Authorship in the age of AI
OpenAI has released more than 720 working papers that collect mathematical advances put together with very little time, a lot of money, and some "mathematical assistants (AI)" that are a marvel. Once everything is duly verified, we will be able to calibrate its importance. For my part, I reserve the most enthusiastic applause for the day that some central questions in the mathematics of the present are solved (such as the Riemann hypothesis or P not equal to NP). It is certain, however, that the ways of doing things in the mathematical profession –do not worry: like the medical one, it will surpass the 3,000-year threshold– will incorporate an intensive use of the new instruments.
Always, the assignment of authorship has been a fundamental aspect of the organization of research. Traditionally, the transmission of its results is done via communications to entities with accredited reputation, such as journals and conferences (also patent offices). The process has two decisive moments that are crystallized with the acceptance of the communication: scientific validity (expert evaluation) and the registration of authors. On this basis, individual curricula and many other forms of recognition are built (employment contracts, awards, election to academic societies, etc.).
It is an organization that has worked, and I believe that its principles continue to be fully valid. But the shake-up from OpenAI poses challenges. I mention two:
1. The signature of the working documents is “OpenAI”. I do not know if the intention is to go to conventional publication with this authorship. If that were the case, and if it were consented to, it would be a very negative thing. It is common in journalism to attribute breakthroughs to the employing institution of those responsible (“Harvard University has demonstrated...”). That they remain unnamed is a lack of respect, but it is not as serious as it would be if the intellectual authors recognized by the scientific accreditation process –which is what counts– were legal entities. When in a public body like CERN (European Organization for Nuclear Research) the Higgs boson is discovered or when DeepMind –a private company– publishes AlphaFold, on the three-dimensional structure of proteins, the publication has the authorship of hundreds of researchers who will be able to proudly exhibit having been part of a great journey. The articles recognize, of course, the role of the funders, of the entity that hosts or hires them, and of the infrastructures they use. But the authors are natural persons. And this is very much in line with the fact that science advances –yesterday, today, and tomorrow– thanks to the free and independent thinking of humans who use tools. It is a premise that we cannot lose.
2. A more subtle problem stems from the fact that “mathematical assistants” need data to train themselves. Their quality depends on it. Therefore, a researcher who uses them is transmitting knowledge that will make them improve. The next user will benefit from it, perhaps decisively. This is what may have happened with OpenAI’s contribution to the solution –in one of the documents– of the Navier-Stokes equations, and which has given rise to complaints about priority (aggravated by a communication practice that minimizes sources). It is not a new problem. When we use a technological platform, we provide material that refines the quality of the equipment and the technicians who manage it. The problem is inherent to the nature of research. It is complicated if the platform is a private company that has its own research agenda and is careless. To overcome the problem, the company would need to adhere to a code of ethics that separates the platform from its own research and treats it with symmetry with external research.
In this framework, researchers will have to adjust their publication strategies. Some, like Andrew Wiles, will shut themselves away for seven years and end up proving Fermat's theorem, but at the risk of someone else beating them to it. Others, like Julia Robinson, will publish partial results towards the resolution of Hilbert's tenth problem to ensure credit, but risking—as was the case—that in the end it is someone else who completes the program. But it must be said that, if we do not ruin it, the system of authorship recognition based on natural persons will include the names that matter on the path towards the solution. Some will have done it from OpenAI in the final stretch, but also in a prominent place will be those who, like the Spaniards Diego Córdoba and Luis Martínez-Zoroa, have designed the solution strategy. Robinson is well recognized today.
