Disclosure and Shame

Max Lieblich, Professor at the University of Washington

The arXiv is full of papers now that include a section, usually either after the acknowledgments or at the end of the paper, entitled “AI disclosure”. I strongly object to this phrase and I think it is a tremendous mistake to use it. Let me tell you why.

The word “disclosure” often brings with it connotations of shame or conflict of interest. I have to disclose otherwise-shadowy information that could have an impact on how other people view my work, my connections, or my incentives. I disclose when I receive money from a company that made a product I am studying in an experiment. I disclose that I was once convicted of stealing and crushing chalk. (These are purely hypothetical; the author has never done a physical experiment with corporate products or been accused or convicted of chalk theft and destruction.) I have to disclose certain defects when I sell a house. A company has to disclose a data breach resulting in theft of my personal information. Disclosure is about revealing information that a person or other entity could or would rather keep secret for its own advantage, or that could bring shame on those making the disclosure.

When telling people that I like chocolate, or that drinking tea helped me wake up in the midst of jet lag, I would not say “I disclose that I enjoy 70% bars made from Madagascar beans” or “I disclose that I had tea this morning”. When I write a paper with a friend, I don’t “disclose” that we wrote it together; we are co-authors. When I get a good idea in a conversation with a colleague, I graciously acknowledge the assistance. It would indeed be shameful not to recognize this help or collaboration. If I were to publish a paper under my name that was entirely produced and written by someone else, that should bring shame and would deserve disclosure.

There is a history of human/machine collaboration in mathematics. Many people have openly used computers in a range of areas for decades. Doron Zeilberger and (his computer) Shalosh B Ekhad have been writing papers together for years. We are now discussing mathematics with and writing papers with AI. Why should we be ashamed of this? What is there to disclose? Perhaps for some people the perception that these tools have agency (they are more like dogs than hammers) lends a tinge of shame to using them. More concerning to me is that the disclosure-shame connection incentivizes early-career mathematicians to hide their use of AI, rather than be clear about how a team of humans and machines produced the work. This has complex implications for training, hiring, and career trajectories.

In many fields, a list of coauthors on a paper includes information about what each author contributed to the work. We have not generally done this in mathematics, but when I look at the “AI disclosure” statements, that is really what I see. I propose that, at the very least, we change the name to “AI contributions”. If an AI model made contributions significant enough that a human doing the same thing would be a coauthor, the model(s) should be listed as coauthors (with their contributions explained). If, on the other hand, the interactions with AI were of a sort that would only merit acknowledgment and not coauthorship were they human, we should simply acknowledge the AI the same way we would humans: in the Acknowledgments section of the paper (with a clear statement of the contributions).

To me, this is how we can build a healthy relationship with our new, increasingly powerful, tools. There is no shame in harnessing the collective knowledge of humanity to ask new questions and find new answers. Let’s embrace it and acknowledge how the teams engaging with mathematics are changing.

Loading…

Filed under Essays

Contribution · Disclosure · Norms

More by Max Lieblich →

New to Proofs and Prompts? Start here →

Thanks for your thoughts. I totally agree with your worry about the phrase ‘AI disclosure’. But I strongly disagree with listing models as a coauthor and I am very worried about how we speak about the use of AI in contrast to any other tools we use. We never list some computer algebra systems, some numeric programs as a coauthor. We never do a ‘disclosure’ of all our searches in google or your favorite search engine. We never list the programs without which we wouldn’t be able to write the stuff, draw the pictures (latex, tixz, gimp, …). But with AI we do – why? I understand that it is all new and we need to learn how to adapt. And I see that there are aspects that are new compared to former technologies and that we do not really understand. But for me it is almost ironically to read when people ‘disclose’ that picture xy was generated by AI, that a text was brought in another format by AI (there are easy command line tools for that by the way we would never even mention) and that an idea/a proof was suggested by model Z. Maybe I am too cynical but especially for the idea/proof part I am wondering whether this ‘disclosure’ just means: I do not know whose ideas I have ‘borrowed’.

We never list some computer algebra systems, some numeric programs as a coauthor. We never do a ‘disclosure’ of all our searches in google or your favorite search engine. We never list the programs without which we wouldn’t be able to write the stuff, draw the pictures (latex, tixz, gimp, …). But with AI we do – why?

Because none of them is able to generate complete papers/proofs.

That is actually not true. Easiest example might be you brute force a counterexample with a computer algebra system. Just because it does not write ‘human looking words around’, doesn’t make it a non-complete proof.

I recommend this enlightening text. https://www.equator.org/articles/arundhati-roy-artificial-intelligence

Thanks for this. I am afraid that most American mathematicians who are ‘AI positive’ are so deeply inculcated in unhinged capitalism that they outright reject any discussion of capitalism and AI as out of scope.

I expect this author to do the same, specifically when they intend to coauthor with AI

Some mathematicians who hold a negative view of AI seem overly intent on positioning themselves in opposition to capitalism. As is the case with so many discussions these days, it appears difficult to find a middle ground.

Excellent, this perfectly encapsulates what AI disclosures are for.

It is also useful to talk about how to handle the incentives potentially created by AI disclosures to behave dishonestly. It might even be useful to discuss the potentially positive role of shame (which is a psychosocial phenomenon that exists for good reasons) in helping to control dishonest behaviour: indeed, if it is generally viewed as shameful to fail to be transparent about one’s research methodology when such transparency is a community norm, then shame might play a useful role in encouraging prosocial behaviour in the face of incentives that militate against prosocial behaviour. It might be useful to discuss the limitations of that approach, perhaps in relation to the decreased (?) role and power of socially-beneficial forms of shame in hyper-individualistic cultures, etc.

Regarding the idea that AI disclosures themselves (rather than a dishonest failure to use them) are inherently a shaming tactic: this may well reflect sincere feelings, but these feelings seem fairly low-priority, and probably kind of idiosyncratic, compared to the obvious concerns that AI disclosures are supposed to be an imperfect mechanism for mitigating.

That also seems fine. Various words would communicate the same information essentially equally clearly, although “disclosure” is more precise than “statement” and I personally dislike needless euphemisms. It’s bizarre, in my view, that the supposed connotations of “disclosure” would evoke a strong enough response to warrant a whole short essay, but I’m sure the OP’s feelings are genuine.

An LLM cannot be a coauthor because an LLM cannot assume the legal responsibilities of an author. For instance, under US law copyright protection cannot be given to non-human authors. It is also not clear how someone could sue an LLM if it violated the law through a publication it “coauthored”, for instance by plagiarism.

Surely you have signed many formal agreements with publishers when you have published papers. Mathematicians often sign them without reading them closely, but especially in today’s climate it would be a good idea to review them. Here is a summary of the agreements used by the AMS: https://www.ams.org/publications/authors/ctp

(this is all in my mind right now since I just got back from an AMS committee on publications meeting where this all was intensely discussed).

Thank you for pointing this out. This should be an obvious issue about LLM coauthorship but for some reason at least myself totally miss it.

Would you prefer “Tool Use Acknowledgement”? Just as we thank colleagues who share their ideas with us, it seems reasonable (at least to me) to thank, or at least acknowledge, the makers of the tools we use to do the research. As just one example, in my subarea it is fairly common for papers to cite SnapPy.

I totally agree! On the other hand, I think most people use the word “disclosure” simply because that is the wording from the Leiden declaration.

I think a disclosure is necessary even from a very selfish viewpoint of the researcher in order to protect him/her from illegal/grey AI actions. For instance if the ideas that AI produces ideas that come from say another person who is also using AI or in the worst case scenario from hacking into someone’s computer, the AI disclosure protects the researcher from liability as well as loss of reputation.

Agree! I’m happy with any of “AI statement,” “Authorial statement,” or going full scientist and having a “Methodology” subsection.

I strongly disagree about listing AI as coauthor.

What I started using, with little forethought, was an “AI usage declaration” paragraph after the “Acknowledgments” paragraph.

Comments are moderated. Read our comment policy.

I actually don’t really understand the argument about “further progress”, which progress are further, which progress are closer, who defines...

That’s a fair point; if you like, you can find my email on the site where the responses are hosted,...

A $5000 prize awarded once every three years at the JMM is a step in the right direction, but a...