Nobody checks whether reference 17 exists

Preprint servers now ban authors for citing papers that don't exist. Checking that a reference is real is still a manual job. That gap is the opportunity.

The pain in one sentence

Researchers, editors, and professors are now personally liable for citations that point at papers which do not exist — and the only way anyone checks today is by hand, one reference at a time.

The receipts

Every claim below links to something you can go and check yourself.

• arXiv now bans you for a year over a fake reference. In May 2026 the preprint server announced an immediate one-year submission ban where it finds incontrovertible evidence of unchecked LLM output — hallucinated references are its first listed example. Coverage in Nature and The Next Web; Times Higher Education quotes integrity experts calling it "welcome but unenforceable" — unenforceable because checking is manual.
• The scale is not hypothetical. A study co-authored by arXiv's founder counted nearly 150,000 hallucinated references in preprints posted in 2025 alone, and a Columbia University team writing in The Lancet found fabricated citations up roughly twelvefold since 2023 (both reported in Nature; accessible summary at SMU Libraries).
• The failure mode is designed to fool a human skimmer. In a Deakin University audit of GPT-4o-drafted literature reviews, about one citation in five was entirely fabricated and 56% were fake or erroneous — and of the fabricated citations that carried DOIs, 64% resolved to real papers about something else, so clicking the link proves nothing (PsyPost).
• Universities are responding with friction, not verification. Princeton's faculty voted in May 2026 to put a proctor in every exam room — ending a 133-year-old honour-code tradition — after roughly 30% of surveyed seniors admitted at least one violation (Times Higher Education, University Herald).
• Even grading is being rationed. Harvard's faculty voted in May 2026 to cap A grades at roughly one fifth of each undergraduate course from 2026–27, an anti-inflation measure passed in the middle of the AI-coursework surge (The Boston Globe, Inside Higher Ed).

Why it persists

The tools that exist answer a different question. Turnitin-class detectors estimate whether prose was written by AI — a probabilistic guess that generates contested accusations and false positives. Reference managers like Zotero format citations; they never ask whether the cited paper exists. So institutions reach for what policy can deploy this semester: proctors, blue books, grade caps.

Meanwhile the question that actually matters — does reference 17 exist, and is it the paper the author says it is? — is deterministic and checkable against free public indexes (Crossref, OpenAlex, PubMed, arXiv). Nobody built the checker because until months ago a phantom reference cost only embarrassment. Now it costs a year of your publishing life, and the penalty regime arrived faster than the tooling.

The wedge

A pre-submission citation-integrity check. Paste a bibliography or upload a manuscript; every reference is cross-checked against Crossref, OpenAlex, PubMed, and arXiv; out comes a flag list — references that exist nowhere, DOIs that resolve to a different paper (the 64% failure mode above), and works that have been retracted since the author last looked.

No machine-learning guesswork, no accusation engine — a deterministic existence check with an explainable answer for every line. The public APIs are free; a working single-bibliography checker is weeks of solo work, not months.

The honest catch: an existence check is also easy for an incumbent to bolt on later, and the deeper problem — a real paper cited for something it never said — is much harder. The edge is being there at the compliance moment with the loop below, not the difficulty of the tech.

How the buyer arrives (no cold outreach)

The buyer here reaches you, at the moment of fear — submission day. Three organic routes, no cold email, no DMs:

1. The check is the channel. A flagged bibliography gets forwarded to every co-author and the supervising PI before anyone hits submit. Each report recruits the next users by itself.
2. Search intent already exists. People are typing the panic into search engines — the arXiv ban, "check if a DOI is real", each journal's new AI-citation policy. A plain explainer page per policy, each ending at the free check, captures intent that costs nothing.
3. This newsletter and the public board. The receipts above are the story; the story travels in the places where academics are already arguing about this in public.

The give-away

You do not need anyone's product to protect yourself this week. The 15-minute manual audit:

1. Every DOI in your bibliography → paste into doi.org. Dead link = red flag. Resolves = read the landing page and confirm the title and authors actually match your citation (remember: 64% of fakes with DOIs point at real, unrelated papers).
2. No DOI → exact-phrase search of the title in Crossref search and Google Scholar. Zero hits anywhere = treat as phantom.
3. Confirm the journal exists and the volume/issue/year are plausible.
4. Spot-check the claimed authors' own publication lists for the paper.
5. Anything you cannot verify in two minutes gets pulled or replaced. arXiv's standard is now your standard.

Forward this to the co-author who "had ChatGPT tidy up the references". They need it more than you do.