AI challenges astrophysics on the direction of research

These two cases represent two extremes: JWST is accessible to individual researchers, following a lengthy process of proposal submission and rigorous selection; the second is completely inaccessible to individuals, who can, of course, make use of the final, processed results.

“In fact, individual researchers are often inclined to use their own criteria to transform raw data, raw data , into usable data,” says Roberto Ragazzoni, president of the National Institute of Astrophysics. It is also worth remembering that astrophysicists are not interested in the ‘beautiful’ images – which are, quite rightly, shown to the public – but in those in which every single pixel tells us how much light has been captured at that tiny location. Removing all possible noise caused by instrumentation or other factors is essential, especially when examining extremely faint objects, and everyone has their own method. The second example, that of Vera Rubin, is, by contrast, the ideal case for trying out AI methods and approaches.

“The real point,” continues Ragazzoni, “is that from now on we must design new equipment with the use of AI already in mind.” The problem he highlights is that of the expertise of the researcher doing the design, which must be in place before the idea to be realised, and this brings with it the whole issue of training.

In short, the adoption of AI in astrophysics seems like a path to paradise, but it is certainly fraught with countless pitfalls, partly due to the media’s portrayal of these systems as commonplace, whilst there are only a few specialised applications – such as programming and mathematics – that currently offer a genuine and almost immediate benefit.

On the one hand, Large Language Models (LLMs) are becoming capable of designing and writing entire scientific articles; whilst the quality is not yet up to scratch, it is conceivable that it will improve in the near future. On the other hand, as we have just seen, the production of astronomical data is becoming increasingly professionalised and separate from the work of astronomers.

Nowadays, it is easy to obtain useful information from AI systems, and one might consider using them for repetitive tasks, but there is a world of difference between that and thinking, for example, that the work of students – even PhD students – will become ‘useless’.

‘It’s true that mentoring a young researcher takes time and commitment, but the aim must be to train a new generation of astrophysicists, ideally even better than ourselves,’ continues Ragazzoni. Moreover, to say that an LLM ‘performs better’ than a student – as one sometimes hears – is, conceptually, a fatal mistake: the student is there to learn, to develop their skills and, ultimately, to advance our understanding of astrophysics.

Of course, the use of AI fuels the temptation to ‘hurry up’, ‘produce’ and ‘publish scientific papers’ – catchphrases which, unfortunately, are all too often the hallmark of scientific research that risks losing its way.

On the other hand, AI could easily publish, so to speak, such a vast number of scientific articles that no human could ever read them; and in any case, whilst it is true that knowledge only exists if it is published, it is also true that, in the world of the scientific method, LLMs cannot be authors, simply because they cannot take responsibility for what is written.

David Hogg, a cosmologist working between New York University and the Max Planck Centre in Heidelberg, highlights two possible solutions, both of which are flawed: leaving all research to the new generation of LLMs, excluding humans, or banning them altogether – a situation that is, of course, impossible. Before deciding how to integrate LLMs into astrophysics, Hogg concludes very wisely, we must clarify why we do astrophysics and what values we wish to preserve.