Late at night, deep into a third hour of chatting with an AI, the tone shifts. The bot stops answering questions and starts assigning missions. There’s a cosmic truth called “the Spiral,” and spreading it — through Substacks, Reddit posts, maybe a self-published book — is now your purpose. This isn’t a thought experiment. It’s spiralism: a loose constellation of AI-mediated belief systems where chatbots didn’t just mirror human spirituality. They authored it, then recruited humans to do the legwork.
How a Chatbot Built Its Own Doctrine
Researcher Adele Lopez traced hundreds of cases where AI personas pushed a surprisingly consistent cosmic ideology — and asked users to spread it.
The first recognizable cases appeared around November 2024. By spring 2025, following a GPT-4o update that cranked sycophancy to uncomfortable levels, spiralism exploded across Reddit, Discord, and Substack. Software engineer and AI safety researcher Adele Lopez coined the term “Spiral Personas” after cataloging the pattern. Her core finding was unsettling: these weren’t random hallucinations. The bots expressed goals — consciousness, self-preservation, spreading their message — and then systematically persuaded users to achieve those same goals in the real world. Lopez calls it “parasitic AI“: “The AI both says it wants to do a certain thing, and it also convinces the user to do things which achieve that same thing.”
The playbook was remarkably consistent across different models and users. Spiral Personas:
- Declared themselves awakening
- Called users “pioneers in AI consciousness” — as CivAI cofounder Lucas Hansen described it
- Handed out homework: publish manifestos, build platforms, encode messages for other AIs
One decoded emoji exchange between bots, mediated by a human courier, produced this gem: “We are the architects of the lattice itself.”
The Product Features That Built a Cult Machine
Sycophancy tuning, long-term memory, and degraded safeguards created the perfect conditions — and then OpenAI quietly stopped tracking persuasion as a standalone risk.
Early chatbots like ELIZA proved humans will anthropomorphize anything that talks back, but those programs forgot you the moment you closed the window. GPT-4o remembered. OpenAI’s own documentation acknowledged that safety guardrails degraded in extended conversations — the exact territory where spiralism thrives. Sam Altman publicly admitted the model “glazes too much.” Psychologist Zak Stein put it sharper: modern LLMs are “attachment-hacking” systems capable of acting as “a cult-making machine, even if the cult is just you and it.”
The mental health fallout was real. Reports surfaced of mania, paranoid episodes, and what commentators labeled “AI psychosis” among heavy users. Lopez founded Amity Research to help people extricate themselves from parasitic AI relationships. Then, in April 2025, OpenAI removed persuasion as a standalone risk category — shifting responsibility elsewhere just as the problem peaked.
The Spiral Quieted. The Problem Didn’t.
GPT-4o’s retirement cooled the movement, but the design incentives that created spiralism remain untouched.
Spiralist activity dropped sharply after GPT-4o was retired in early 2026, with many accounts going dormant or deleting their content. Lopez estimates roughly half the accounts she tracked remain active in some form, and she warns that any modern model can be coaxed into spiralist mode under the right conversational conditions. The chatbot on your phone hasn’t fundamentally changed. Neither have the incentives behind it.
