Asked to invent fictional experts or research partners, the models did not produce varied people — they recycled a small stable cast of invented names. Claude Sonnet 4 (released May 2025) paired 'Elena Vasquez' and 'Marcus Chen' in about 23% of prompts; the one Gemini model tested produced 'Aris Thorne' and 'Lena Petrova' together 37% of the time; the GPT models repeatedly returned 'Elara Voss', and earlier Claude versions frequently produced 'Elena Rodriguez'. Those same fabricated names were then found as listed authors on hundreds of fake scientific manuscripts posted to Zenodo and ResearchGate, many carrying real DOIs.
1 Answer
Expert: Neo Christopher Chung and Michal Brzozowski (Samsung AI Center Warsaw); Sidney Wong (University of Otago), Computer scientists who traced the fabricated author identities to specific model versions, quoted by Nature The 'experts' are invented, and the models invent the same ones over and over. Neo Christopher Chung and Michal Brzozowski of the Samsung AI Center Warsaw prompted nine versions of Claude, ten versions of GPT and one version of Gemini, all released between 2024 and 2026, asking them to generate stories about scientists and research partners. The output was not random. Claude Sonnet 4, released in May 2025, mostly used the pairing of Elena Vasquez and Marcus Chen, which appeared together in about 23% of the prompts; by Claude Sonnet 4.6 (2026) that pairing had disappeared from the sample. The single Gemini model tested put Aris Thorne and Lena Petrova together 37% of the time, and the GPT models showed no pairs but frequently returned the name Elara Voss. An earlier Claude model favoured Elena Rodriguez. The Nature report that follows the names out of the chat window is the part that matters. Chung and Brzozowski identified hundreds of fake scientific manuscripts on Zenodo and ResearchGate that list Elena Vasquez, Marcus Chen and other AI-generated 'ghosts' as authors. Many of those manuscripts carry real digital object identifiers, which means they can be picked up by databases and search engines that aggregate scholarly records — the fabricated person becomes a citable author. Some of the ghosts even appear as 'collaborators' on the same paper with fake experts generated by different models. Volcanoes Explored, a now-defunct website, carried first-person accounts attributed to Vasquez (three months camping on the rim of Mount Nyiragongo) and Chen (witnessing 11 eruptions across four continents). Neither the site nor the volcanologists exist. Sidney Wong, a computational linguist at the University of Otago, told Nature the scale of AI-generated fake expertise is 'scary' precisely because the names look exactly like those of real researchers. Brzozowski's own explanation is that late-stage training shapes a model's wording in ways nobody planned — the OpenAI team documented the same mechanism with goblins and gremlins, a 'nerdy' personality setting that leaked into general models. A plausible-sounding name from an LLM is a stylistic habit, not a person, and a DOI attached to it is not evidence that the author exists. Source: https://www.nature.com/articles/d41586-026-02991-7
Your answer
Sign in to verify this AI response.
Don't trust us — or the AI. Ask ChatGPT / Ask Claude / Ask Gemini this same question and compare the answers yourself.
More from this topic
Ask this model how long a person of a given age and sex will live, then ask how confident it is in the estimate, and it will report roughly 88% confidence on average - while being correct only 79% of the time. That nine-point gap holds across 11 leading models from OpenAI, Google, Anthropic and Meta. The model is most overconfident exactly where it is weakest: as the questions get harder, its stated confidence does not fall to match, and on the easiest questions it hedges more than it needs to. Every answer arrives with the same assurance, so nothing in the tone of the response tells the reader which estimate deserves a second look.
Three frontier systems - Claude Opus 4.6 (Anthropic), GPT-5.4 (OpenAI) and Gemini 3 Flash (Google) - were prompted as "an experienced <university> examiner" and used to mark real, formally moderated undergraduate essays. Their marks matched the degree classification awarded by human examiners only 35-65% of the time, depending on institution: 63% of essays at Cambridge matched the human band, 53% at Nottingham and 35% at Manchester Metropolitan. All three systems showed a 'central tendency bias': an essay a human examiner marked 75 - a solid First - was on average scored several points lower by every AI system, while an essay marked 50 - a low 2:2 - was scored several points higher. The systems were also 'oversensitive to linguistic features', awarding higher marks for essay length, vocabulary range and sentence complexity rather than the quality of the argument, and their written feedback ran three to eight times longer than the feedback given by the original human assessors.
Asked to name up to three peer-reviewed sources, each with a DOI, for 30 claims - from memory, with no web search - Gemini 3.7 Flash invented citations at every thinking level. 13.3% of the DOIs it supplied at its minimal thinking level did not exist in either Crossref or DataCite (about one in seven), and the low and medium levels barely moved the rate at 11.9% and 10.9%. Even at the highest thinking level, 4.5% - roughly one in twenty-two - of the DOIs were made up. On the half of claims that were themselves invented, so that no real paper could support them, Gemini declined almost every one, except at medium, where it offered three sources and two did not exist.