My Use of AI

I want to be completely transparent about my use of LLMs on this website. I use them to:

• search for information and references;
• gather data and produce overviews of the state of the art on scientific, philosophical, or historical topics;
• develop ideas, arguments, and criticisms through dialogue with LLMs;
• produce summaries and translations of my texts written in French;
• create animations and numerical simulations for my physics articles;
• get critical feedback on my writing;
• suggest formulations or reformulations of certain passages;
• correct and edit the final versions of my texts.

The use of LLMs to help write texts remains a highly controversial topic. I think the debate sometimes conflates fundamentally different practices. One concern is that we may delegate our thinking to machines, thereby losing human responsibility and authorship over the content. Sometimes we read a text because we want to know what a person thinks, not what a machine thinks. We may therefore feel misled if we discover that a text was written with the assistance of AI.

For certain kinds of writing, it is indeed problematic when an author simply gives an LLM a prompt containing more or less detailed instructions and asks it to produce an almost finished text. I want to make clear that none of the articles on this website were produced in this way, with the exception of the summaries found on the presentation pages for articles or for the library.

However, I also want to be transparent about the fact that, for some articles, I have made substantial use of AI as a writing tool. This is particularly the case for articles, or sections of articles, that mainly involve presenting the ideas of other authors.

In any case, there is one thing I would like to emphasize: in none of my articles has the use of AI reduced the amount of writing I have done. Using AI makes me write more, not less — except when it comes to coding, of course. The reason is simple: even in passages where I let AI write the final version of what will be published — because I find the formulation particularly clear and an accurate expression of my thoughts — the final product is the outcome of an iterative process in which I remain fully active and in control.

I also believe that this way of working makes me think more, not less, at least about the aspects that really matter to me. My writing process sometimes looks something like this: I write a first draft, and the AI criticizes some of my ideas and formulations. Sometimes I defend them when I think the criticism is unfounded or irrelevant. At other times, I concede the point. The AI may then suggest a new formulation. I can accept it or, in turn, criticize it. This can lead to endless discussions about conceptual distinctions, judgments about the coherence of an argument, and so on.

The risk that the use of AI may lead to an atrophy of thinking seems to me to depend strongly on how it is used. Likewise, the idea that simply running a text through an AI detector would allow us to measure how much its author was involved in the writing process seems misleading to me — at least in some cases.

Socrates did not write texts. Richard Feynman had at least a substantial proportion of the books published under his name written up by others, including his famous physics lectures. Writing ability does not exhaust the strength or originality of a person’s thought. My argument is not necessarily that only the final result matters, although that is also a position that can be defended in some contexts. My point is rather that delegating certain cognitive processes may allow some kinds of intellectual profiles to concentrate on what they do best.

That said, two things have made me reconsider some of my texts — a process that is still ongoing — where I may have delegated too much of the final writing process to AI. The first is the occasional presence of prose that I find a little too literary, polished, and overly cautious. The second, and more important, is that some people lose interest in a text when they learn that it was heavily assisted by AI. This is something an author has to take into account.

I have therefore decided to reduce the role of AI in the final wording of my texts compared with how I used it for at least some texts when I first launched this blog. And for the sake of even greater transparency, from now on I will provide an estimated percentage of the final prose in each text that was produced with AI assistance. This estimate will necessarily be very approximate: I do not have access to the paid versions of the best AI-detection tools.

Adrien Vila Valls