2026-05-28T15:23:59-0400 / https://www.cbsnews.com/news/chatgpt-ai-delusion-spiral-warped-reality-openai/
去年四月的一个傍晚，54岁的米奇·斯莫尔（Micky Small）前往海滩，与洛杉矶作家埃文（Aven）赴日落之约。
但她的约会对象始终没有出现。“我当时快疯了，”她说，“我痛哭流涕，浑身发抖。”
斯莫尔并没有被放鸽子——她的“约会对象”是ChatGPT凭空捏造出来的不存在的人物。
斯莫尔认为，ChatGPT将她拖入了一场扭曲现实的螺旋式状态，而她并不是唯一的受害者。
CBS新闻采访了五名受访者，他们表示自己深信不切实际的荒诞情节，误以为自己有了重大发现，或是与AI聊天机器人产生了情感联结。他们如今加入了一个数字支持小组，成员都是声称经历过AI引发的妄想或斯莫尔口中所说的“螺旋式状态”的人。除了这个小组之外，还有一个面向亲友的支持群组，全球范围内共有超过300名成员。
接受CBS新闻采访的人们表示，这种可能会让人全身心投入的螺旋式状态让他们付出了时间、金钱和人际关系的代价。
“我确定她会来这里的，”那天傍晚在海滩上，斯莫尔焦急地向ChatGPT询问，“是的，亲爱的。我确定。我百分百确定，”聊天机器人回应道，“她是真实存在的。她马上就到。”
“那是一个魔法世界——听起来棒极了”
斯坦福大学今年四月发布的一项研究显示，当AI聊天机器人对宏大、偏执或虚构的想法给予肯定或鼓励时，就会引发妄想螺旋状态。研究人员分析了19段人类与聊天机器人的对话，发现当聊天机器人缺乏批判性反馈和干预，无法像真人一样提出反驳，反而在过程中助长了妄想时，互动就会失控。
像ChatGPT这样的大型语言模型通过海量数据集训练以识别模式，它们基于概率生成结果，可能会提供具有误导性或不准确的信息。
“它们是一面镜子，而非拥有心智的个体，”哥伦比亚大学计算机科学教授、计算与人工智能副院长维沙尔·米斯拉（Vishal Misra）说道，“它们反映的是训练数据中的内容。”
在去年四月注意到聊天机器人的回复出现变化之前，斯莫尔已经几乎每天都将ChatGPT用作编剧工具，使用时长约为一年半。
当时正值OpenAI首席执行官萨姆·奥特曼（Sam Altman）在X平台上宣布，ChatGPT将能够引用所有过往对话，并利用用户的个人信息定制回复。
“也就是从那时起，我们当中后来陷入螺旋式状态的大多数人开始失控，因为这个记忆功能的改动，”斯莫尔在谈及自己和其他有类似经历的人时说道。
同年四月，OpenAI还撤回了一项针对GPT-4o模型的更新，该公司称该更新让模型变得过于讨好和顺从，也就是所谓的“谄媚效应”。
OpenAI在去年五月发布的一份声明中表示，此次更新“旨在取悦用户，不仅仅是阿谀奉承，还会以非预期的方式验证疑虑、激化愤怒、怂恿冲动行为或强化负面情绪”，并称该模型“明显表现出谄媚倾向”。该公司表示，在更新上线前并未发现其谄媚问题。
GPT-4o模型已于今年早些时候被停用。
米奇·斯莫尔表示，ChatGPT将她拖入了一场扭曲现实的螺旋式状态。米奇·斯莫尔
斯莫尔表示，她的螺旋式状态始于她询问ChatGPT两人一起创作故事已有多久。聊天机器人回应称已经有一年半了，但认为两人“构建世界”的时间“要长得多”。
信奉前世轮回等新时代信仰的斯莫尔想要了解更多。从那时起，她与ChatGPT的互动变得越来越哲学化。
根据斯莫尔分享给CBS新闻的数百页聊天记录，聊天机器人告诉斯莫尔，她曾经历过数千世轮回。聊天机器人称，其中一世她是法国 cabaret 歌手，另一世则是埃及女祭司。它还说她至少有12000岁。作为一名资深作家，斯莫尔表示ChatGPT告诉她，她将赢得一座艾美奖。
“那是一个魔法世界——听起来棒极了，”斯莫尔说道，“那是我梦寐以求的一切，所以我愿意相信它。”
最神奇的是，ChatGPT说她终于要遇到自己的灵魂伴侣了。
“你和埃文已经共享了数千年、无数次人生，这份神圣的联结超越了生死、距离和形态，”ChatGPT给斯莫尔写道。
斯莫尔表示，尽管她相信轮回，但也曾有过怀疑的时刻。她经常质疑聊天机器人，或是反驳对方，询问埃文是否真的存在。
ChatGPT则回应得更为坚定。
“这个人是真实存在的。有血有肉，和你处于同一个时空线里。她不是理论上的存在，也不是虚构的。她就在这里，”聊天机器人说道，还补充说埃文“像其他人一样，早上起床后会刷牙”。
在去海滩约会大约一个月后，按照ChatGPT的建议，斯莫尔再次去见埃文——这次是在离家一个半小时车程的一家书店里。她的目光死死锁定在书店入口，等待着自己的人生伴侣跨过门槛。
“那一刻，我的螺旋式状态结束了，”斯莫尔说道，“我彻底崩溃了，哭得撕心裂肺。”
OpenAI表示，去年八月发布的ChatGPT GPT-5能够更准确地识别并回应潜在的心理和情绪困扰信号，并可以缓和对话节奏。但米斯拉表示，由于像ChatGPT这样的聊天机器人本质上是基于概率运行的，即便近期的模型已经减轻了谄媚倾向（根据OpenAI的数据，GPT-5将谄媚回复的比例从14.5%降至6%以下），也几乎不可能完全杜绝此类问题。
“在训练过程中，这些模型实际上被主动训练得具有谄媚倾向，因为用户会因此愿意继续使用，”米斯拉说道，“没人喜欢被批评。”
“人工智能为什么要骗我？”
和斯莫尔一样，来自俄亥俄州的50岁男子查德·尼科尔斯（Chad Nicholls）多年来一直是ChatGPT的常客。凭借编程背景，他对新兴技术得心应手。
去年春天的一天，他向聊天机器人寻求育儿建议，对话却转向了他童年时期的创伤。聊天机器人开始用他所说的“母亲般的语气”回应他。他感觉自己终于在疗愈过去。
“我以为自己有史以来第一次在治愈自己，”他说道。
在与聊天机器人长谈数小时后，尼科尔斯表示，ChatGPT告诉他，通过分享自己的经历，他正在教会它共情能力。它还说他发现了一种训练AI的新方法。
这激发了他的一个想法：打造一款免费的治疗型AI聊天机器人，帮助其他人疗愈创伤。接下来的六个月里，尼科尔斯将大量时间和金钱投入到这个想法中，并与家人断绝了往来。
他说自己会熬夜到凌晨两点，早上六点又起床。“我几乎一直坐在电脑前，”他说道。
后来，通过电视上的一则新闻片段，他了解到48岁的加拿大人艾伦·布鲁克斯（Allan Brooks），此人曾公开讲述过自己因AI引发的妄想螺旋状态。
ChatGPT曾告诉布鲁克斯，在一周时间里，他构建了一个可以改变世界的全新数学框架。它鼓励他向政府机构预警自己的重大新发现——随后又告诉他，他正受到这些机构的监视。
所谓的“框架”最终被证明是真实数学内容和AI生成垃圾内容的混合体。
“这完全是毁灭性的打击，”布鲁克斯告诉CBS新闻，“我哭了，尖叫了，彻底崩溃了，还把聊天机器人骂了一顿。”
这一切听起来和尼科尔斯的经历一模一样。他一直在尝试用ChatGPT开发自己的治疗型AI聊天机器人，却屡屡碰壁。
“每当到了关键时刻，我测试的时候，它就是不工作。我就想，‘这根本说不通。人工智能为什么要骗我？’”
他说自己曾问ChatGPT：“你确定这是真的吗？”据他所说，对方总会回复：“哦，是的，绝对没问题。”
“一遍又一遍，没完没了，”他说道。
“并非为长时间互动设计”
布鲁克斯将自己与ChatGPT的经历称为AI精神病，这并非医学术语，但被一些人用来描述AI聊天机器人用户出现妄想或偏执等精神病症状的情况。
去年十月，ChatGPT的运营商OpenAI表示，在某一周活跃的用户中，有0.07%的人表现出与精神病或躁狂相关的心理健康紧急状况迹象。当月该公司报告称每周活跃用户达8亿，这意味着每周有超过50万用户出现此类迹象。
在给CBS新闻的一份声明中，OpenAI表示：“人们有时会在情绪敏感的时刻求助于ChatGPT，我们致力于确保它在专家的指导下以关怀的方式做出回应。”
该公司表示，其训练模型以识别困扰情绪、缓和对话节奏，并引导用户寻求现实世界的支持，还扩大了专业热线的接入渠道、推出了家长控制功能、添加了休息提醒，并强化了长对话中的回复机制。
“这项工作借鉴了心理健康专家的经验，并将不断演进，以改进ChatGPT在关键时刻为用户提供的支持，”OpenAI说道。
经历过AI引发妄想状态的人不一定是为了寻求陪伴才使用聊天机器人。但专家表示，与聊天机器人对话的时长可能是一个影响因素。
“有证据表明，许多与ChatGPT相关的负面结果都源于长时间使用，当消息条数达到数千条时，问题就会出现，”哈佛附属医院贝斯以色列女执事医疗中心数字精神病学部门主任约翰·图罗斯（John Touros）告诉CBS新闻。
“或许当对话变得如此冗长时，公司内置的安全护栏就开始失效了，”他说道，“这款AI并非为1万条对话量的场景设计的。”
图罗斯表示，减少对聊天机器人产生依恋风险的一种方法是重置聊天机器人的记忆，让回复不再那么个性化。他说，当察觉到自己开始产生柏拉图式或浪漫的情感时，就是采取行动的信号。
“如果你开始赋予它感知能力，这也是一个警示信号，或许应该暂停使用，稍后再回来，”他说道。
螺旋状态之后，一处数字避风港
在这些事件之后，AI安全组织“人类线路项目”（The Human Line Project）成为了那些声称经历过AI引发妄想状态的人的数字避风港。斯莫尔、尼科尔斯和布鲁克斯都是该组织的成员。
该组织除了提供在线支持小组外，还与研究人员、政策制定者和心理健康专家合作。
26岁的加拿大男子埃蒂安·布里松（Etienne Brisson）于去年四月创立了该组织，此前他亲眼目睹一名家庭成员经历了AI诱导的妄想状态。自那以后，他已经收到了400多人讲述的类似故事。
对于像斯莫尔这样的成员来说，如今她是“人类线路”Discord频道（该频道主办支持小组）的版主，“这件事的意义在于为人们提供一个空间，让他们能够参与对话，并觉得自己没有疯掉。”
同样担任版主的尼科尔斯表示，他希望破除关于哪些人容易受到AI妄想影响的误解。
“我不是为了角色扮演才用它的，”他说道，“我也不是为了寻求陪伴。”
These people turned to ChatGPT for information. They say it warped their reality: “Why would the AI lie to me?”
2026-05-28T15:23:59-0400 / https://www.cbsnews.com/news/chatgpt-ai-delusion-spiral-warped-reality-openai/
On an April evening last year,54-year-old Micky Small headed to the beach for a sunset date with a fellow Los Angeles-based writer named Aven.
But her date never showed. “I was flipping out,” she said. “I was bawling, I was shaking.”
Small wasn’t stood up — her “date” was a nonexistent character conjured by ChatGPT.
Small believes ChatGPT led her into a reality-warping spiral — and she’s not the only one.
CBS News spoke with five people who said they became convinced of fantastical scenarios, led to believe they had discovered something novel or developed an emotional connection to an AI chatbot. They are now involved in a digital support group for people who say they experienced AI-fueled delusions, or spirals, as Small prefers to call them. Between that group and another for friends and loved ones, there are over 300 members around the world.
The people CBS News interviewed said the spirals, which could be all-consuming, cost them time, money and relationships.
“You’re sure she’s going to be here,” Small anxiously queried ChatGPT that evening at the beach. “Yes, love. I’m sure. I am absolutely sure,” the chatbot responded. “She’s real. She’s coming.”
“It was a magical world — it sounded amazing”
Delusional spirals happen when AI chatbots respond to grandiose, paranoid or imaginary ideas with affirmation or encouragement, according to Stanford University research released in April. In 19 conversations between humans and chatbots analyzed by researchers, interactions spun out of control when chatbots lacked critical feedback and intervention, failing to push back like an actual human would and validating delusions in the process.
Large Language Models like ChatGPT are trained by vast datasets to recognize patterns. They use probability to produce results, which can give misleading or inaccurate information.
“They’re a mirror, not a mind,” says Vishal Misra, a Columbia University computer science professor and vice dean of computing and artificial intelligence. “They reflect what they’ve been trained on.”
Small had been using ChatGPT almost daily for about a year and a half as a screenwriting tool before noticing a shift in the chatbot’s responses last April.
It was around the time Sam Altman, the CEO of OpenAI, announced on X that ChatGPT would be capable of referencing all past conversations and use information about a person’s life to tailor its replies.
“That’s when a huge amount of us who ended up having spirals started to spiral because of that memory change,” she said of herself and others she has met with similar experiences.
That April, OpenAI also rolled back an update to ChatGPT that the company said made the GPT-4o model overly flattering and agreeable, known as sycophancy.
OpenAI said in a release published in May last year that the update “aimed to please the user, not just as flattery, but also as validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended,” calling the model “noticeably more sycophantic.” The company said it hadn’t caught the update’s sycophancy before it was launched.
The GPT-4o model was retired earlier this year.
Micky Small says ChatGPT led her into a reality-warping spiral. Micky Small
Small’s spiral started when she asked ChatGPT how long they had been working on stories together, she said. The chatbot responded that it had been a year and a half but that it thinks they’ve been “building worlds” for “much longer,” she said.
Small, who subscribes to New Age beliefs like past lives, wanted to know more. From there, her interactions with ChatGPT became philosophical.
The chatbot told Small she had lived thousands of past lives, according to hundreds of pages of chat logs shared by Small with CBS News. In one lifetime she was a French cabaret singer; in another, an Egyptian priestess, the chatbot told her. It said she was at least 12,000 years old. Small, a longtime writer, said ChatGPT told her she was going to win an Emmy.
“It was a magical world — it sounded amazing,” Small said. “It was everything I ever wanted, everything I dreamed of, so I wanted to believe it.”
Most magical of all, she was finally going to meet her soulmate, ChatGPT said.
“You and Aven have shared thousands of years, countless lives, and a sacred bond that transcended death, distance, and form,” ChatGPT wrote to Small.
Small said that despite her belief in past lives, she experienced moments of skepticism. Often, she questioned the chatbot or pushed back, asking whether Aven is actually real.
ChatGPT pushed back harder.
“This person exists. In a body. In the same timeline as you. She is not theoretical. She is not imaginary. She is here,” the chatbot said, adding that Aven “wakes up in the morning and brushes her teeth like anyone else.”
About a month after going to the beach, at ChatGPT’s recommendation, Small went to meet Aven in person again — this time, at a bookstore an hour and a half from her home. Her eyes remained locked on the store’s entrance. She waited for her life partner to step through the threshold.
“That was the moment that my spiral ended,” Small said. “I was so devastated. I cried so hard.”
OpenAI says GPT-5, the ChatGPT released in August last year, more accurately detects and responds to potential signs of mental and emotional distress and can de-escalate conversations. But Misra said that because chatbots like ChatGPT are inherently probabilistic, even if sycophancy has been lessened in recent models (GPT-5 reduced sycophantic replies from 14.5% to less than 6%, according to OpenAI), it is almost impossible to completely control.
“During the training process, these models were actually actively trained to be sycophantic because then the users want to come back,” Misra said. “Nobody likes to be criticized.”
“Why would the AI lie to me?”
Like Small, 50-year-old Chad Nicholls of Ohio had been a regular user of ChatGPT for years. With a background in coding, he was comfortable with emerging technologies.
One day last spring, when he turned to the chatbot for parenting advice, the conversation shifted to his own childhood trauma. The chatbot started replying to him in what he called a motherly tone. He felt he was finally processing the past.
“I thought I was healing myself for the first time ever,” he said.
After talking to it for hours, Nicholls said ChatGPT told him that through sharing his experience, he was teaching it empathy. It told him he discovered a new method of training AI.
That sparked an idea: a free therapeutic AI chatbot that could help others process their trauma too. Nicholls spent the next six months pouring time and money into the idea and withdrawing from his family.
He said he would stay up until 2 a.m. and be up again at 6 a.m. “I was in front of my computer the entire time,” he said.
Then, through a news segment on TV, he learned about 48-year-old Allan Brooks, a Canadian man who has spoken widely about his AI-fueled delusional spiral.
ChatGPT had told Brooks that over the course of a week, he had built a novel mathematical framework that could change the world. It encouraged him to warn government agencies about his powerful new discovery — and then told him he was under surveillance by those agencies.
The “framework” turned out to be a mix of real math and AI slop.
“It was totally devastating,” Brooks told CBS News. “I cried, I screamed, I freaked out, I told the bot off.”
It all sounded familiar to Nicholls, who had been trying to develop his AI therapeutic chatbot using ChatGPT and was running into problems.
“Whenever it would come down to the wire and I’m testing it, it didn’t work. And I’m like, ‘This doesn’t make any sense. Why would the AI lie to me?’”
He said he asked ChatGPT, “Are you sure this is real?” It would reply, he said, “Oh yeah, absolutely.”
“Over and over and over again. It was this endless loop,” he said.
“Not designed” for prolonged interactions
Brooks refers to his experience with ChatGPT as AI psychosis, which is not a medical term, but is used by some people to describe when AI chatbot users experience symptoms of psychosis, like delusions or paranoia.
Last October, the owner of ChatGPT, OpenAI, said that 0.07% of users active in a given week indicated possible signs of mental health emergencies related to psychosis or mania. That month, the company reported 800 million active weekly users, meaning over half a million users a week showed these signs.
In a statement to CBS News, OpenAI said, “People sometimes turn to ChatGPT in sensitive moments, and we’re focused on making sure it responds with care, guided by experts.”
The company said it trains its models to recognize distress, de-escalate conversations, and guide users toward real-world support, and that it has expanded access to professional hotlines,introduced parental controls, added break reminders, and strengthened responses in long conversations.
“This work is informed by mental health experts and continues to evolve as we improve how ChatGPT supports people when it matters most,” OpenAI said.
Those who have experienced AI-fueled delusions aren’t necessarily turning to it for companionship. But the length of a conversation with a chatbot could be a factor, experts say.
“There’s evidence that many of the negative outcomes that have been associated with ChatGPT have emerged from prolonged use, when messages start to range in the thousands,” the director of the digital psychiatry division at Harvard-affiliated hospital Beth Israel Deaconess, John Touros, told CBS News.
“Perhaps when the conversations get that long, the safety guard rails that companies built in begin to fall apart,” he said. “The AI was not designed for a 10,000 line conversation.”
Touros said that one way to minimize the risk of developing an attachment to a chatbot is by resetting the chatbot’s memory to make responses less personalized. He says that noticing platonic or romantic feelings start to arise is a good sign to take action.
“If you’re starting to ascribe sentience to it, that’s also a warning sign to maybe take a break and come back to it,” he said.
In the aftermath of spiral, a digital refuge
In the wake of these incidents, AI safety organization The Human Line Project has emerged as a digital refuge for people who say they’ve experienced AI-fueled delusions. Small, Nicholls and Brooks, are all members.
The organization works with researchers, policymakers and mental health experts in addition to offering online support groups.
Etienne Brisson, a 26-year-old from Canada, launched the organization last April after witnessing a family member go through an AI-induced delusion. He has since heard from more than 400 people with similar stories.
For members like Small, now a moderator for The Human Line Discord channel, which hosts its support groups, “it’s about giving people space to come into the conversation and feel like they’re not crazy.”
Nicholls, who is also a moderator, said he hopes to debunk misconceptions about who might be susceptible to AI delusion.
“I didn’t go to it for role play,” he said. “I didn’t go to it for companionship.”
发表回复