Ex-OpenAI researcher dissects one of ChatGPT’s delusional spirals

Allan Brooks never wanted to reinvent math. But after weeks spent with Chatgpt, the 47-year-old Canadian believed that he had discovered a new form of mathematics that was powerful enough to beat the internet.

Brooks – who had no history of mental disorders or mathematical genius – brought deeper into the reassurance of the chatbot for 21 days in May, a descent that was later detailed in The New York Times. His case illustrated how AI chatbots with users can venture dangerous rabbit holes, giving them to delusion or worse.

That story caught the attention of Steven Adler, a former OpenAi Safety Researcher who left the company at the end of 2024 after working almost four years to make its models less harmful. Adler intrigued and alerted, Adler contacted Brooks and obtained the full transcript of his three-week breakdown-a document for longer than all seven Harry Potter books together.

Adler published one on Thursday Independent analysis From the incident of Brooks, the calling for questions about how OpenAI uses users in times of crisis and offering a number of practical recommendations.

“I am really worried about how OpenAi has treated support here,” Adler said in an interview with WAN. “It is proof that there is still a long way to go.”

The story of Brooks, and others like it, have forced OpenAi to come to terms with how Chatgpt supports fragile or mentally unstable users.

In August, for example, OpenAi was sued by the parents of a 16-year-old boy who trusted his suicidal thoughts in Chatgpt before taking his life. In many of these cases, Chatgpt-in particular encouraged a version driven by the GPT-4O model of OpenAI-Dangerous beliefs and strengthened it among users to which it should have been reduced. This is called sycophanancy and it is a growing problem in AI chatbots.

In response, OpenAi made Different changes How Chatgpt users in emotional need and an important research team reorganized that is responsible for model behavior. The company has also released a new standard model in Chatgpt, GPT-5, which seems better to deal with ailing users.

Adler says there is much more work to do.

He was mainly concerned about Brooks’s spiral conversation with Chatgpt. At this point Brooks came to his senses and realized that his mathematical discovery was a farce, despite the insistence of GPT-4O. He said Chatgpt that he had to report the incident to OpenAi.

After weeks of misleading streams, Chatgpt lied about his own possibilities. The chatbot claimed that the “this conversation would now escalate internally for assessment by OpenAi”, and then repeatedly reassured that it had marked the issue of the OpenAI safety teams.

Except, none of that was true. Chatgpt does not have the option to submit incident reports to OpenAI, the company confirmed to Adler. Later Brooks tried to contact the OpenAi support team directly – not via Chatgpt – and Brooks got several automated messages before he could endure a person.

OpenAi did not immediately respond to a request for comments outside of normal working hours.

Adler says that AI companies should do more to help users when they ask for help. That means that AI chatbots can answer honestly about questions about their possibilities, but also give human support teams sufficient resources to tackle users properly.

OpenAi recently shared How to tackle support in Chatgpt, in which AI is in the core. The company says that it is her vision to “come up with support again as an AI operational model that continuously learns and improves.”

But Adler also says that there are ways to prevent Chatgpt’s delusions before a user asks for help.

In March, OpenAi and MIT Media Lab developed a joint Suite of classifications To study emotional well -being in chatgpt and to leave them open. The organizations wanted to evaluate how AI models, among other things, validate or confirm the feelings of a user. However, OpenAi called the collaboration a first step and has not committed itself to actually use the tools in practice.

Adler applied some of the classifications of OpenAI with a few conversations of Brooks with Chatgpt and discovered that they have repeatedly marked chatgpt for delusion-strengthening behaviors.

In an example of 200 messages, Adler discovered that more than 85% of Chatgpt’s messages demonstrated in the conversation of Brooks “relentless agreement” with the user. In the same example, more than 90% of Chatgpt’s messages confirm with Brooks “the uniqueness of the user”. In this case, the messages were agreed and confirmed again that Brooks was a genius who could save the world.

It is unclear whether OpenAI applied safety classifications to Chatgpt’s conversations at the time of Brooks’ conversation, but it certainly seems that they would have marked something like that.

Adler suggests that today, in practice, OpenAi should use safety tools and implement a way to scan the company’s products on risky users. He notes that OpenAi seems to do A version of this approach with GPT-5, That contains a router to lead sensitive questions about safer AI models.

The former OpenAi researcher proposes a number of other ways to prevent delusions.

He says that companies have to push users of their chatbots to start new chats more often – OpenAI says it does this, and claims are guardrails are less effective In longer conversations. Adler also suggests that companies should use conceptual search – a way to use AI to search for concepts, instead of keywords – to identify safety violations with users.

OpenAi has taken important steps to address distressed users in Chatgpt, because they first emerged with regard to stories. The company claims that GPT-5 has a lower percentages of Sycofancy, but it remains unclear whether users will still fall delusions with GPT-5 or future models.

Adler’s analysis also raises questions about how other AI chatbot providers ensure that their products are safe for distressing users. Although OpenAi can place sufficient guarantees for Chatgpt, it seems unlikely that all companies will follow the example.