Eliza Ward: Hey — glad you're here for this one because I need a gut check.
Brian Reed: Sure — what are we gut-checking?
Eliza Ward: Springer Nature. October 30th report. Fabricated citations found in published books — and a named researcher, Manolya Kavakli-Thorne, confirmed she did not write work attributed to her. Her description: deep learning hallucinations.
Brian Reed: Okay. So the AI didn't just invent a fake citation — it invented her as an author.
Eliza Ward: That's the part that keeps landing on me, actually. And there's a thread going back to a bioinformatics professor named Ali Sharifi-Zarchi reportedly flagging AI-generated manuscripts at Springer Nature before this — but that event is in secondary reporting, no direct dated source.
Brian Reed: So it might be the same story or it might be a separate earlier incident — we just don't know yet.
Eliza Ward: We don't know yet. The confirmed piece is the books, the named researcher, and Springer Nature saying it's investigating with corrections and retractions coming.
Brian Reed: And here's what I want to understand — Springer Nature's own stated principle is that AI can assist but human judgment has to remain accountable. So how did this get into a finished book?
Eliza Ward: Right — but that question gets weirder from Springer to, say, NeurIPS. GPTZero ran an analysis in January 2026 of papers accepted to NeurIPS 2025 — that's the premier AI conference — and found at least 100 confirmed hallucinated citations across 53 papers. Papers that had already passed peer review.
Brian Reed: Hold on. Passed peer review and still had fake citations in them?
Eliza Ward: Caught retroactively, yeah.
Brian Reed: Actually, let me just say this plainly because I think people need to hear what an AI hallucinated citation actually is. It's not a misquote. It's not a wrong page number. The paper the AI cited — it just doesn't exist. It sounds real, the author names sound real, the journal name sounds real. Like if a student handed in a bibliography where half the books were invented but plausible enough that the librarian didn't pull them off the shelf to check. No one checked.
Eliza Ward: And 100 of those, across 53 NeurIPS papers. That's not a rogue submission. That's a pattern inside accepted, reviewed work at the top of the field.
Brian Reed: Which — I mean, that's the part that makes this different from the Springer story, right? The Springer case is one publisher, books, identity fraud on named researchers. The NeurIPS finding is, wait, this is inside the conferences that set the agenda for AI research itself.
Eliza Ward: arXiv moved on this — announced a ban on papers containing AI-hallucinated references. And Nature published a correspondence October 28th arguing these should be classified as research misconduct. Actually — there's a small thing here — a correction came out October 30th adding a co-author, Eve Veromann, who was omitted from the original piece.
Brian Reed: The misconduct label is the live question, though. Because if that classification sticks, every researcher using AI becomes personally responsible for verifying every citation the tool spits out. That's — I don't know if that's workable at scale.
Eliza Ward: Science AAAS reported October 21st that this is still an unresolved debate — not settled. So arXiv banning and Nature calling it misconduct and the research community actually agreeing are three different things right now.
Brian Reed: But the settled framing is the part I don't buy. Because arXiv banning something on a preprint server — that's not enforcement in the pipeline where careers actually get made. Peer-reviewed journals, conference proceedings, that's where your tenure case lives. arXiv banning hallucinated citations doesn't touch any of that.
Eliza Ward: That's the exact gap. And the misconduct label has a deeper problem — intent. Misconduct classifications almost always require showing the author knew. So what do you do with a postdoc at eleven p.m., finishing a literature review, lets an AI tool auto-fill citations to hit a deadline, and three of those references just don't exist? They didn't know. The AI fabricated them and they submitted.
Brian Reed: And that's not hypothetical — that's the intent question made concrete.
Eliza Ward: Nature 652, 26 through 29, 2026 — scientists are still actively debating this. Whether the intent bar is even met. So when people say 'Nature called it misconduct, this is settled' — no. Nature published a correspondence arguing it should be. The field is still arguing back.
Brian Reed: Hang on — so the arXiv ban and the Nature correspondence are being read as consensus, when they're actually just... the loudest positions in an unresolved debate?
Eliza Ward: That's the circulating take that doesn't hold. And — wait, actually there's another piece of this I want to flag. The Ali Sharifi-Zarchi framing. He's showing up in secondary summaries as the person who flagged AI manuscripts at Springer Nature. But there's a Nature letter by someone named Ayyoob Sharifi — different person, titled 'Tackle fake citations generated by AI.' We should be honest that we're not certain these are the same story.
Brian Reed: No, that matters. If we're attributing original flagging to one person and we've conflated two different people, that's — I mean, that's the same credibility problem we're covering.
Eliza Ward: Science AAAS, October 21st — journal rules are fragmented, non-harmonized, vague. There's no enforceable cross-journal mechanism. Which means even if the misconduct debate resolves, there's nothing to enforce it through. And the gatekeeping evidence — what IEEE's investigation actually found versus the 'largest science crisis' framing — is genuinely complicated, and we need to get into it.
Brian Reed: Right — because those two pictures can't both be accurate at the same time.
Eliza Ward: And that contradiction is exactly where Benjamin Skuse's investigation sits. October 27th, IEEE Spectrum — he actually tried to get AI-generated papers past IEEE's defenses himself. Freelance writer, applied mathematician, goes in as a probe. And the finding was — human review plus integrity-checker software is, right now, keeping pace.
Brian Reed: Keeping pace with what, though? That's the part that doesn't land for me.
Eliza Ward: That's the question. Because the same body of reporting has a separate source describing this as — and I'm quoting — 'arguably the largest science crisis of all time,' literature 'flooded with fraudulent papers on an industrial scale.' Those two descriptions cannot both be true simultaneously.
Brian Reed: So one source says cat-and-mouse, we're holding. Another says the flood is already through the gates. And — wait, Bogdan Ibanescu, right? Named researcher, still uncertain whether a citation attributed to him was fabricated by AI. That uncertainty is sitting inside the permanent record right now.
Eliza Ward: Yeah. And NeurIPS 2025 — those 53 papers cleared peer review. So 'keeping pace' can't mean catching everything before it lands. It might just mean catching more than nothing.
Brian Reed: Which is — I mean, that's a really uncomfortable place to park a claim. 'We're keeping pace' as, we are not yet losing completely.
Eliza Ward: Neither of us can resolve that tonight — I want to be clear about that. What we can say: the policy side isn't catching up either. Science AAAS, October 21st — researchers want AI disclosure harmonization, universal cross-journal standards. What they actually have is a patchwork. Journal-by-journal, vague, non-enforceable across institutions.
Brian Reed: So the concrete watch item is actually whether any cross-journal standard gets traction — because without that, even if IEEE holds, Springer's review chain and NeurIPS's review chain are each running their own version of the cat-and-mouse game with no shared floor.
Eliza Ward: No universal standard yet exists. That's the confirmed fact. Everything else — whether the gatekeeping actually scales, whether the misconduct label sticks — that's still open. Watch for whether harmonization moves from a research community ask to an actual adopted policy.
Brian Reed: Springer Nature's corrections and retractions are still forthcoming. That investigation is ongoing. So fabricated citations are sitting in finished, published books right now, and the fix hasn't happened yet.
Eliza Ward: And no universal disclosure standard exists to prevent the next batch. That's not speculation — Science AAAS confirmed it October 21st. The patchwork is the current reality.
Brian Reed: So what would actually settle this? I mean — I don't think it's catching the next bad paper. It's whether retractions can repair a record that, I mean, other researchers already built on. Cited those citations. Ran experiments off that literature.
Eliza Ward: That's the question we don't have an answer to. A retraction notice doesn't reach every paper that cited the original. It doesn't update someone's mental model of a field.
Brian Reed: No. It really doesn't.
