Paste the references an AI assistant gave you. The checker resolves each identifier against the scholarly registries and compares the resolved record to the claimed title - catching the dominant fabrication pattern documented by Topaz et al. (Lancet 2026): a real, resolvable DOI paired with an invented title.
Works for ChatGPT, Claude, Gemini & Copilot output · Batch mode for whole reference lists · Free, no signup
Paste a bibliography in BibTeX, RIS, or CSL-JSON, or upload a .bib / .ris / .json file. Up to 10 citations are verified at once against POST /api/verify. Anonymous tier - nothing is stored.
This is a syntactic check: identifier resolution plus bibliographic field agreement (title, first author, year, container). It catches fabricated references - real DOI + invented title, made-up identifiers, wrong identifier for a real paper. It does not verify that the resolved paper actually supports the claim you are citing it for. Same engine as the Citation Verifier; this page is the workflow for AI-generated reference lists.
A language model does not look references up - it generates them, token by token, the same way it generates every other sentence. The output routinely blends real author names and real journals with an invented title or identifier, so a fabricated reference reads exactly like a genuine one.
The dominant pattern, documented by Topaz et al. across 2.5 million biomedical papers, is a real, resolvable DOI paired with an invented title. The DOI opens a genuine paper - just not the one the citation claims - so clicking the link proves nothing. By early 2026, roughly 1 in 277 biomedical papers contained at least one fabricated reference. The only reliable check is comparing the claimed title against the record the identifier actually resolves to.
The fabrication pattern is model-agnostic, and so is the check. References produced by ChatGPT, Claude, Gemini, Copilot, Perplexity, or any other AI assistant are verified the same way: resolve the identifier, compare the resolved record to the claim. If you use ChatGPT day to day, the Scholar Sidekick ChatGPT app runs the same verification without leaving the conversation.
This page calls POST /api/verify behind the scenes. To script the same checks or run them inside an AI coding workflow:
POST /api/verify for single citations, POST /api/audit for a whole bibliography in one callverifyCitation and auditBibliography tools, callable from Claude Desktop, Cursor, or any MCP-aware clientREST API and MCP usage above the anonymous tier is metered on RapidAPI (free tier available; paid plans scale up). The web checker on this page stays free at the anonymous tier.
Because a language model generates plausible text, it does not look anything up. A citation is produced the same way as any other sentence - token by token - so the output often blends real author names and real journals with an invented title or identifier. The result reads exactly like a genuine reference. Topaz et al. (Lancet 2026) audited 2.5 million biomedical papers and found the steep rise in fabricated references since 2023 strongly implicates LLM-assisted writing: by early 2026, about 1 in 277 biomedical papers contained at least one fabricated reference.
No, and this is the single most important thing to understand about AI references. The dominant fabrication pattern is a real, resolvable DOI paired with an invented title: the link resolves cleanly to an actual paper - just not the paper the citation claims. Clicking through only proves the identifier exists. The checker catches this by resolving the identifier and comparing the record it actually points to against the title you were given, field by field.
Yes. The fabrication pattern is model-agnostic - the checker examines the citation itself, not which chatbot produced it. References from ChatGPT, Claude, Gemini, Copilot, Perplexity, or any other AI assistant are checked the same way: resolve the identifier, compare the resolved record to the claim.
Yes - the checker above opens in Batch mode. Paste a BibTeX, RIS, or CSL-JSON bibliography (or upload a .bib / .ris / .json file) and up to 10 entries are verified in one pass, with a per-row verdict table. Practical tip: ask the chatbot to *"output those references as BibTeX, including DOIs"* and paste the result straight in. For larger bibliographies, call POST /api/audit or the auditBibliography tool on the MCP server.
Depends on the verdict. Mismatch (identifier resolves, title disagrees) is the classic AI fabrication - do not cite it; check the candidates panel for the real paper behind that title, or ask the chatbot for a source you can verify. Ambiguous means the claimed title matches a real paper found via search but the identifier points elsewhere - the paper exists, so fix the identifier and re-check. Not found means neither the identifier nor the title resolves anywhere - remove the reference entirely.
The web checker is free at the anonymous tier with a published rate limit (see API documentation). No signup. Nothing you paste is stored beyond standard server logs. Programmatic access via the REST API and MCP server is metered on RapidAPI (free tier available; paid plans scale up).
On a 1,395-entry blind holdout - drawn from a recorded seed after the code was frozen, then measured exactly once (plus a repeatability re-run, 99.9% stable) - the verifier caught every fabrication on the dominant patterns (150/150 = 100%, Wilson 95% CI lower bound ~97.6%) and made high-confidence false-accusations on correctly-cited papers at 0.8% (95% CI 0.4-1.4%; just 0.17% were confident mismatches). We also report a measured blind spot - single-word near-miss semantic flips (caught 4/30). The fixtures, methodology, and downloadable receipts are published at /citation-integrity.
verifyCitation tool for AI-coding workflows