For partners 
Artificial intelligence is increasingly helping students and scientists write academic texts, which is creating a serious problem in the community: according to a study by the journal Nature, papers are increasingly containing references to works that do not actually exist and never did. An example from a sample of 100 publications: at least one non-existent reference was found in 65 texts!
The study was conducted jointly with the company Grounded AI. First, 4,000+ scientific articles published in 2025 by major publishers were analyzed; a special tool flagged texts with suspicious references, and then experts manually checked some of them. The result turned out to be unexpectedly high: many sources simply cannot be found in any scientific database! The leading theory is the use of neural networks in preparing the materials.
Sometimes this happens due to author inattentiveness, but often the reason is different. AI can fabricate sources to make scientific text look convincing, and such errors are not always obvious, because a fake reference looks plausible but leads nowhere. If the author of an article does not double-check the authenticity of a source, they risk significantly lowering the level of expertise of their material.
To test the theory, the researchers conducted a separate test with the GPT-4o model. It was asked to write several scientific reviews — as a result, nearly 20% of the references turned out to be completely fabricated, and another roughly 45% of real sources contained serious errors, such as incorrect identifiers. It turns out that even genuine references can prove to be inaccurate and unreliable!
The AI problem extends beyond a handful of scientific articles — according to researchers' estimates, we may already be talking about tens of thousands of scientific works containing gross errors, and this directly affects the quality of science: it becomes harder to verify information, trust in publications declines, and errors can propagate further down the research chain. The Nature editorial team urges authors of publications to more carefully verify their reference lists, not to rely entirely on texts generated by neural networks, and to strengthen editorial and peer-review checks.
