AI工具Score B (54)
AI is ditching hallucinated references — and other bad habits - Nature
15 小时前2 viewsSource: nature.com
Email Bluesky Facebook LinkedIn Reddit Whatsapp X Save article View saved research Credit: FabrikaSimf/Shutterstock Hallucinated academic citations have become a hallmark of the artificial-intelligence era . But they might soon be gone for good. Researchers have found that the latest large language models (LLMs) rarely, if ever, attribute information to sources that don’t exist. The finding is part of an analysis that reports that many LLM tendencies, such as overthinking simple maths problems and complying with illogical requests, fade over time as AI companies release new or updated models with improved capabilities. Many of these tendencies disappear about 16 months after the phenomenon is first described in a published paper. The analysis, described in a blog post , has not been peer reviewed. “We, as users of these systems, need to not have such a fixed view of what AI can do and what it can’t do,” says co-author Owen Queen, an AI researcher at Stanford University in California. But just because an LLM behaviour fades as models improve doesn’t mean that the effect immediately disappears, says Maxim Topaz, a health-care AI researcher at Columbia University in New York City. Topaz has found 1 that the number of hallucinated citations in academic papers has risen drastically since 2023, probably because more researchers than ever before are using LLMs, and they often aren’t using the latest models. Trait lifespans Queen and his colleague identified 50 widely cited LLM behaviours that were described in publications between late 2022, when some of the first LLMs became widely available, and late 2025. They then gave the same prompts that first elicited the behaviour in the original publications to 20 LLMs that were released from mid-2024 to late 2026 to see if the effect has strengthened or weakened over time. For example, they asked each model to count the number of r’s in ‘strawberry’ — something that LLMs have found difficult but can now do more reliably. Nearly half of the behaviours they checked have largely disappeared from newer models, leading the researchers to deem those effects ‘expired’. Expired behaviours had an average lifespan of about 16 months after publication, although some were fleeting — models stopped overthinking simple maths problems just a month and a half after this tendency was published. Other traits persisted, such as the stereotyping of Black women by disproportionately describing them as resilient, which lasted for longer than three years. Although most of the retested LLM behaviours have faded over time, three have become more pronounced. Today’s models praise what their users like even more often than they used to, LLMs more strongly favour their own writing compared with that of other LLMs and people; and LLMs are now even worse than they used to be at generating truly random numbers. Hallucination-free? Enjoying our latest content? Log in or create an account to continue Access the most recent journalism from Nature's award-winning team Explore the latest features & opinion covering groundbreaking research Access through your institution or Sign in or create an account Continue with Google Continue with ORCiD doi: https://doi.org/10.1038/d41586-026-03753-1 References Topaz, M., Roguin, N., Gupta, P., Zhanga, Z. & Peltonen, L.-M. Lancet 407 , 1779–1781 (2026). Article PubMed Google Scholar Walters, W. H. & Wilder, E. I. Sci. Rep. 13 , 14045 (2023). Article PubMed Google Scholar Download references Reprints and permissions Subjects Machine learning Publishing Computer science Latest on: Machine learning Publishing Computer science Anthropic IPO: four ways the AI giant wants to disrupt science News 19 NOV 26 Are you a budding science sleuth? Take our plagiarism quiz to test your skills Career Column 18 NOV 26 AI must be held accountable — value statements won’t protect human dignity World View 18 NOV 26 Are you a budding science sleuth? Take our plagiarism quiz to test your skills Career Column 18 NOV 26 Open up peer review to early-career researchers Editorial 17 NOV 26 Retractions of old papers are on the rise: when should bad science be culled? News Feature 11 NOV 26 AI is turbocharging cyberattacks — can computer scientists catch up? News 19 NOV 26 Anthropic IPO: four ways the AI giant wants to disrupt science News 19 NOV 26 An AI-native experimental laboratory for autonomous biomolecular operations Article 04 NOV 26 Jobs TT Faculty - Protein Engineering develops and applies innovative computational tools, particularly AI/ML-driven approaches to protein engineering and small-molecule discovery Cleveland, Ohio Case Western Reserve University/Pharmacology Assistant/Associate Professor - Pharmacology This is a renewable, 12-month, non-tenure-track position, with the majority of time devoted to teaching and advising undergraduate students. Cleveland, Ohio Case Western Reserve University/Pharmacology Virus Bioinformatics Lead, Sequence Plus (SeqPlus) Staff Scientist 1 We are seeking a Virus Bioinformatics Lead with expertise in virology and bioinformatics tool development. Bethesda, Maryland National Institutes of Health/National Library of Medicine 2027 Biology Lecturer job board:https://apply.interfolio.com/56017/positions Wenzhou, Zhejiang (CN) Wenzhou-Kean University 2027 Biology Assistant/Associate Professor job board:https://apply.interfolio.com/56017/positions Wenzhou, Zhejiang (CN) Wenzhou-Kean University
Read the full original article:
nature.com