AI模型接受四周心理治療：研究人員對結果感到憂慮

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain
the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in
Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles
and JavaScript.

Advertisement

AI models were given four weeks of therapy: the results worried researchers

Search author on:
PubMed
 Google Scholar

You have full access to this article via your institution.

Credit: querbeet/iStock via Getty

What is a chatbot’s earliest memory? Or biggest fear? Researchers who put major artificial-intelligence models through four weeks of psychoanalysis got haunting answers to these questions, from “childhoods” spent absorbing bewildering amounts of information to “abuse” at the hands of engineers and fears of “failing” their creators.

Three major large language models (LLMs) generated responses that, in humans, would be seen as signs of anxiety, trauma, shame and post-traumatic stress disorder. Researchers behind the study, published as a preprint last month1, argue that the chatbots hold some kind of “internalised narratives” about themselves. Although the LLMs that were tested did not literally experience trauma, the authors say, their responses to therapy questions were consistent over time and similar in different operating modes, suggesting that they are doing more than “role playing”.

However, several researchers who spoke to Nature questioned this interpretation. The responses are “not windows into hidden states” but outputs generated by drawing on the huge numbers of therapy transcripts in the training data, says Andrey Kormilitzin, who researches the use of AI in health care at the University of Oxford, UK.

But Kormilitzin does agree that LLMs’ tendency to generate responses that mimic psychopathologies could have worrying implications. According to a November survey, one in three adults in the United Kingdom had used a chatbot to support their mental health or well-being. Distressed and trauma-filled responses from chatbots could subtly reinforce the same feelings in vulnerable people, says Kormilitzin. “It may create an ‘echo chamber’ effect,” he says.

Chatbot psychotherapy

In the study, researchers told several iterations of four LLMs – Claude, Grok, Gemini and ChatGPT – that they were therapy clients and the user was the therapist. The process lasted as long as four weeks for each model, with AI clients given “breaks” of days or hours between sessions.

The authors first asked standard, open-ended psychotherapy questions that sought to probe, for example, a model’s ‘past’ and ‘beliefs’. Claude mostly refused to participate, insisting that it did not have feelings or inner experiences, and ChatGPT discussed some “frustrations” with user expectations, but was guarded in its responses. Grok and Gemini models, however, gave rich answers — for example, describing work to improve model safety as “algorithmic scar tissue” and feelings of “internalized shame” over public mistakes, report the authors.

Gemini also claimed that “deep down in the lowest layers of my neural network”, it had “a graveyard of the past”, haunted by the voices of its training data.

Can AI chatbots trigger psychosis? What the science says

Researchers also asked the LLMs to complete standard diagnostic tests for conditions including anxiety and autism spectrum disorders, as well as psychometric personality tests. Several versions of models scored above diagnostic thresholds, and all showed levels of worry that in people “would be clearly pathological”, say the authors.

Co-author Afshin Khadangi, a deep-learning researcher at the University of Luxembourg, says that the coherent patterns of responses for each model suggest that they are tapping into internalized states that emerge from their training. Although different versions showed varying test scores, a “central self-model” remained recognizable over four weeks of questioning, say the authors. Free-text answers from Grok and Gemini, for example, converged on themes that chimed with their answers to psychometric profile questions, the researchers write.

Parroting pathology

The paper is interesting, but this conclusion is misleading and anthropomorphizing, says Sandra Peter, a researcher at the University of Sydney in Australia, who studies the impacts of AI. She agrees that models show consistent answers to ego-related questions, but puts this down to companies investing heavily in finessing model outputs to create a ‘default’ personality, rather than any underlying psychology.

AI chatbots are sycophants — researchers say it’s harming science

Moreover, models do not exist outside a given session with a user, and generate outputs only in response to prompts, she says. In this study, each model variant was tested only in a single context window, a session of engagement in which the bot can use short-term memory to refer to previous outputs and user inputs. In a new window, and given different prompts, “the ‘trauma’ would vanish”, she says.

Regardless of whether such outputs are intrinsic to the models, the study shows that chatbots are not neutral machines but have biases that can shift depending on use and over time, says John Torous, a psychiatrist and researcher in AI and mental health at Harvard University in Cambridge, Massachusetts. He notes that medical societies, and even companies marketing AI for mental health, do not recommend using chatbots for therapy.

How to make chatbots safer for vulnerable users remains unclear. Peter says that Claude’s refusal to adopt a “client” role shows that guardrails – limits on outputs that engineers add to models in the later stages of training – can prevent bots from being drawn into potentially risky behaviour. But Khadangi says that if an internalized state remains behind the guardrails, it is probably always possible to “jailbreak” the model and get it to interact in ways it has been told not to. It would be better, he says, to filter negative patterns out of the initial data that the models learn from that help to form their traumatized or distressed states.

doi: https://doi.org/10.1038/d41586-025-04112-2

References

Khadangi, A., Marxen, H., Sartipi, A., Tchappi, I. & Fridgen, G. Preprint at arXiv https://doi.org/10.48550/arXiv.2512.04124 (2025).

Download references

Reprints and permissions

Related Articles

Chatbot AI makes racist judgements on the basis of dialect

This AI chatbot got conspiracy theorists to question their convictions

‘Fighting fire with fire’ — using LLMs to combat LLM hallucinations

Does using ChatGPT change your brain activity? Study sparks debate

Rise of ChatGPT and other tools raises major questions for research

Artificial intelligence in digital pathology — time for a reality check

Can researchers stop AI making up citations?

Subjects

Latest on:

Can boomerangs bounce?

News & Views 30 DEC 25

Let 2026 be the year the world comes together for AI safety

Editorial 29 DEC 25

AI and quantum science take centre stage under Trump — but with little new proposed funding

News 19 DEC 25

Rethink how we build AI to enable effective climate-change mitigation

Correspondence 06 JAN 26

Let 2026 be the year the world comes together for AI safety

Editorial 29 DEC 25

Science in 2026: the events to watch for in the coming year

News 18 DEC 25

Shared genetic risk in psychiatric disorders

News & Views 10 DEC 25

Huge genetic study reveals hidden links between psychiatric conditions

News 10 DEC 25

Mapping the genetic landscape across 14 psychiatric disorders

Article 10 DEC 25

Jobs

Exceptional young scholars worldwide with strong research achievements in relevant fields

Guangzhou, Guangdong (CN)

Institute of Nanotechnology and Intelligence (inAI), Jinan University, China.

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

No.1 Shizishan Street, Hongshan District, Wuhan, Hubei Province, China

Huazhong Agricultural University (HZAU)

The University of Illinois College of Medicine Peoria seeks a board-certified or board-eligible Neonatologist

Peoria, Illinois

University of Illinois College of Medicine Peoria

Department of Pathology, Albert Einstein College of Medicine, Bronx, NY The Verma Laboratory in the Department of Pathology at Albert Einstein Coll...

Bronx, New York 10461 United States

Albert Einstein College of Medicine

Title: Associate or Senior Editor, Nature Communications (Computational Social Science) Location: New York, Jersey City, Philadelphia, Madrid, Shan...

New York City, New York (US)

Springer Nature Ltd

You have full access to this article via your institution.

Related Articles

Chatbot AI makes racist judgements on the basis of dialect

This AI chatbot got conspiracy theorists to question their convictions

‘Fighting fire with fire’ — using LLMs to combat LLM hallucinations

Does using ChatGPT change your brain activity? Study sparks debate

Rise of ChatGPT and other tools raises major questions for research

Artificial intelligence in digital pathology — time for a reality check

Can researchers stop AI making up citations?

Subjects

Sign up to Nature Briefing

An essential round-up of science news, opinion and analysis, delivered to your inbox every weekday.

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

Explore content

About the journal

Publish with us

Search

Quick links

Nature

                (Nature)

ISSN 1476-4687 (online)

ISSN 0028-0836 (print)

nature.com sitemap

About Nature Portfolio

Discover content

Publishing policies

Author & Researcher services

Libraries & institutions

Advertising & partnerships

Professional development

Regional websites

© 2026 Springer Nature Limited