The arXiv made a decision last week that shouldn’t have been controversial.
If a research paper goes out with hallucinated citations in it, and there’s clear evidence the authors never checked the AI-generated results before submitting, the authors can be banned from the platform for up to a year. Thomas Dietterich, who chairs arXiv’s computer science section, said it plainly: if you can’t trust that an author checked their own paper, you can’t trust anything in it.
That’s not a new standard. That’s the oldest standard there is. Your name goes on the work, you stand behind the work. Nobody invented that rule for AI. AI just made it easier to break without anyone noticing for a while.
The pushback came fast, and it came from people who should know better. One economics professor asked what happens when a citation is in a language he doesn’t read, or covers material he doesn’t understand, but a co-author does. An AI researcher called the policy too strict, arguing that a paper is long, mistakes happen even to careful people, and treating every slip the same as a serious false citation is unfair. A biotech founder went further and called the whole thing gatekeeping.
I want to take that seriously instead of waving it off, because there’s a real question buried inside the complaining. Not every mistake is the same size. A copy-paste error in a supplementary table isn’t the same thing as a fabricated source that doesn’t exist. Lumping those together is sloppy thinking, and arXiv’s critics aren’t wrong to point that out.
But here’s where the complaint runs out of road. The question was never whether AI makes mistakes. Of course it does. The question is who’s supposed to catch the mistake before it goes out under a human name. And the answer has to be the human, every time, because the alternative is a discipline where nobody is actually responsible for what gets published — not the author, who can point at the tool, and not the tool, which was never capable of standing behind anything in the first place. That’s not a stricter world. That’s a world with no accountability left standing in it at all.
This is the same question I’ve been building the Faust Baseline around, just wearing different clothes. Somebody has to be the one who checked. Somebody has to be the one whose name means something when it’s attached to a claim. A system can disclose what it is and isn’t sure of. A system can flag where its own confidence runs thin. What a system can’t do is take the place of the person who’s supposed to look at the output and decide whether it’s true before putting their name underneath it. That decision has to stay a human decision, or the word “author” stops meaning anything.
The complaints about language barriers and co-authors and long papers are really just complaints about how hard the job is now. They’re not wrong that it’s harder. Checking AI-assisted work takes real time, and time is the one thing academics have always been short on. But “this is hard” and “this shouldn’t be required” are two different arguments, and the people pushing back on arXiv keep sliding from the first one into the second without ever actually making the case for it.
I don’t think arXiv got every detail right. A one-size penalty for a fabricated source and a sloppy citation probably does need a second look. But the core of it is right, and it’s worth saying plainly: if your name is on it, you checked it. If you didn’t check it, your name shouldn’t be on it. That was true before any of us had heard of a hallucination, and it’s still true now that we have.
Challenge this response?
“The Faust Baseline Codex 3.5”
micvicfaust@gmail.com
Post Library – Intelligent People Assume Nothing
”AI Baseline Governance”
Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited. © 2026 The Faust Baseline LLC
