Can AI make up sources?
Wretify is a clean, minimalist writing app for nonfiction writers, built for focus — with an AI that has read everything you've written and refuses to write it for you.
The short answer
Yes, and often. When researchers checked 636 citations produced by ChatGPT, 55% of GPT-3.5’s and 18% of GPT-4’s were fabricated, and many of the real ones had errors (Walters and Wilder, 2023). A language model writes what a citation usually looks like; it does not look one up. Every citation in your book has to come from a real record that you have opened.
How often do AI tools fabricate citations?
In the 2023 study, researchers asked ChatGPT for short literature reviews on 42 topics and checked all 636 citations: 55% of GPT-3.5’s were fabricated against 18% of GPT-4’s, and among the real ones, 43% and 24% had substantive errors (Walters and Wilder, 2023). In a 2025 study of mental-health reviews written by GPT-4o, 20% of 176 citations were fabricated and 45% of the real ones contained errors — “nearly two-thirds” fabricated or inaccurate (Linardon and colleagues, 2025).
Has it got better?
Newer models do better on some tasks, but the problem has not disappeared, and results vary by topic and question. The safe assumption for a writer is the old one: a citation you have not opened is a citation you have not checked.
Have made-up sources reached published work?
Yes. In December 2025, Retraction Watch reported that an independent check of one paper in the Journal of Academic Ethics found 19 of its 29 references appeared to be fabricated. That is a single case, not a measured rate — but it shows fabricated sources can pass through review.
How do I stop AI from inventing my sources?
- Never cite a source you have not opened and read, at least the part you cite.
- Find sources through real databases — OpenAlex, Crossref, Google Scholar, your library — not from a chatbot’s memory.
- Check the DOI or link resolves to the paper you are citing.
- Keep each source attached to the sentence it supports, so it cannot drift to another claim.
How does Wretify handle sources?
Wretify’s AI never writes a citation from memory. It searches 250 million real records in OpenAlex, the reference is written from the record it retrieved, a quoted line is kept only if it appears word for word in the source, and when nothing supports a claim it says so. How references and citations work in Wretify →
Sources
- Walters & Wilder (2023), Fabrication and errors in the bibliographic citations generated by ChatGPT, Scientific Reports 13 — doi.org/10.1038/s41598-023-41032-5
- Linardon et al. (2025), Citation fabrication in mental health research using LLMs, JMIR Mental Health 12 — doi.org/10.2196/80371
- Retraction Watch (2 December 2025), The case of the fake references in an ethics journal — retractionwatch.com/2025/12/02/fake-references-chatgpt-journal-academic-ethics-springer-nature-whistleblowing/
- OpenAlex — openalex.org/
Updated .