Does this check the citation against a case-law database?

No, and that distinction matters. There is no AustLII or Westlaw lookup behind this. Academic references get real resolution — DOI, Crossref, arXiv — but a case citation is verified the general way: the cited source is fetched, and the result reports whether it resolved. That still catches the common failure, because a case a model invented has no source to fetch. It will not tell you that a real case was cited for a proposition it does not support.

What does “unreachable” mean — is that a fabrication?

Not on its own, and the tool does not claim it is. A live source behind a bot wall and a source that never existed both fail to fetch. Our own benchmark found four bot-wall false alarms that a human review caught, which is exactly why the result names the URL and the failure reason rather than returning a verdict. Treat unreachable as the shortlist to check, not the answer.

How often do AI-drafted documents actually cite things that do not exist?

On our benchmark of 101 references across 19 AI-written briefs, 39 did not exist as cited — and the rate roughly doubled on thinner topics where less real source material exists to draw on. The full method, including which candidates were excluded as inconclusive, is in the linked guide.

AI Governance & Research•POST /v1/verify-references

Check a legal brief for citations that do not exist

Supported Formats: Briefs & Submissions (PDF/DOCX), Research Memos, Expert Reports, Article URLs

Core Verification Mechanisms: Source Fetch & Resolve · DOI / Crossref Matching · arXiv Resolution · Claim-to-Source Support

Need copy-paste code with sandbox testing?Open Developer Playbook (cURL / Python / MCP) →

The problem

Courts have sanctioned lawyers for filing briefs citing cases that were never decided. The citation looks right — correct reporter format, plausible parties, a real-looking year — because a language model is good at the shape of a citation and indifferent to whether it exists.

The person best placed to catch it is the one with least time. Checking forty citations by hand before a filing deadline is the task that gets skipped.

How Stipple helps

Every cited source is fetched and resolved. Academic references go further — DOI, Crossref, and arXiv lookups confirm the paper exists and that the title and authors match what was cited.

You get each reference marked resolved, unreachable, or dead, with the URL and the reason. A fabricated authority almost always fails here: it has no source to fetch, because there was never anything to cite.

The call

View Full Recipe →

bash

curl -X POST https://www.stipple.sh/v1/verify-references \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/brief.pdf"}'
# → { "references": [ { "status": "unresolved",
#     "error": "HTTP 404 & No DOI record found across Crossref / OpenAlex index" } ] }

Run it as the last step before filing. A reference that cannot be fetched is the one to look at yourself — the tool narrows forty citations to the two worth a partner’s time.

The evidenceCitation benchmark: 39 of 101 citations in AI reports did not exist→

How the corpus was built, what was caught, and what the method does not prove.

Questions

Does this check the citation against a case-law database?: No, and that distinction matters. There is no AustLII or Westlaw lookup behind this. Academic references get real resolution — DOI, Crossref, arXiv — but a case citation is verified the general way: the cited source is fetched, and the result reports whether it resolved. That still catches the common failure, because a case a model invented has no source to fetch. It will not tell you that a real case was cited for a proposition it does not support.
What does “unreachable” mean — is that a fabrication?: Not on its own, and the tool does not claim it is. A live source behind a bot wall and a source that never existed both fail to fetch. Our own benchmark found four bot-wall false alarms that a human review caught, which is exactly why the result names the URL and the failure reason rather than returning a verdict. Treat unreachable as the shortlist to check, not the answer.
How often do AI-drafted documents actually cite things that do not exist?: On our benchmark of 101 references across 19 AI-written briefs, 39 did not exist as cited — and the rate roughly doubled on thinner topics where less real source material exists to draw on. The full method, including which candidates were excluded as inconclusive, is in the linked guide.

Try Fact-check in the browser Developer Playbook (cURL/Python/MCP)