On 27 August 2026, Wharton professor Ethan Mollick posted that he had found "a number of AI slop papers put on preprint sites with my name on them, which I have never written nor seen," and that other academics had found the same. He added that "AI paper factories of at least okay quality are coming soon." The post quoted Andrew Gelman's blog item about University of Chicago statistician Nicholas Polson, whose SSRN page listed 258 papers in 2026. Mollick's line "I wouldn't assume that this is all real" raised impersonation as a possible explanation.
This matters to builders because preprint servers and DOI repositories feed training corpora, retrieval-augmented generation (RAG) indexes and research agents. If bylines, venues and dates can be forged in bulk and still receive real DOIs, then an author's name and a DOI stop being trust signals. Any pipeline that treats them as evidence inherits the contamination.
What we checked
- Can Mollick's claim, that AI slop papers bearing his name sit on preprint sites, be independently verified, including which platforms and identifiers? unresolved. The post exists as quoted, dated 27 August 2026, but it names no platform, title or identifier. A Zenodo API query for creators named 'Mollick' returned 16 records, and none was attributed to Ethan Mollick. I found no independent report that documents his case. [1][13]
- Is fabrication or misattribution of AI-generated papers on preprint and repository infrastructure independently documented at scale? partly confirmed. Fabricated authorship is documented at scale: 1,655 ghost-authored Zenodo records with real DOIs, and 15 fake SSRN papers credited to invented authors. Papers falsely credited to real, named researchers are documented case by case, mostly in fake journals. The scale of that second problem has not been measured. [5][6][7][8]
What we found
- The post is genuine and was published on 27 August 2026. It quotes Gelman's post on Polson and gives no list of affected papers or platforms. The newsroom marked the item as fresh, but it is about two months old. [1][2]
- The Polson case that set off the post was not impersonation. Coauthor Vadim Sokolov wrote on Gelman's blog that the papers were "genuine preprints produced and posted by the named researchers." He also acknowledged that AI involvement varied and "was not consistently disclosed." The Washington Post headline says SSRN removed 257 papers. The College Fix, citing the Post, says 253 and quotes SSRN: "The most significant issue was the unusual submission activity." [2][3][4]
- A 2026 arXiv preprint, not peer reviewed, identifies 1,655 Zenodo records that claim two nonexistent journals. The records carry real DataCite DOIs, and their publication dates were backdated: they claim 2020 to 2023, while the DataCite registration shows March to April 2026. There were 991 uploads in March 2026 and 666 in April. The authors trace many of the bylines to LLM default names, such as 'Elena Vasquez' for Claude and 'Aris Thorne' for Gemini. They report that as of May 2026 none of the records was indexed by Semantic Scholar. [5]
- Retraction Watch reported 15 SSRN papers that rewrote real papers, apparently with an LLM, and credited them to nonexistent authors at real universities. The reference lists appeared built to inflate other researchers' citation counts. Elsevier said all 15 were removed for plagiarism. Who uploaded them is unknown. [6]
- Using real people's names is also documented. An apparently AI-generated paper in a sham 'Journal of Digital Health Implementation' listed Eric Topol and other real researchers, and five of the six authors told Retraction Watch they had never seen it. Its DOI pointed to a Zenodo page that was later removed. In January 2025 the Mainichi reported a predatory journal publishing AI-generated papers under the names of three Japanese researchers without their consent. [7][8]
- Platforms have tightened their rules. Since 21 January 2026, arXiv no longer accepts an institutional email address alone as grounds for endorsing new submitters. Nature reported in May 2026 that arXiv imposes a one-year ban for incontrovertible signs of unchecked LLM output. SSRN lists 'untrue authorship claims' among the issues its integrity team screens for. Zenodo's April 2026 policy excludes AI-generated content without a verifiable basis in human research, though its spam page notes its moderators "are not domain experts." [9][10][11][12]
What it means
Treat an author name, a venue and a DOI as unverified claims, not as provenance. Before a preprint enters a corpus or supports an agent's answer, run four checks. Compare the claimed publication date with the DataCite or Crossref registration timestamp. Check that the claimed journal has a real ISSN and a Crossref record. Look for a related identifier pointing to a publisher DOI. Prefer authors linked through ORCID. Repeated LLM default names such as 'Elena Vasquez' are a cheap red flag. They are not proof, and real people share those names.
The more common failure may be the Polson pattern rather than forgery: real authors posting large volumes of lightly checked AI output. Rank by account behaviour, such as submission rate, and by verification signals, not by the reputation of the name. If you generate synthetic research text, vary the names you generate, because default name sets leave fingerprints. Researchers should monitor their own names across SSRN, ResearchGate and Zenodo. Mollick's specific complaint is plausible given the documented pipelines, but it should not be repeated as fact without the platform and identifiers.
Still open
- Which platforms and papers carry the misattributed items Mollick mentions, and have the other academics he spoke with documented their cases?
- How common is the use of real researchers' names on preprint servers, as opposed to invented ghost authors?
- Has Zenodo removed or tombstoned the 1,655 ghost-authored records and revoked their DOIs?
- How do arXiv's ban and SSRN's removals treat people listed as authors without their consent?