Raphael Wimmer, a human-computer interaction researcher at the University of Regensburg, generated a complete scientific paper in 54 seconds.
The paper described an experiment Wimmer never performed. It included methodology, results, and citations. He used OpenAI’s Prism, a LaTeX-based scientific writing workspace built on GPT-5.2 that launched in late January 2026 and is free for all ChatGPT users. Wimmer’s point wasn’t to commit fraud. He was demonstrating how trivially the tool could be misused.
The demonstration landed in a scientific publishing ecosystem already buckling under AI-generated content. At ICLR 2026, reviewers identified more than 50 submissions containing hallucinated citations (references to papers that don’t exist, by authors who never wrote them). A study by Pangram Labs found that over half of ICLR peer reviews showed signs of LLM assistance. ICML 2026 received more than 24,000 submissions, a volume that makes manual screening for AI-generated content practically impossible.
OpenAI positions Prism as a writing and formatting tool, not a research tool. That distinction matters on paper. It matters less in practice when the tool can produce a complete paper, fabricated experiment included, faster than most researchers can write an abstract. As Ars Technica noted, OpenAI’s marketing “blurs that line” between assistance and generation.
The institutional responses so far amount to asking reviewers to check harder. That’s the equivalent of asking a dam to hold by requesting the water slow down. Conference organizers face a volume problem that no amount of reviewer diligence can solve when the cost of generating a plausible submission dropped to zero.
The deeper concern isn’t the obviously fake papers. Those can be caught. It’s the competent-looking submissions that pass initial screening, consume reviewer time, and pollute the scientific record with unverified claims wrapped in properly formatted LaTeX. The infrastructure that distinguishes real science from generated content doesn’t exist yet. The tools that generate the content already ship for free.
Source: Nature | [Ars Technica / MIT Technology Review] | Feb 2026