Why There's Still No Tool to Verify AI-Generated Citations

The Problem

Among the hallucinations LLMs produce, one kind is the easiest to handle: a fake “pointer to external reality.” A citation to a paper that does not exist, a case-law number that was never issued, a function or package name that lives nowhere. They share one trait — they are verifiable. Just check whether the target exists in the real registry.

The problem is that no dedicated tool does that check automatically. Researchers Google references line by line, lawyers look up case law in databases by hand, and engineers wait for the IDE to underline a symbol in red. General-purpose “hallucination detectors” essentially ask an LLM again, “is this real?” — verifying a hallucination with a hallucination. What is actually needed is a deterministic verifier that checks directly against an authoritative registry, not another model.

Why Now

In May 2026, arXiv began banning authors for one year if their paper contains hallucinated citations. Such citations have risen tenfold since 2023, hitting 1 in every 277 papers, and NeurIPS 2025 saw over 100 surface in 53 papers that had cleared three or more reviewers. The meaning of the policy: verification demand now carries a price tag — “one year of no publishing.”

When a cost is vague, nobody pays for a tool. When the cost is explicit, willingness to pay for tools that reduce it appears. arXiv triggered that shift in academia, and domains with the same structure — hallucinated case law, hallucinated regulatory citations, hallucinated APIs — in law, compliance, and software are likely to follow. Whoever builds the verifier first becomes that domain’s standard.

How to Build It

The MVP picks just one domain. Academic citations have the lowest verification difficulty — the arXiv, Crossref, and PubMed APIs are already public, and the ground truth is unambiguous. The flow is simple.

flowchart LR
    A[LLM-generated document] --> B[Extract references]
    B --> C[Match against real registry]
    C --> D{Target exists?}
    D -->|Exists| E[Pass]
    D -->|Missing| F[Flag fake reference]
    F --> G[Pre-submission report]

The core technology is not an LLM but two things. First, a parser that extracts references accurately from the document (citation formats, footnotes, hyperlinks). Second, matching logic that checks against domain registry APIs and catches “similar but different” entries. Ship it as a gate right before submission — a word-processor plugin, a CI step, a submission-form check — and the tool embeds naturally into the workflow.

Success Criteria (optional)

The key assumption is that verifiable hallucinations alone create willingness to pay. Even without semantic verification (a real source but a false conclusion), if existence checking alone lets a lab avoid an arXiv ban, research labs and law firms will pay. Attach the gate free to the first ten labs or firms, and measure how many “fake references that would have slipped through” it catches. More than one per submission is a basis for converting to paid.