted if the purported discovery or achievement has led to an improvement of our understanding of the world around us.

I do not discuss which account is best in capturing cognitive progress. Instead, in the paper, I show that, when it comes to the use of AI in chemistry, it is not evident that all four accounts are straightforwardly satisfied. Some are, but others— not so much!

Let’s start with the easiest one: the problem-solving account. Here we have plenty of examples from chemistry where AI indeed managed to solve puzzles that have troubled chemists. Take for instance the protein folding problem. AlphaFold managed to predict the structure of millions of proteins; a pursuit which hadn’t been achieved for decades (not for the lack of trying). The truth likeness account seems to be satisfied as well. Just think of how many new substances AI models have predicted with specific desired properties, most often for medical applications. Aren’t these discoveries new facts we uncover? In a sense, yes.

Things start to get more complicated with the epistemic and noetic accounts. In part, this is because of a problem that has already troubled philosophers of AI: the so-called problem of opacity. Very crudely, this has to do with the fact we just don’t know how AI models manage to make the predictions they do. Even though highly successful and accurate, it is unclear to us how the models use the data by which they are trained and produce the predictions they do. This is problematic because it implies that— strictly speaking— AI does not give us a justification for the truths it delivers and, in some sense, does not produce (at least not on its own) a better understanding of the target phenomena.

This is, in a nutshell, the thesis of the paper. I realise I have left out crucial scientific details that support these claims; so if you find that what I am saying is not convincing, please do read the paper! In any case, even if I am wrong, it is worth considering that the hype around AI might be a bit overblown. Indeed, it has been immensely helpful to chemical practice; but whether it has already revolutionised our understanding of chemical phenomena (and, I would venture, the natural world more broadly), I’m not so sure we can say that just yet.

*   Of course, one could point out how such advances have been also detrimental to humanity in some respects. Thank you, Sarah Hijmans, for pointing this out!!

*   Research for this paper was supported by the project PExAI that is funded by the Hellenic Foundation for Research and Innovation (H.F.R.I.; project number 22712).

*   The paper: Seifert, Vanessa A.. "Chemical progress in the age of AI" Pure and Applied Chemistry, 2026. https://doi.org/10.1515/pac-2026-0748

References & Further Reading

Dellsén, Finnur & D'Alessandro, William (forthcoming). Scientific Progress in the Age of AI. Cambridge: Cambridge University Press.

Laudan, L. Progress and its Problems; Routledge and Kegan Paul: London, 1977.

Seifert, Vanessa A.. "Chemical progress in the age of AI" Pure and Applied Chemistry, 2026. https://doi.org/10.1515/pac-2026-0748