The Sentence That Made Me Laugh — Then Terrified Me

“Pregnant women developed uncontrollable cravings for prime numbers.”

That sentence is in a published scientific article. In an obstetrics journal. Peer-reviewed. Accepted with “minor revisions” and praised as “very well written and interesting.”

When I read the story of Pascual D. Diago — a mathematics education professor at the University of Valencia — my first reaction was laughter. The second was a chill down my spine. Because what he did isn’t just a joke. It’s a diagnosis of the state of science in 2026 — and its consequences for AI.

The Professor’s Experiment

Diago was tired of spam from predatory scientific journals — those that send emails soliciting articles, charge publication fees, and approve anything without real review. The last straw came when he received an invitation from the Clinical Journal of Obstetrics and Gynecology — completely outside his field (mathematics education).

Instead of ignoring it, he decided to stress-test the system. He used ChatGPT to generate the most absurd paper possible for obstetrics.

The result was a masterpiece of absurdity: the central thesis claimed that teaching pelvic geometry to pregnant women reduced birth anxiety by 13.7%. The text mixed the Fibonacci sequence into labor graphs. And the cherry on top: pregnant women developed cravings for prime numbers.

The graphs explained nothing. Conclusions were patently impossible. Cited sources contained coded messages generated by AI essentially saying “I am lying, this is a fake text.” As Diago wrote on Retraction Watch: “A cursory glance at the abstract or sources would have been enough to realize nothing made sense.”

The timeline: in October 2025, he submitted under the pseudonym “Pascual Chiago.” On November 12, a “Susan Lee” (not listed on the journal’s staff) demanded urgent response to review comments within 24 hours. Diago resubmitted the exact same file five minutes later, randomly highlighting passages in yellow, making zero actual changes. Within an hour, he received final acceptance. An invoice for $2,949 followed. He didn’t pay — but the article was published anyway. And Diago was invited to speak at a gynecology conference.

20,000 Predatory Journals (And This Isn’t the Exception)

The first reaction many people have is: “that’s a scam journal, it doesn’t count.” But the problem is far bigger.

An estimated 20,000+ predatory scientific journals operate globally. The business model is simple: charge publication fees from researchers desperate for CV volume, completely ignoring any real review. Researchers under “publish or perish” pressure pay — and their CVs grow with publications nobody read.

But the garbage isn’t confined to bottom-tier journals.

NeurIPS 2025: The Garbage Reached the Top

In January 2026, GPTZero scanned 4,841 of the 5,290 papers accepted at NeurIPS 2025 — one of the world’s most prestigious AI conferences, with a 24.52% acceptance rate.

The result: 100+ completely hallucinated citations spread across 51 accepted papers. References to authors, papers, and DOIs that don’t exist. AI-generated fabrications ranging from obvious placeholders (names like “John Doe and Jane Smith”) to “uncanny valley” citations that looked completely legitimate — real author names paired with fake paper titles and fabricated DOIs.

Each paper passed review by at least 3 reviewers. Each beat 15,000+ competing submissions. Each was publicly presented at NeurIPS in November 2025.

GPTZero coined the term “vibe citing” — just like vibe coding produces code that seems to work but breaks under pressure, vibe citing produces references that look legitimate but crumble under verification.

And a Retraction Watch data point from 2026 is even more alarming: an analysis of PubMed-indexed papers found that 1 in every 277 papers published in the first 7 weeks of 2026 references a paper that doesn’t exist. That number was 1 in 458 in 2025, and 1 in 2,828 in 2023. The jump coincides precisely with massive adoption of AI writing tools.

The Garbage Loop: Why This Matters for AI

Here’s where the story gets truly concerning — and connects to everything I’ve written about training data.

Frontier language models are trained by scraping the internet for “high-quality” data — and scientific articles are among the most valued sources. PubMed, arXiv, institutional repositories — all treated as “gold standard” by training pipelines.

When the scientific ecosystem is flooded with fake texts — papers with hallucinated citations, invented data, impossible conclusions — tech companies end up using this “academic garbage” to train next year’s models.

It’s a feedback cycle: AI generates fake data → humans publish without reading → the next AI learns from these lies → generates more sophisticated fake data → humans publish → …

As I discussed in the End of the Infinite Internet post: quality public data is running out. And now we discover that part of the data we thought was high-quality (academic papers) is contaminated. The “gold” has impurities.

ScienceDirect published a paper in April 2026 dedicated entirely to this problem: “Hallucinations in generative AI: A threat to scholarly integrity and the urgent need for publisher-led academically supervised verification.” The title says it all.

What I Take from This

First: “published” doesn’t mean “true.” This should be obvious, but in practice many people treat scientific articles as verified facts. In 2026, we need an additional layer of skepticism — even with peer-reviewed papers.

Second: the crisis is systemic, not anecdotal. Diago is the joke. NeurIPS is the warning. PubMed at 1 in 277 is the epidemiological data point. The problem isn’t one predatory journal — it’s the entire ecosystem under pressure of volume, speed, and tools that facilitate fabrication.

Third: citation verification should be automated and mandatory. GPTZero is working with ICLR and other publishers to integrate hallucination checks as a formal publication step. This should be standard at every serious venue.

Fourth: AI trainers need much more rigorous data curation. If academic papers are contaminated, training pipelines need filters that don’t exist today at scale. “AI-generated academic garbage training the next AI” is the most dangerous loop in the ecosystem.

Conclusion: Who Will Review the Reviewers?

Diago’s experiment proves that human intelligence failed at its most basic curation function. The moment we turned scientific publishing into a monetized assembly line, we opened the floodgates for digital garbage to overtake our knowledge base.

If we don’t create rigorous audit mechanisms — human and technological — for scientific articles, we’ll soon be unable to distinguish a real medical discovery from a chatbot’s statistical hallucination.

And the final irony: the math professor who proved the system is broken was invited to give a talk on obstetrics. That says everything we need to know.

Share if this changed how you read papers:

Pregnant women craving prime numbers. Fibonacci in labor. Accepted with “minor revisions.” If science can’t filter this, how can we expect AI trained on this data to be reliable?

Read Also