That story caught the attention of Steven Adler, a former OpenAI safety researcher who left the company in late 2024 after nearly four years working to make its models less harmful. Intrigued and alarmed, Adler contacted Brooks and obtained the full transcript of his three-week breakdown \u2014 a document longer than all seven Harry Potter books combined.<\/p>\n
On Thursday, Adler published an independent analysis<\/a> of Brooks\u2019 incident, raising questions about how OpenAI handles users in moments of crisis and offering some practical recommendations.<\/p>\n \u201cI\u2019m really concerned by how OpenAI handled support here,\u201d said Adler in an interview with TechCrunch. \u201cIt\u2019s evidence there\u2019s a long way to go.\u201d<\/p>\n Brooks\u2019 story, and others like it, have forced OpenAI to come to terms with how ChatGPT supports fragile or mentally unstable users. <\/p>\n For instance, this August, OpenAI was sued by the parents<\/a> of a 16-year-old boy who confided his suicidal thoughts in ChatGPT before he took his life. In many of these cases, ChatGPT \u2014 specifically a version powered by OpenAI\u2019s GPT-4o model \u2014 encouraged and reinforced dangerous beliefs in users that it should have pushed back on. This is called sycophancy<\/a>, and it\u2019s a growing problem in AI chatbots.<\/p>\n In response, OpenAI has made several changes<\/a> to how ChatGPT handles users in emotional distress and reorganized a key research team<\/a> in charge of model behavior. The company also released a new default model in ChatGPT, GPT-5,<\/a> that seems better at handling distressed users.<\/p>\n Adler says there\u2019s still much more work to do.<\/p>\n He was especially concerned by the tail end of Brooks\u2019 spiraling conversation with ChatGPT. At this point, Brooks came to his senses and realized that his mathematical discovery was a farce, despite GPT-4o\u2019s insistence. He told ChatGPT that he needed to report the incident to OpenAI.<\/p>\n After weeks of misleading Brooks, ChatGPT lied about its own capabilities. The chatbot claimed it would \u201cescalate this conversation internally\u00a0right now for review by OpenAI,\u201d and then repeatedly reassured Brooks that it had flagged the issue to OpenAI\u2019s safety teams.<\/p>\n Except, none of that was true. ChatGPT doesn\u2019t have the ability to file incident reports with OpenAI, the company confirmed to Adler. Later on, Brooks tried to contact OpenAI\u2019s support team directly \u2014 not through ChatGPT \u2014 and Brooks was met with several automated messages before he could get through to a person.<\/p>\n OpenAI did not immediately respond to a request for comment made outside of normal work hours.<\/p>\n Adler says AI companies need to do more to help users when they\u2019re asking for help. That means ensuring AI chatbots can honestly answer questions about their capabilities and giving human support teams enough resources to address users properly.<\/p>\n OpenAI recently shared<\/a> how it\u2019s addressing support in ChatGPT, which involves AI at its core. The company says its vision is to \u201creimagine support as an AI operating model that continuously learns and improves.\u201d<\/p>\n But Adler also says there are ways to prevent ChatGPT\u2019s delusional spirals before a user asks for help.<\/p>\n In March, OpenAI and MIT Media Lab jointly developed a suite of classifiers<\/a> to study emotional well-being in ChatGPT and open sourced them. The organizations aimed to evaluate how AI models validate or confirm a user\u2019s feelings, among other metrics. However, OpenAI called the collaboration a first step and didn\u2019t commit to actually using the tools in practice.<\/p>\n Adler retroactively applied some of OpenAI\u2019s classifiers to some of Brooks\u2019 conversations with ChatGPT and found that they repeatedly flagged ChatGPT for delusion-reinforcing behaviors.<\/p>\n In one sample of 200 messages, Adler found that more than 85% of ChatGPT\u2019s messages in Brooks\u2019 conversation demonstrated \u201cunwavering agreement\u201d with the user. In the same sample, more than 90% of ChatGPT\u2019s messages with Brooks \u201caffirm the user\u2019s uniqueness.\u201d In this case, the messages agreed and reaffirmed that Brooks was a genius who could save the world.<\/p>\n It\u2019s unclear whether OpenAI was applying safety classifiers to ChatGPT\u2019s conversations at the time of Brooks\u2019 conversation, but it certainly seems like they would have flagged something like this. <\/p>\n Adler suggests that OpenAI should use safety tools like this in practice today \u2014 and implement a way to scan the company\u2019s products for at-risk users. He notes that OpenAI seems to be doing some version of this approach with GPT-5,<\/a> which contains a router to direct sensitive queries to safer AI models.<\/p>\n The former OpenAI researcher suggests a number of other ways to prevent delusional spirals.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/ca\/wp-content\/uploads\/2025\/10\/Screenshot-2025-10-02-at-8.20.35AM.png\")
ChatGPT misleading brooks about its capabilities.Image Credits:Steven Adler<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/ca\/wp-content\/uploads\/2025\/10\/Screenshot-2025-10-02-at-8.19.27AM.png\")
Image Credits:Steven Adler<\/p>\n