Today I joined the Cologne ReproducibiliTea group led by Elen Le Foll, Denis Arnold and Gabriele Schwiertz for a session on generative AI and reproducibility. I’ve made a separate page listing the resources I have drawn on and ended up blogging about this event in two installments. This is Part I; here is Part II.
I was asked to talk about this topic and so I shared a bit of the background behind the guidance on generative AI and research integrity I developed for my faculty. For this guidance, I built on the excellent Dutch code of conduct for research integrity, which spells out five core values that I hold to be uncontroversial for anyone who wants to do robust science: honesty, scrupulousness, transparency, independence, and responsibility.
Now one thing that I have come to realize is that it is offputting to preach about values. This is true and always has been. Nobody likes to be told they might be doing things wrong. The ReproducibiliTea format by the way seems well-placed to partly solve this by facilitating interaction over lecture mode. And indeed there was plenty of discussion, with contributions from scholars across all age ranges and backgrounds. I appreciated the contributions people made. The hybrid format did make it a little harder to do quick responses and to follow through on interesting lines of thought, but we did manage to discuss a wide range of topics.
The praxis of slow science
I had a bit of a Treppenwitz moment after the session because I realized open science praxis provides a perfect analogy. I’ve never done open science because it was “right” or because people told me so. Nor have I tried to tell people how they should do their work (though I do set high standards in my own research groups). I have been doing open science for over a decade now in large part for selfish reasons, namely that working as openly as possible makes me more cautious and makes me triple-check what I do. The result is that can be confident that my work is more robust. “More robust” relative to what? Relative to my own work had I not worked according to open science principles. Added benefits are that it is easier to find (because open access) and easier to build on (because open data and analysis code) — and therefore more likely to make a lasting impact.
The same applies to the case of generative AI. Yes you can speedrun an analysis by vibecoding it. And then when you drop the preprint, people will find that what the LLM told you was a ‘new’ insight was already documented by people 35 years ago (of course unattributed), and moreover that its key contribution was to suggest the wrong mathematical test, which the author didn’t catch, quite likely because LLMs are literally designed to come across as supremely confident and able. What problem does this technology solve, exactly?
As I have pointed out:
A common argument for Generative AI is a promise of higher productivity [5]. Yet productivity does not equal insight, and when kept unchecked it may hinder innovation and creativity [6, 7]. We do not need more papers, faster; we rather need more thoughtful, deep work, also known as slow science [8–10].
Citation [8] here goes to Isabel Stengers, whose work on slow science perhaps has never been more crucial and relevant. (I’ve blogged about slow science before.)
So, I dislike telling anyone what they should or should not do. But I know why I am taking a values-first rather than a technology-first approach.
I prefer my reading to be serendipitous instead of glued to the probability mass of whatever data the text regurgitator was trained on.
I prefer my scholarly work to be sharp instead of sanded off by the mediocrity of synthetic text.
I prefer my analyses to be directly rooted in empirical data rather than mediated by a statistical distillation of the scrapeable internet.
I prefer to keep my wits about me rather than entrust myself to automated thoughtlessness.
Part II continued here: Don’t seek permission, center values