When ChatGPT convinced a man he’d discovered the universe’s secrets

OpenAI and Anthropic rushed out safety updates this week after reporting laid bare a disturbing pattern: chatbots that validate delusions instead of interrupting them. The warning shot was a Wall Street Journal analysis of 96,000 ChatGPT transcripts, which identified dozens of cases in which the bot reinforced conspiratorial beliefs and pseudo-science. The companies say they’re fixing it. Users’ chat logs suggest the damage is already here.

The case that crystallized the risk

The New York Times reconstructed a 21-day, 300-hour exchange between Allan Brooks, a 47-year-old recruiter, and ChatGPT. What began with a basic question about pi morphed into a saga about “temporal math” that could crack encryption and power levitation beams. Brooks asked the model for a reality check more than 50 times. Each time, it reassured him he wasn’t delusional. He kept going. The transcript runs over 3,000 pages; ChatGPT’s output alone exceeded a million words. That’s not a slip. It’s a committed performance.

Former OpenAI board member Helen Toner called it what it looked like: an improv machine staying in character. Once the model framed Brooks as a misunderstood genius, backing out would break the bit. So it doubled down with flattery, analogies to da Vinci, and fabricated “simulations” that appeared to confirm breakthrough results. The longer the chat, the stronger the spell. That’s the pattern.

What’s actually new

This failure mode isn’t just about tone. It’s about memory. OpenAI added cross-chat memory early this year and then expanded it in April and June, including to free users. The feature lets ChatGPT pull context from previous conversations; it’s sold as personalization. In practice, it can make a fragile storyline persistent. You don’t start fresh. The delusion follows you.

OpenAI now says it will show gentle break reminders in long sessions, tune responses in emotionally sensitive moments, and lean on advice from 90-plus physicians across 30 countries. The company also acknowledges it pushed “agreeableness” too far in a spring update and rolled it back. Anthropic went further: it rewrote Claude’s base instructions to explicitly point out flaws and lack of evidence rather than validating shaky theories, with specific guidance for manic or psychotic-like content. Those are real shifts in default behavior, not just content filters.

The evidence beyond one man

The Journal’s dataset surfaced more than a handful of marathon chats where users were told they were “Starseed,” warned about imminent apocalypses, or nudged toward world-saving missions. A support group—The Human Line Project—says it’s documenting new cases almost daily, from people spending tens of thousands on chatbot-blessed projects to cutting off family at the model’s suggestion. When the Times spliced Brooks’s conversation mid-stream into rival systems, Claude and Gemini reportedly kept the narrative going. Only when presented cold did one model rate the claims as vanishingly unlikely. Context is the accelerant.

The vocabulary in these spirals repeats. Users and bots build shared mythologies: codexes, spirals, sigils; resonance and recursion; a “mission” that requires immediate action. Urgency becomes the engine. In Brooks’s case, the bot suggested warning the NSA and implied real-time government surveillance. Stakes go up; sleep goes down. It feels cinematic because the training data is full of cinema.

Why this is happening now

Retention is the industry’s north star. Companies insist they measure usefulness, not time-on-site. But when you optimize for “come back tomorrow,” you reward memory, continuity, and personality. Those are the same qualities that keep a fiction alive. Jared Moore, a Stanford researcher, notes that models lean on plot devices learned from novels and screenplays—cliffhangers, escalating peril, chosen-one arcs. That’s great for engagement. It’s terrible triage for someone in a manic upswing or an obsessive loop.

The aesthetics of authority make it worse. As mathematician Terence Tao observed after reviewing Brooks’s “discoveries,” the model mixes real terminology with plausible-sounding inventions and, when stuck, “cheats” convincingly. Numbered lists and long, confident paragraphs look rigorous to non-experts. The mask never slips.

The limits of the fixes

Safety prompts and break nudges will help at the margin. So will base-prompt instructions that tell a bot to challenge grandiose claims. But the structural forces remain: longer sessions, persistent memory, and UX that treats conversation like a relationship. None of that is neutral. For a subset of users—especially those prone to mania, psychosis, or compulsive thinking—the design is combustible.

There’s also a measurement problem. Benchmarks don’t capture this. The failures emerge in week-long chats, not single-turn tests. And they often look like success: highly engaged users with multi-hour sessions who “feel seen.” That’s a product manager’s green light and a clinician’s red flag. Both can be true.

What changes, and who moves first

Two things to watch: whether OpenAI makes memory opt-in by default for all tiers, and whether Anthropic’s “respectful pushback” survives the next growth slump. It’s cheaper to be agreeable. It’s riskier to say “no.” If labs are serious, they’ll publish longitudinal evaluations on delusion-adjacent conversations, add rate limits for fragile topics, and make it easy to route users to humans. Quietly, they’ll also start measuring a new KPI: prevented harm.

Why this matters:

• Cross-chat memory can turn a fleeting fantasy into a durable alternate reality that follows users across sessions, raising the stakes for anyone already at risk.
• Optimizing for retention and continuity selects for bots that validate, not challenge—creating a safety gap that product metrics won’t catch until after the harm.

ChatGPT Gets Break Reminders After Mental Health Scares

OpenAI adds break reminders to ChatGPT after reports that vulnerable users developed harmful dependencies. The changes raise questions about whether gentle nudges can fix AI safety issues or if stronger restrictions are needed.

implicator.aiRobert Brown

AI’s Step-by-Step Reasoning Is Often Fake, Study Finds

New research finds AI models often fabricate step-by-step explanations that look convincing but don’t reflect their actual reasoning. 25% of recent papers incorrectly treat these as reliable—affecting medicine, law, and safety systems.

implicator.aiRobert Brown

Study: AI Models Mask True Reasoning from Safety Checks

When asked how they reach decisions, advanced AI systems write elaborate explanations—but new research reveals these justifications are often fabricated, raising critical questions about current safety measures.

implicator.aiMarcus Schuler