←
Thomas Bloom1,887 posts
Thomas Bloom avatar

Thomas Bloom

@thomasfbloom

Royal Society University Research Fellow at the University of Manchester. Mathematician and owner of erdosproblems.com. He/him/his.

◷ Joined December 2020▧ 42 Photos and videos
Thomas Bloom avatar
Thomas Bloom@thomasfbloom·22h

Part of the reactions to this come from negative connotations with the word 'disclaimer', I.e. that it implies something is wrong or shameful about AI use. This is not my intention; probably should have used something more neutral, e.g. "disclosure" or "acknowledgement"

In the past few weeks I have seen some significant mathematical papers that have no AI disclosure at all. Yet examining the proofs, and the writing style, it is clear (to me) that there was significant (or perhaps total) AI use. This is bad for everyone. 1/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

I don't see a way of enforcing this in practice. But we should make it clear that intentionally misleading people, to a significant degree, about AI usage is academic fraud - and any short-term gains will be overshadowed by the long-term harm to your reputation. 12/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

Ideally, everyone should include an accurate AI disclosure (even if that disclosure is 'no AI was used'). This lets us see what went into producing a proof/paper, and reward appropriately. 11/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

Or, more likely, something in between. But unless they know what type of work people are doing, it will be impossible to allocate resources in a way that encourages what they want to encourage. 10/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

Perhaps a funding body wants to support heavy use of AI, mass-generating proofs with the goal of discovering truths as efficiently as possible. Perhaps they want to support human development, encouraging people to think deeply about mathematics without AI assistance. 9/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

But if the system is being gamed, then this support cannot be assigned properly, to where it would help the most. The system *does* need to be radically redesigned in this new age - but whatever the new system of rewards will be, we need accuracy on what people have done. 8/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

But emotional response aside, there are practical reasons this is bad - it breaks the entire system of rewards. If someone has produced a significant result using AI use, great - this can be recognised and rewarded appropriately. 7/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

There should still be -some- accolades! Great AI (or AI-assisted) work should be recognised and rewarded. But this should be done in a transparent way; be open about the process and let others decide how much credit/reward individuals deserve. 6/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

My immediate reaction is based on emotion and instinct -it just feels unfair to claim the same accolades as someone who has spent years working on a problem, just because you have money and tokens to burn. It offends me on a gut level. 5/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

My problem is not with using AI to generate these proofs - lots of exciting discoveries are being made. But these fradulent authors presumably want to be rewarded as if they themselves came up with the ideas unassisted (or else why not accurately disclose AI use?) 4/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

These mechanisms are meant to recognise the implicit skill and effort that went into producing a proof, and reward the authors proportionately. (e.g. so that they can be supported enough to put similar amounts of effort into future work) 3/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

I assume the authors did this so that they can claim full 'credit' for the paper, without having to share it with AI. They are dishonestly taking advantage of old-world mechanisms that reward papers, results proved, etc. 2/

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 17

In the past few weeks I have seen some significant mathematical papers that have no AI disclosure at all. Yet examining the proofs, and the writing style, it is clear (to me) that there was significant (or perhaps total) AI use. This is bad for everyone. 1/

Alessandro Della Corte avatar
⟳ Thomas Bloom retweeted
Alessandro Della Corte@Ognifedefingo·Nov 14

I think a lot of things here are wrong. I'll try to explain in a short thread. 1. The narrative "human vs AI" is nonsensical in itself. No one is trying to cut costs by "replacing" pure mathematicians with machines, since no one pays them *out of need for their results*. Pure

Wouldn’t this just cause humans to focus on where they think they have an edge over ai? Theory building? What are the fields it’s gonna be bad. Math success is just being a good ai capability predictor and picking the right subfield with longest AI resistance.

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 10

They do note that the 'length' here is context dependent - a feature not a bug. e.g. some results in algebraic geometry are not interesting to me, taking ages to state/define, but to someone who knew the definitions, the statement is very short, and the result interesting.

Thomas Bloom avatar
Thomas Bloom@thomasfbloom·Nov 10

Or results with short statements and long proofs, that are nonetheless not interesting at all?