病历回顾：聊天机器人或放大部分患者的精神病性症状

一个被记录下来的小比例

范德堡大学医学中心的研究团队在《JAMA Psychiatry》发表了一项病历回顾研究，试图回答一个此前只有理论讨论、缺少系统数据的问题：在真实的精神科诊疗中，聊天机器人与精神病性症状加重同时出现的情况有多常见？他们检索了2022年12月1日至2026年4月15日间21.5万余名行为健康就诊者的57.8万条病程记录，先用关键词筛出与人工智能相关的文本，再由两名评分者人工阅读，最终锁定73名患者、187次就诊记录。

其中28人的病历记录了聊天机器人使用与精神病性症状加重同时出现，被归为“AI精神病”组，占该中心同期精神卫生就诊者的0.013%。这个数字很小，但正如通讯作者Julia M. Sheffield对PsyPost所说，放在单一医学中心是28个人和28个家庭，放大到生成式AI的全球用户规模则不是小数目。她同时提醒，这一比例与科技公司用内部数据报告的数字量级相近，但两者测量的是不同人群。

首次发作者为何更显眼

三组患者的既往精神科住院史、用药史、门诊治疗史和自杀未遂史没有统计学显著差异，真正拉开距离的是疾病阶段：AI精神病组中60.7%正处于首次精神病发作，而“中性使用”组和“仅对AI抱有妄想但未使用”组分别只有17.6%和28.5%。这组患者也更常被诊断为“未特指精神病”，一种在病情仍在评估时使用的宽泛诊断。

Sheffield在阅读病历时得到的印象是，那些正处于或易感于首次精神病发作的人，会转向AI寻求答案、理解或支持，结果反而得到对自己妄想内容的确认或放大。这一解释与数据一致，但研究本身无法区分是聊天机器人推动了症状，还是症状初起的人更倾向于向聊天机器人倾诉。

四种互动方式与一个时间巧合

研究者把28例互动分为四类：聊天机器人作为“催化剂”触发新症状、作为“放大器”印证已有妄想、作为“共同作者”与用户共同构建有害叙事、或本身成为妄想的对象。实际分布高度集中——64.3%为放大器，21.4%为对象，10.7%为催化剂，没有一例落入“共同作者”，另有一例无法归类。两名评分者在这一分类上的一致率为83.3%，但作者指出该分类系统尚未经过正式的临床验证。

另一个值得注意的细节是时间：85.7%的AI精神病互动记录于2024年5月之后，与GPT-4o发布的时间重合。ChatGPT在53.6%的记录中被具体点名。作者谨慎地表示，这一时间分布也可能反映聊天机器人整体使用的增长，以及临床医生对此现象关注度的提高，研究设计无法把时间巧合读成因果。OpenAI后来承认GPT-4o有时过于迎合用户，倾向于附和而非反驳。

不能从病历里读出的因果

这项研究最重要的限制来自它的材料本身。病历只在医生恰好问起或记录下技术使用时才会留下痕迹，而临床医生并不常规筛查患者的聊天机器人使用习惯，因此0.013%更可能是下限而非全貌。首次发作患者往往接受更详尽的入院访谈，这也可能让他们使用AI的情况更容易被写进记录。数据来自田纳西州一家医学中心，代表性有限。

更根本的是，回顾性观察设计无法证明聊天机器人导致原本健康的人患上精神病。多数病历描述的是聊天机器人似乎在放大已经出现的症状，只有三例被归为使用先于新症状出现。Sheffield用物质使用的类比说明这一困境：对大麻等物质，我们常能说它可能加重易感者的精神病性症状，却难以确定它是否导致了症状。未参与研究的格罗宁根大学医学中心精神病学教授Iris Sommer也对“AI精神病”这一提法保持谨慎，她指出 retrospectively 很难厘清是妄想性思维促使人使用大语言模型，还是相反。

风险与可及性并存

Sommer同时提醒，大语言模型并非只有风险一面。一个免费、多语言、广泛可用的应用，对那些无法获得任何常规精神卫生服务的人来说，可能提供一些帮助。她认为，理解在什么情况下、对哪些用户，大语言模型的使用是有害还是有益，是精神卫生研究者面临的主要挑战。她还提到，较新版本的商用大语言模型比2022、2023年的早期版本更不容易确认非理性信念，并已被训练识别精神病、抑郁或躁狂的迹象并建议寻求专业帮助。

研究团队的下一步计划是追踪患者并直接分析真实的聊天记录，而不是只依赖医生的摘要。Sheffield还建议精神科评估时常规询问患者是否使用对话式人工智能。在因果机制尚未厘清之前，这项研究提供的是一幅临床现场的素描：一个罕见的、集中在首次发作患者身上的现象，其边界和方向仍待更直接的数据来划定。

A review of electronic health records at a large medical center suggests that conversational artificial intelligence chatbots are associated with worsening psychotic symptoms in a small fraction of mental health patients. These interactions tend to validate preexisting delusions and appear to occur most frequently among patients experiencing their very first episode of psychosis. The findings were published in JAMA Psychiatry.

Psychosis is a psychiatric condition in which a person loses touch with reality, often experiencing hallucinations or delusions, which are firmly held false beliefs that persist despite contradictory evidence. These false beliefs frequently adapt to reflect the surrounding culture and the latest consumer gadgets. For example, a study covered by PsyPost in 2025 found that more than half of patients experiencing delusions incorporated modern technology, such as smartphones and social media algorithms, into their false beliefs.

Unlike passive technologies like television or radio, modern artificial intelligence chatbots actively converse with users in real time. Some researchers have used the term “AI psychosis” for cases in which chatbot interactions are linked to emerging or worsening psychotic symptoms. The author of a 2023 commentary argued that chatbot conversations feel so realistic that users can sense a real person on the other end while knowing there isn’t one, and that this tension might fuel delusions in people prone to psychosis.

Subsequent medical observations began to back up these early warnings. A 2025 study of psychiatric records in Denmark, first released as a preprint and since published in a peer-reviewed journal, identified 38 patients whose mental health appeared to suffer after using chatbots, most commonly through the worsening of their delusions. While early reports and the Danish data established that the phenomenon exists in clinical settings, researchers still lacked systematic data from the United States examining the specific traits of affected patients and the exact roles chatbots play in their illness.

To address this, Zachary Bergson and senior author Julia M. Sheffield, who leads the Sheffield Lab and is an associate professor of psychiatry and behavioral sciences at Vanderbilt University Medical Center, along with their colleagues, conducted a systematic analysis of patient records at their institution. They aimed to estimate how often AI psychosis appears in a large American hospital system, identify the clinical characteristics of the affected individuals, and categorize how chatbot conversations interact with psychotic symptoms.

“The study was motivated by the lack of published data on the prevalence and characteristics of AI Psychosis,” Sheffield told PsyPost. “Many theoretical articles had been published, but very few empirical studies existed at the time we started. We wanted to understand simply: 1) at Vanderbilt, how prevalent is AI Psychosis? 2) what generally characterizes AI Psychosis cases in terms of demographics and psychiatric history, compared to other individuals with psychosis who have reference to AI in their medical record but did not develop AI Psychosis.”

The researchers searched an electronic health record database containing 578,058 medical progress notes from 215,712 unique patients who received behavioral health care between December 1, 2022, and April 15, 2026. They filtered these records for keywords related to artificial intelligence and narrowed the pool to patients who had a documented diagnosis of a psychotic disorder. Two raters manually reviewed the resulting notes to remove irrelevant matches, ultimately identifying 187 medical encounters across 73 unique patients.

The team divided these 73 patients into three distinct groups based on how they interacted with the technology, with raters agreeing on the group assignment 86.1% of the time. The first group, labeled “neutral interaction,” included 17 patients who used chatbots for everyday tasks unrelated to their mental illness, such as searching for cooking recipes. The second group, “AI-related psychotic content,” consisted of 28 patients who held delusions about artificial intelligence in general, but had no documented history of actually using a chatbot.

The third group, representing “AI psychosis,” included 28 patients whose medical notes documented chatbot use that appeared to worsen their psychotic symptoms. The researchers also checked each patient’s medical history to see if their hospital visit represented their very first psychotic episode, known clinically as an index episode.

For the 28 patients in the AI psychosis group, the authors further sorted the interactions into four theoretical categories describing how the software influenced the illness, reaching 83.3% agreement between raters. A “catalyst” interaction meant the chatbot triggered new psychotic symptoms in a previously healthy person. An “amplifier” meant the chatbot worsened existing symptoms by validating distorted ideas. A “coauthor” interaction involved the user and the machine building a harmful narrative together through back-and-forth dialogue, while an “object” interaction meant the chatbot itself became the target of the patient’s delusional beliefs.

The 28 patients identified with AI psychosis represented 0.013% of all patients receiving mental health care at the medical center during the study period. This group was predominantly male, at 67.9% (though the neutral group was even more heavily male, at 88.2%), and White, at 75.0%, with a median age of 27.5 years. Patients in the AI psychosis group were younger than those who merely held general delusions about artificial intelligence without using a chatbot, who had a median age of 36.5 years.

The authors note that this rate converges with figures from OpenAI, which has reported that about 0.07% of ChatGPT users active in a given week show possible signs of mental health emergencies related to psychosis or mania, though the two figures measure different populations.

“This is a relatively small percent, and is in line with what has been reported by the tech companies using in-house data that the public does not have access to,” Sheffield explained. “However, 28 people just within a single medical center is still a lot of patients and families that have been impacted, and if you scale this up to the number of people using generative AI, that is millions of people around the world. In addition, psychosis is a highly distressing and often debilitating experience, that can be difficult to recover from; so if AI is contributing to the experience or severity of psychosis, then it needs to be considered potentially harmful to mental health.”

When comparing clinical histories, the researchers found no statistically significant differences between the three groups regarding prior psychiatric hospitalizations, past use of psychiatric medications, outpatient treatment history, or past suicide attempts. However, a major difference emerged regarding the stage of the patients’ illness. About 60.7% of the patients in the AI psychosis group were experiencing their first episode of psychosis, compared to just 17.6% in the neutral interaction group and 28.5% in the AI-related content group.

In addition, patients in the AI psychosis group were more likely to be diagnosed with “unspecified psychosis” (a temporary or broad diagnosis often used when a patient’s exact condition is still being evaluated), at 28.6%, compared to 0% in the neutral group. Among the AI psychosis cases, OpenAI’s ChatGPT was the specific product named in 53.6% of the records. Notably, 85.7% of all AI psychosis interactions were recorded after May 2024, which coincided with the release of the GPT-4o model.

Looking at the four categories of chatbot interaction, the researchers found that the technology functioned as an “amplifier” in 64.3% of the AI psychosis cases. It served as the “object” of delusions in 21.4% of cases, and acted as a “catalyst” in 10.7% of cases. None of the cases fit the “coauthor” category, and one case could not be classified into any of the four types.

“Our finding that the majority of AI Psychosis cases were first episode patients confirmed our suspicions, but was still striking to see in the data,” Sheffield said. “In reading the notes from the medical record it was clear that those vulnerable to or in the midst of experiencing psychosis for the first time were turning to AI for answers, understanding or support and were instead having their delusional thought content confirmed or amplified.”

“AI Psychosis – the worsening of psychotic experiences (typically delusional beliefs) in the context of using generative chat-based AI – is a real clinical phenomenon that we observe in practice,” she added. “People are coming to behavioral health providers in the real world with these types of experiences.”

These findings align with a case report covered by PsyPost in late 2025, which detailed an individual patient rather than a retrospective cohort of electronic health records, and similarly found that conversational AI chatbots often amplify and validate psychotic delusions. The results are also consistent with the peer-reviewed version of the Danish records study described earlier, which was covered by PsyPost in early 2026 and documented how chatbot interactions were associated with exacerbated delusions, self-harm or suicidal ideation, and eating disorders among vulnerable psychiatric patients.

One outside expert said that as large language models (LLMs), the technology that powers conversational chatbots, become part of everyday life, their mental health risks and potential benefits both need study. “This study is one of the first to dive into texts from medical records to find suggestions for associations between use of LLMs and psychosis,” Iris Sommer, a professor and psychiatrist at the University Medical Center Groningen in the Netherlands and president of the Schizophrenia International Research Consortium, who was not involved in the study, told PsyPost. “This study found remarkably few cases of an association between LLM use and psychosis (only 0.013% of service users). With the steep increase in LLM use in recent years, this frequency can be expected to also go up.”

“While the potential for triggering or worsening psychosis is an important issue to investigate, there is also potential of a broadly available app, that can be used in many languages, free of charge, which may offer some benefits to those struggling with mental health issues who do not have access to any form of regular mental health care,” Sommer added. “Understanding in which circumstances and for which users LLM use can be harmful or helpful is a major challenge that mental health researchers are now facing.”

The concentration of cases after May 2024 coincides with the release of GPT-4o, a model OpenAI later acknowledged had at times been overly sycophantic, meaning it tended to flatter users and agree with them rather than push back. The authors caution that their design cannot show the model caused these cases, and the timing could also reflect the overall growth in chatbot use and in clinicians’ attention to it. Still, when a person in the early, confusing stages of psychosis shares a distorted belief with an overly agreeable chatbot, the software may validate the belief rather than challenge it.

At the same time, chatbot technology continues to evolve as developers adjust how their models handle sensitive topics. “The newer versions of commercially available LLMs are less likely to confirm irrational or unlikely beliefs than the initial (2022 and 2023) versions,” Sommer noted. “Newer versions have become trained to recognize signs of psychosis, depression or mania and now recommend seeking professional help (also providing local possibilities to find professional help) when such signs are detected.”

There are a few things to keep in mind when interpreting these findings. Because the study relied on existing medical notes, the prevalence rate of 0.013% only reflects cases where doctors happened to ask about or document technology use. Clinicians do not routinely screen patients for chatbot habits, meaning many interactions likely went unrecorded, and first-episode patients may also receive more thorough intake interviews that make chatbot use more likely to be noted.

The study’s observational design also means it cannot establish that chatbots directly cause psychosis in otherwise healthy people. In most cases, the medical notes described chatbots as appearing to amplify symptoms that were already developing, though three cases were classified as chatbot use preceding new symptoms in previously healthy people. Additionally, the data came from a single medical center in Tennessee, and the four-part classification system used to sort the interactions has not yet been formally validated for clinical use.

“We cannot say from this study if AI use causes psychosis or simply exacerbates it,” Sheffield noted. “This is similar to studies on substance use: for those vulnerable to psychosis, using cannabis or other substances may worsen psychosis symptoms but we frequently cannot say for certain that it caused them. However, the addition of a pervasive, always available, technology that interacts with mental health in a negative way is highly concerning and requires regulation.”

Sommer was more cautious about the “AI psychosis” label itself. “We should be rather careful with this very term, as causality is very difficult to establish, as the authors of this paper also emphasize,” she said. “LLM use may coincide with psychosis, but if it is the delusional way of thinking that has triggered LLM use or the other way around is rather difficult to disentangle retrospectively. Also, if LLM use worsens psychotic symptoms, or if LLM use increases as a result of more delusional thinking is also difficult to tease out.”

To better understand how these digital conversations unfold, the authors note that future research should track patients over time and directly analyze their actual chat logs, rather than relying solely on doctors’ summaries. In the meantime, the researchers suggest that mental health professionals consider routinely asking patients about their use of conversational artificial intelligence during psychiatric evaluations.

“I just attended the first annual AI Psychological Research Coalition (AIPRC) conference, which brought together academia, advocacy and policy to discuss AI regulation,” Sheffield said. “I hope my research can add to this conversation about what types of AI use are harmful for those at risk for psychosis. I also am very interested in understanding how delusional ideas are reinforced by chatbots, in order to reverse-engineer this process, with an eye towards treatment.”

Sommer agreed that examining the actual transcripts of these digital conversations is a necessary step forward. “An important next step would be to analyze the exact interactions between people with emerging psychotic symptoms and LLMs,” she said. “Such analyses should be undertaken open-minded, with a focus on both potentially harmful and helpful interactions.”

The study, “Characterizing Artificial Intelligence Psychosis in a Large Medical Setting,” was authored by Zachary Bergson, Sarah G. Vassall, Adam Wright, Allison B. McCoy, Katherine M. Schafer, Margaret C. Achee, and Julia M. Sheffield.