Luca Ambrogioni

I think a key question for minority language revitalization that most movements fail is, "What kind of careers can you pursue while speaking *only* this language?"—and usually the only two answers are "fisherman" or "government bureaucrat".

The penalty is a 1-year ban from arXiv followed by the requirement that subsequent arXiv submissions must first be accepted at a reputable peer-reviewed venue. 4/

The backlash against arXiv is a bit odd. All they're asking is that you read your papers before submitting them.

Note that any of my coauthors can submit or update something on arxiv without my direct consent Minor edits are submitted by one author routinely with the tacit approval of the PI and collaborators With this policy, it becomes a huge risk

There are important distinctions getting blurred in the conversation about arXiv and generative AI. arXiv’s written policy does not prohibit LLM use (see attached). In fact, it directly acknowledges that authors may use these tools, while emphasizing that authors remain fully responsible for the entire submission. But, the recent X thread on enforcement raises deeper questions and seems to mingle evidence of LLM usage with more severe misconduct. What counts as evidence of misconduct versus evidence of imperfect editing using tools that are explicitly allowed? If a manuscript contains hallucinated references, fabricated claims, invented data, or misleading content, that is a serious problem. Such authors have failed in their responsibility, regardless of whether the source was an LLM or not and deserve penalties. But if a paper contains a leftover drafting comment from an LLM - or a stray human comment judged as similar to an LLM, e.g., to a postdoc "Please add the newest data here in a short paragraph" - that also seems to meet criteria for the proposed bans. I feel that treating every visible AI artifact as evidence that the entire paper is untrustworthy risks overshooting the balance between allowing AI slop to run wild and curating impactful research. This is a very tough balance to define, and ArXiv is ground zero for many of these changes and they have done a noble job trying to figure things out. But, circling back to the beginning, this blurring is especially confusing as official arXiv policy says that generative AI use is allowed. If that is the policy, then the presence of some sign of AI-assisted drafting should not be considered the core violation. In my opinion, a more durable standard would separate things like minor drafting artiacts, or sloppy manuscript preparation from the much worse problems of hallucinated references, or worst of all - fabricated claims, figures, or data. I want to conclude by saying thank you to arXiv for their brilliant work as a whole; they are a foundational capability that has allowed AI and science across the board to flourish. But, there is work to be done on this policy. Also, can we please write this into an official arXiv policy instead of trying to infer from Prof. Dietterich's X post alone? :)

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.

Although I generally feel something should be done about arXiv, I can understand the backlash given the severity of the proposed punishment. Could we at least institute a points-based system, like for a driving licence? A direct one-year ban, plus losing the ability to share unpublished work on arXiv, seems too harsh. That said, failing to notice that a reference is incorrect is different from submitting a paper that contains an AI prompt, which could literally be detected by reading the said paper.

seems pretty intense for a mistake that could possibly happen to someone just reformatting text before a deadline (not the hallucinated references, but LLM meta-comments.)