Attention Is All You Need Has Five Fake Twins on Crossref

1. The discovery

Query Crossref for “Attention Is All You Need” — the 2017 Vaswani et al. paper that introduced the Transformer, cited more than 140,000 times^[1] — and five separate records come back. Same title, same eight authors, byte for byte. All five list the publication year as 2025. All five carry a DOI under the prefix 10.65215, registered not to a computer-science venue but to an entity named the Shenzhen Medical Academy of Research and Translation, via a Chinese preprint host.

We checked each one directly rather than taking the search result at face value: the DOIs resolve, the Crossref records are real and indexed, not an artifact of our own tooling. The actual 2017 NeurIPS paper has no Crossref DOI at all — NeurIPS does not systematically register one for its papers, unlike venues indexed through the ACL Anthology. A title search against Crossref alone therefore has nothing genuine to return; the five fabrications are the only candidates on offer.

2. Why this matters beyond one paper

Anything that verifies citations against registry data is exposed to this mechanism — ours, or anyone else's. If the correct source does not answer, and every registry has its own coverage gaps and occasional downtime, a single spam duplicate can be the only voice in the room. Agreement across independent sources is the meaningful signal, not any one source's word in isolation; a claim resting on exactly one unfamiliar registry warrants deliberate skepticism before it is treated as settled.

3. The pattern is not rare

We selected ten of the most-read AI papers on Hugging Face's Daily Papers listing this week — real preprints with real citation lists, not curated for this piece — and ran their actual .bib/.bbl files, taken directly from each paper's own arXiv source, through the same thirteen-registry checks the workspace uses. 831 references in total; more than 300 flagged for closer review.

A meaningful share of that is noise rather than findings, and it is worth stating plainly rather than folding it into a bigger number: Semantic Scholar's keyless tier shares one rate-limited pool with every anonymous caller on the internet, and it throttled heavily partway through this run. A reference that would ordinarily clear with two independent sources in agreement can read as unconfirmed for reasons that have nothing to do with the citing paper. Real, specific, independently verifiable problems were nonetheless present throughout:

• One paper's citation year read 1609 — not a year, but a fragment of an AAAI DOI (10.1609/…) that some bibliography-generation tool had evidently mistaken for one. The identical pattern, digits lifted from an identifier rather than an actual year, appeared independently in three separate papers.
• Another paper cited a diffusion-model paper by DOI. That DOI resolves to an unrelated paper on human-factors modelling for autonomous driving, credited to four different authors with no connection to diffusion models.

4. OpenAI's and Meta's own citations are broken too

This is not confined to fast-moving preprints from smaller teams. A paper in the sample cites GPT-4o and Llama 3 as baseline models; we pulled the official arXiv BibTeX export both companies publish for their own papers, directly from arXiv's own /bibtex/ endpoint, not a third-party citation manager.

OpenAI's own GPT-4o System Card export reads, in full:

A bare colon, listed as a co-author, in the citation OpenAI itself publishes for its own paper.

Meta's Llama 3 Herd of Models export is broken a different way — real people's names split in half:

One author, Guangyi Zhang, becomes two nonexistent people: “Guangyi” and “Zhang.” The same fault recurs later in the same list for Yu Wang. Somewhere in the tooling that generated these official citations, the literal string “and” landed inside a name rather than only between names — and it shipped, from two of the best-resourced labs in the industry, into every paper that copies the citation as published.

The corruption does not stop at the manuscript. DataCite, the registry to which arXiv reports, has ingested the identical broken author lists as the record of standing. A tool that treats a registry's word as ground truth would confirm both entries as correct, because the fault originates at the source, not somewhere downstream of it.

5. What this does not mean

It does not mean citation verification is a lost cause, nor that these registries are unreliable on the whole — for most of what they cover, most of the time, they are not. It means the registry says so and it is true are different claims, and worth treating as different claims regardless of which tool is doing the asking.

To check what your own bibliography's actual reference file looks like under the same procedure: strictcite.com. The free tier requires no card.