Thomas G. Dietterich reminds arXiv authors of content responsibility

@tdietterich @ziv_ravid @arxiv @AIXIV Yes, I think that makes perfect sense.

@tdietterich @arxiv Very good, please strongly enforce it.

Attention @arxiv authors: Our Code of Conduct states that by signing your name as an author of a paper, each author takes full responsibility for all its contents, irrespective of how the contents were generated. 1/

@andrewwhite01 @tdietterich @arxiv each author takes full responsibility.

Making humans responsible for their AI use seems like an incredibly reasonable way to address problems & opportunities in the use of AI for academic research, at least in the short term (autonomous scientific work will require different solutions).

@tallinzen @ziv_ravid Everyone is worried about scaling, as if building a mountain of PDFs that no one will ever look at is the pinnacle of doing research ¯\_(ツ)_/¯

This is definitely well-intentioned.

But it also reminds me of the ACL anonymity policy, which ended up creating a lot of friction for many legitimate researchers.

If arXiv starts enforcing this, will it require even stricter screening with the already limited amount of human review capacity?

False positives are already a frustrating issue on arXiv: good papers get put on hold without clear explanations. I worry this policy could lead to even more unnecessary holds and longer review delays.

In the end, it feels like the vast majority of researchers may end up paying the price for the actions of a very small minority.

We must think about how to handle the fact that LLMs can generate papers without any human intervention, but this is not the way. Also, I want to see the arXiv ban senior PIs who upload 40 papers a year.

A few more thoughts: I agree with the concern that Arxiv will become full of millions of AI-slop papers written by agents. **But** this is not the answer. Not only because banning an author for life is not fair (it might that you write only some sections and ofcourse that you can't check all the references one by one), but also because this is not the real problem. AI is writing our papers at this point, and the only question is at what level. If I ask the model to rephrase a sentence and it keeps, " Sure, I will do it for you", it doesn't mean anything about the quality of the paper. The same as we didn't punish papers with grammar errors in the past. Moreover, I can create entire fake papers that will pass this bar from one prompt, and soon, we will have systems that will catch these mistakes automatically. At the end, we need to come up with a semi-automatic system that will check for correctness without the question of whether it was generated by AI (the question of whether we want humans to be responsible for each paper is a different story)

@tdietterich @arxiv Why are meta comments from the LLM such a bad thing? What is the difference with human comments?

@yoavartzi I understand the intention, but this particular version does not scale and does not really incentivize quality. It will simply encourage more superficial checks and then disputes around borderline conditions.

Reputation based mechanisms for accountability seems to be a likely direction esp for agentic use cases.

@tdietterich @arxiv Will a non-corresponding author get a 1 year ban if there are hallucinated references in an arxiv submission?

the thread

This was posted 15 minutes after I uploaded my latest paper to arxiv.

Then I spent the rest of the evening reminding myself that correlation does not imply causation :-D

A neat example of updating our professional norms in the face of new technology. Bold, necessary move by arXiv to push back against the flurry of AI slop content published on its repository.

A neat example of updating our norms in the face of new technology. Bold, necessary move by arXiv to push back against the flurry of AI slop content published on arxiv repository.