Hallucinated Citations Are Polluting the Scientific Literature. What Can Be Done?

Why It Matters

Fake citations undermine research reproducibility and can damage journal credibility, prompting a need for systematic detection across the scholarly ecosystem.

Key Takeaways

• AI models fabricate plausible but nonexistent citations.
• “Frankenstein” references blend real fragments with invented data.
• Hallucinated DOIs undermine scholarly verification processes.
• Peer review struggles to detect sophisticated fake references.
• Tools like Grounded AI aim to screen and flag fakes

Pulse Analysis

The rapid adoption of large language models such as ChatGPT has transformed academic drafting, but it has also introduced a subtle form of misinformation: fabricated references. Researchers and legal professionals have inadvertently inserted these invented citations into manuscripts, trusting the models’ surface realism. As a recent Nature investigation with Grounded AI reveals, the problem is escalating; entire reference entries can mimic authentic titles, author lists, and journal details, creating a veneer of credibility that deceives even seasoned reviewers. This erosion of citation integrity threatens the reproducibility cornerstone of scientific discourse.

Grounded AI’s researchers label these fabrications “Frankenstein” citations because they splice genuine fragments with invented metadata. A fake DOI may point to a non‑existent article, while volume and page numbers align with real journals, making manual checks labor‑intensive. Traditional peer‑review pipelines rely on author honesty and quick visual scans, which are insufficient against algorithmic precision. Consequently, journals risk publishing articles that cite non‑existent work, inflating citation metrics and misleading subsequent studies that build on phantom literature. Institutions therefore face reputational risk when such errors slip through.

To curb the spread of bogus references, publishers are deploying AI‑driven screening tools that cross‑verify DOIs, author identities, and journal archives in real time. Digital Science’s scientometrics team advocates mandatory reference validation steps before manuscript acceptance, while funding agencies consider penalizing papers with unverifiable citations. Authors, too, must adopt a verification mindset, treating every generated reference as a hypothesis to be tested. As detection technologies mature, the academic ecosystem can restore confidence in the citation chain, preserving the credibility essential for innovation and public trust.