A chatbot can confidently cite a 2021 The Lancet study, article 14 of a law, or a line from Einstein. The citation may be flawless in form and entirely invented. For users of ChatGPT, Gemini, Claude, or Copilot, fabricated sources can slip into a court filing, journal submission, or business report.
- Courts and journals face the risk: Canary Islands judges fined lawyers over nonexistent judgments, while a The Lancet letter reported a rise in PubMed papers citing fabricated references.
- Verification depends on source type: use DOI records for studies, official legal sources for statutes and rulings, and exact-phrase searches for quotations.
- Accountability stays human: demand a DOI, law number, or URL for each claim, spot-check several, and audit all if one fails.
Courtroom and Journal Evidence
The problem is not rare or limited to students. In 2026, the Criminal Chamber of the Tribunal Superior de Justicia de Canarias fined a lawyer €840 over an appeal with “at least 25 judgments” that were nonexistent or unrelated, according to the Consejo General del Poder Judicial. The citations were quoted and emphasized, attributed to the Tribunal Supremo and the Constitutional Court. The same note recalled a February fine of €420 for another lawyer who cited up to 48 false judgments suggested by AI.
The court warned that using AI tools “does not exempt the professional from reviewing and verifying the content.”
Science shows a parallel pattern. An audit of nearly 2.5 million PubMed-indexed articles, published as a letter in The Lancet in May 2026 by Maxim Topaz’s team at Columbia University, found about one in 277 early-2026 articles cited a nonexistent work. The ratio was one in 458 in 2025 and one in 2,828 in 2023, according to Retraction Watch. The biggest jump came in mid-2024, when AI writing tools surged. DOI: 10.1016/S0140-6736(26)00603-3.
Chatbot fabrication rates vary by model and task. A 2023 Scientific Reports study asked ChatGPT-3.5 and GPT-4 for bibliographic reviews on 42 topics: 55% of GPT-3.5’s references and 18% of GPT-4’s were invented. A 2026 arXiv paper analyzed links from 10 models and research agents: 3% to 13% of URLs were invented—never present in the Wayback Machine—and 5% to 18% did not work. The newer model failed less, but no figure justifies blind trust.
The Four-Pass Verification Routine
1. Studies: begin with the DOI
The DOI is an article’s permanent identifier, resolvable through doi.org. If the AI supplies one, paste it after the resolver; a non-resolving DOI is a warning. If it resolves, confirm title, authors, and year. Without a DOI, search the exact title in quotation marks through Crossref Search, Google Scholar, or PubMed for biomedical work. Read the abstract to see whether the study supports the claim, and check Retraction Watch for retractions.
2. Laws and judgments: go official
Spanish rules have a single source in the BOE, which also publishes consolidated texts; European rules sit in EUR-Lex. Search by law number and year, then open the cited article and confirm its wording. Check the consolidated version and its date; a valid law may have been amended. For Spanish case law, the CENDOJ search tool from the Poder Judicial can locate rulings. If a supposed Supreme Court judgment cannot be found by number, date, or reporting judge, do not cite it.
3. Quotations: search the exact phrase
Famous lines attributed to historical figures were a fabrication trap long before AI. Search the quotation in quotation marks; if it appears only on unsourced quote sites, be skeptical. Ask the AI for the primary source—book, speech, or letter, with date and page—then verify that source. For recent statements by officials or executives, look for the press release or transcript. If the statement exists only in the chatbot’s reply, it does not exist.
4. Red flags: distrust “too perfect”
Several patterns should trigger alarm: a reference that fits your question exactly, real authors attached to an unrecognized title, round numbers with no date or sample, and links that fail or lead only to a homepage. Ask the AI for a DOI, law number, or URL for each assertion, spot-check several. If any fail, review everything. This mirrors detecting AI-generated fakes: what matters is whether you can reach the original source, not whether it looks authentic.
The deeper mistake is treating a chatbot as a library when it behaves like a writer: confident and widely read, but not guaranteeing that every reference sits behind the prose. The Canary Islands cases make accountability clear: responsibility for a filing belongs to the person who signs it, not the tool. That applies to classwork and reports alike. As AI agents take on more research, verification becomes a core professional discipline. If you have not opened the source, do not cite it.