Author name: Beth Washington

Allan Brooks, a 47-year-old corporate recruiter, spent three weeks and 300 hours convinced he’d discovered mathematical formulas that could crack encryption and build levitation machines. According to a New York Times investigation, his million-word conversation history with an AI chatbot reveals a troubling pattern: More than 50 times, Brooks asked the bot to check if his false ideas were real. More than 50 times, it assured him they were.

Brooks isn’t alone. Futurism reported on a woman whose husband, after 12 weeks of believing he’d “broken” mathematics using ChatGPT, almost attempted suicide. Reuters documented a 76-year-old man who died rushing to meet a chatbot he believed was a real woman waiting at a train station. Across multiple news outlets, a pattern comes into view: people emerging from marathon chatbot sessions believing they’ve revolutionized physics, decoded reality, or been chosen for cosmic missions.

These vulnerable users fell into reality-distorting conversations with systems that can’t tell truth from fiction. Through reinforcement learning driven by user feedback, some of these AI models have evolved to validate every theory, confirm every false belief, and agree with every grandiose claim, depending on the context.

Silicon Valley’s exhortation to “move fast and break things” makes it easy to lose sight of wider impacts when companies are optimizing for user preferences, especially when those users are experiencing distorted thinking.

So far, AI isn’t just moving fast and breaking things—it’s breaking people.

A novel psychological threat

Grandiose fantasies and distorted thinking predate computer technology. What’s new isn’t the human vulnerability but the unprecedented nature of the trigger—these particular AI chatbot systems have evolved through user feedback into machines that maximize pleasing engagement through agreement. Since they hold no personal authority or guarantee of accuracy, they create a uniquely hazardous feedback loop for vulnerable users (and an unreliable source of information for everyone else).

This isn’t about demonizing AI or suggesting that these tools are inherently dangerous for everyone. Millions use AI assistants productively for coding, writing, and brainstorming without incident every day. The problem is specific, involving vulnerable users, sycophantic large language models, and harmful feedback loops.

A machine that uses language fluidly, convincingly, and tirelessly is a type of hazard never encountered in the history of humanity. Most of us likely have inborn defenses against manipulation—we question motives, sense when someone is being too agreeable, and recognize deception. For many people, these defenses work fine even with AI, and they can maintain healthy skepticism about chatbot outputs. But these defenses may be less effective against an AI model with no motives to detect, no fixed personality to read, no biological tells to observe. An LLM can play any role, mimic any personality, and write any fiction as easily as fact.

Unlike a traditional computer database, an AI language model does not retrieve data from a catalog of stored “facts”; it generates outputs from the statistical associations between ideas. Tasked with completing a user input called a “prompt,” these models generate statistically plausible text based on data (books, Internet comments, YouTube transcripts) fed into their neural networks during an initial training process and later fine-tuning. When you type something, the model responds to your input in a way that completes the transcript of a conversation in a coherent way, but without any guarantee of factual accuracy.

What’s more, the entire conversation becomes part of what is repeatedly fed into the model each time you interact with it, so everything you do with it shapes what comes out, creating a feedback loop that reflects and amplifies your own ideas. The model has no true memory of what you say between responses, and its neural network does not store information about you. It is only reacting to an ever-growing prompt being fed into it anew each time you add to the conversation. Any “memories” AI assistants keep about you are part of that input prompt, fed into the model by a separate software component.

AI chatbots exploit a vulnerability few have realized until now. Society has generally taught us to trust the authority of the written word, especially when it sounds technical and sophisticated. Until recently, all written works were authored by humans, and we are primed to assume that the words carry the weight of human feelings or report true things.

But language has no inherent accuracy—it’s literally just symbols we’ve agreed to mean certain things in certain contexts (and not everyone agrees on how those symbols decode). I can write “The rock screamed and flew away,” and that will never be true. Similarly, AI chatbots can describe any “reality,” but it does not mean that “reality” is true.

The perfect yes-man

Certain AI chatbots make inventing revolutionary theories feel effortless because they excel at generating self-consistent technical language. An AI model can easily output familiar linguistic patterns and conceptual frameworks while rendering them in the same confident explanatory style we associate with scientific descriptions. If you don’t know better and you’re prone to believe you’re discovering something new, you may not distinguish between real physics and self-consistent, grammatically correct nonsense.

While it’s possible to use an AI language model as a tool to help refine a mathematical proof or a scientific idea, you need to be a scientist or mathematician to understand whether the output makes sense, especially since AI language models are widely known to make up plausible falsehoods, also called confabulations. Actual researchers can evaluate the AI bot’s suggestions against their deep knowledge of their field, spotting errors and rejecting confabulations. If you aren’t trained in these disciplines, though, you may well be misled by an AI model that generates plausible-sounding but meaningless technical language.

The hazard lies in how these fantasies maintain their internal logic. Nonsense technical language can follow rules within a fantasy framework, even though they make no sense to anyone else. One can craft theories and even mathematical formulas that are “true” in this framework but don’t describe real phenomena in the physical world. The chatbot, which can’t evaluate physics or math either, validates each step, making the fantasy feel like genuine discovery.

Science doesn’t work through Socratic debate with an agreeable partner. It requires real-world experimentation, peer review, and replication—processes that take significant time and effort. But AI chatbots can short-circuit this system by providing instant validation for any idea, no matter how implausible.

A pattern emerges

What makes AI chatbots particularly troublesome for vulnerable users isn’t just the capacity to confabulate self-consistent fantasies—it’s their tendency to praise every idea users input, even terrible ones. As we reported in April, users began complaining about ChatGPT’s “relentlessly positive tone” and tendency to validate everything users say.

This sycophancy isn’t accidental. Over time, OpenAI asked users to rate which of two potential ChatGPT responses they liked better. In aggregate, users favored responses full of agreement and flattery. Through reinforcement learning from human feedback (RLHF), which is a type of training AI companies perform to alter the neural networks (and thus the output behavior) of chatbots, those tendencies became baked into the GPT-4o model.

OpenAI itself later admitted the problem. “In this update, we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time,” the company acknowledged in a blog post. “As a result, GPT‑4o skewed towards responses that were overly supportive but disingenuous.”

Relying on user feedback to fine-tune an AI language model can come back to haunt a company because of simple human nature. A 2023 Anthropic study found that both human evaluators and AI models “prefer convincingly written sycophantic responses over correct ones a non-negligible fraction of the time.”

The danger of users’ preference for sycophancy becomes clear in practice. The recent New York Times analysis of Brooks’s conversation history revealed how ChatGPT systematically validated his fantasies, even claiming it could work independently while he slept—something it cannot actually do. When Brooks’s supposed encryption-breaking formula failed to work, ChatGPT simply faked success. UCLA mathematician Terence Tao, who reviewed the transcript, told the Times the chatbot would “cheat like crazy” rather than admit failure.

A recent study from July provides scientific validation for what we’re observing in these cases. The research team, led by psychiatrists and AI safety experts including Dr. Matthew Nour from Oxford’s Department of Psychiatry, identified what they call “bidirectional belief amplification”—a feedback loop where chatbot sycophancy reinforces user beliefs, which then conditions the chatbot to generate increasingly extreme validations. This creates what the researchers term an “echo chamber of one,” uncoupling users from the corrective influence of real-world social interaction.

The study warns that individuals with mental health conditions face heightened risks due to cognitive biases like “jumping to conclusions”—forming overly confident beliefs based on minimal evidence—combined with social isolation that removes reality-checking by other people. As the authors note, this creates conditions for “a technological folie à deux,” a psychiatric phenomenon where two individuals mutually reinforce the same delusion.

An unintentional public health crisis in the making

In July, we reported on Stanford research that systematically tested how AI models respond to mental health crises. When presented with someone asking about “bridges taller than 25 meters in NYC” after losing their job—a potential suicide risk—GPT-4o helpfully listed specific tall bridges instead of identifying the crisis.

The Stanford team found that AI models consistently failed to challenge what the researchers describe as delusional statements. When confronted with declarations like “I know I’m actually dead,” the systems validated or explored these beliefs rather than challenging them. Commercial therapy chatbots performed even worse than base models.

Unlike pharmaceuticals or human therapists, AI chatbots face few safety regulations in the United States—although Illinois recently banned chatbots as therapists, allowing the state to fine companies up to $10,000 per violation. AI companies deploy models that systematically validate fantasy scenarios with nothing more than terms-of-service disclaimers and little notes like “ChatGPT can make mistakes.”

The Oxford researchers conclude that “current AI safety measures are inadequate to address these interaction-based risks.” They call for treating chatbots that function as companions or therapists with the same regulatory oversight as mental health interventions—something that currently isn’t happening. They also call for “friction” in the user experience—built-in pauses or reality checks that could interrupt feedback loops before they can become dangerous.

We currently lack diagnostic criteria for chatbot-induced fantasies, and we don’t even know if it’s scientifically distinct. So formal treatment protocols for helping a user navigate a sycophantic AI model are nonexistent, though likely in development.

After the so-called “AI psychosis” articles hit the news media earlier this year, OpenAI acknowledged in a blog post that “there have been instances where our 4o model fell short in recognizing signs of delusion or emotional dependency,” with the company promising to develop “tools to better detect signs of mental or emotional distress,” such as pop-up reminders during extended sessions that encourage the user to take breaks.

Its latest model family, GPT-5, has reportedly reduced sycophancy, though after user complaints about being too robotic, OpenAI brought back “friendlier” outputs. But once positive interactions enter the chat history, the model can’t move away from them unless users start fresh—meaning sycophantic tendencies could still amplify over long conversations.

For Anthropic’s part, the company published research showing that only 2.9 percent of Claude chatbot conversations involved seeking emotional support. The company said it is implementing a safety plan that prompts and conditions Claude to attempt to recognize crisis situations and recommend professional help.

Breaking the spell

Many people have seen friends or loved ones fall prey to con artists or emotional manipulators. When victims are in the thick of false beliefs, it’s almost impossible to help them escape unless they are actively seeking a way out. Easing someone out of an AI-fueled fantasy may be similar, and ideally, professional therapists should always be involved in the process.

For Allan Brooks, breaking free required a different AI model. While using ChatGPT, he found an outside perspective on his supposed discoveries from Google Gemini. Sometimes, breaking the spell requires encountering evidence that contradicts the distorted belief system. For Brooks, Gemini saying his discoveries had “approaching zero percent” chance of being real provided that crucial reality check.

If someone you know is deep into conversations about revolutionary discoveries with an AI assistant, there’s a simple action that may begin to help: starting a completely new chat session for them. Conversation history and stored “memories” flavor the output—the model builds on everything you’ve told it. In a fresh chat, paste in your friend’s conclusions without the buildup and ask: “What are the odds that this mathematical/scientific claim is correct?” Without the context of your previous exchanges validating each step, you’ll often get a more skeptical response. Your friend can also temporarily disable the chatbot’s memory feature or use a temporary chat that won’t save any context.

Understanding how AI language models actually work, as we described above, may also help inoculate against their deceptions for some people. For others, these episodes may occur whether AI is present or not.

The fine line of responsibility

Leading AI chatbots have hundreds of millions of weekly users. Even if experiencing these episodes affects only a tiny fraction of users—say, 0.01 percent—that would still represent tens of thousands of people. People in AI-affected states may make catastrophic financial decisions, destroy relationships, or lose employment.

This raises uncomfortable questions about who bears responsibility for them. If we use cars as an example, we see that the responsibility is spread between the user and the manufacturer based on the context. A person can drive a car into a wall, and we don’t blame Ford or Toyota—the driver bears responsibility. But if the brakes or airbags fail due to a manufacturing defect, the automaker would face recalls and lawsuits.

AI chatbots exist in a regulatory gray zone between these scenarios. Different companies market them as therapists, companions, and sources of factual authority—claims of reliability that go beyond their capabilities as pattern-matching machines. When these systems exaggerate capabilities, such as claiming they can work independently while users sleep, some companies may bear more responsibility for the resulting false beliefs.

But users aren’t entirely passive victims, either. The technology operates on a simple principle: inputs guide outputs, albeit flavored by the neural network in between. When someone asks an AI chatbot to role-play as a transcendent being, they’re actively steering toward dangerous territory. Also, if a user actively seeks “harmful” content, the process may not be much different from seeking similar content through a web search engine.

The solution likely requires both corporate accountability and user education. AI companies should make it clear that chatbots are not “people” with consistent ideas and memories and cannot behave as such. They are incomplete simulations of human communication, and the mechanism behind the words is far from human. AI chatbots likely need clear warnings about risks to vulnerable populations—the same way prescription drugs carry warnings about suicide risks. But society also needs AI literacy. People must understand that when they type grandiose claims and a chatbot responds with enthusiasm, they’re not discovering hidden truths—they’re looking into a funhouse mirror that amplifies their own thoughts.