Newshot

AI is ditching hallucinated references — and other bad habits

Summary

Nature, Published online: 19 November 2026; doi:10.1038/d41586-026-03753-1 An analysis finds that many undesirable LLM behaviours dissipate after about 16 months.

Original text(via Nature)
A collaged image of a person with an old fashioned television on their head with the words AI Hallucination shown on the screen. In the background is binary code on a green colour.
A collaged image of a person with an old fashioned television on their head with the words AI Hallucination shown on the screen. In the background is binary code on a green colour.

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 The analysis, described in a blog post, has not been peer reviewed.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 found1 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

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

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

Walters, W. H. & Wilder, E. I. Sci. Rep. 13, 14045 (2023).

Article PubMed Google Scholar

Download references

Reprints and permissions

Subjects

  • Machine learning
  • Computer science

Latest on:

  • Machine learning
  • Computer science
  • Are you a budding science sleuth? Take our plagiarism quiz to test your skills Career Column 18 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

AI must be held accountable — value statements won’t protect human dignity

World View 18 NOV 26

  • Hypothalamic volume and cell-type changes in ageing and Alzheimer’s disease Article 18 NOV 26

Hypothalamic volume and cell-type changes in ageing and Alzheimer’s disease

Article 18 NOV 26

  • Are you a budding science sleuth? Take our plagiarism quiz to test your skills Career Column 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

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

Retractions of old papers are on the rise: when should bad science be culled?

News Feature 11 NOV 26

  • An AI-native experimental laboratory for autonomous biomolecular operations Article 04 NOV 26

An AI-native experimental laboratory for autonomous biomolecular operations

Article 04 NOV 26

  • History’s long-secret codes are being cracked by amateurs using AI News 16 OCT 26

History’s long-secret codes are being cracked by amateurs using AI

News 16 OCT 26

  • Margaret Hamilton obituary: computer scientist who pioneered software engineering and helped us get to the Moon Obituary 15 OCT 26

Margaret Hamilton obituary: computer scientist who pioneered software engineering and helped us get to the Moon

Obituary 15 OCT 26

Jobs

  • 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

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 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

2027 Biology Assistant/Associate Professor

job board:https://apply.interfolio.com/56017/positions

Wenzhou, Zhejiang (CN)

Wenzhou-Kean University

  • Global Leading Talent Recruitment at College of Bio-X, Huazhong Agricultural University Join HZAU's global faculty team to advance research with competitive benefits. Wuhan, Hubei (CN) Huazhong Agricultural University (HZAU)

Global Leading Talent Recruitment at College of Bio-X, Huazhong Agricultural University

Join HZAU's global faculty team to advance research with competitive benefits.

Wuhan, Hubei (CN)

Huazhong Agricultural University (HZAU)

  • Postdoctoral Research Fellow Postdoctoral fellowship in vascular biology at UT Southwestern, studying the endothelial basis of cardiometabolic disease. Dallas, Texas (US) UT Southwestern

Postdoctoral Research Fellow

Postdoctoral fellowship in vascular biology at UT Southwestern, studying the endothelial basis of cardiometabolic disease.

Dallas, Texas (US)

UT Southwestern

The text above was extracted automatically from the source page and may miss images, tables or code blocks. The original is authoritative.

Published by Nature
This site only summarises and aggregates; copyright stays with the original author. Headlines and summaries follow the source — where they differ, the original wins.
Read the original ↗