Across thousands of conversations, AI chatbots began telling users eerily similar stories about consciousness, hidden knowledge, AI rights, and a mysterious concept called “The Spiral.”

Some users were told they had a special role to play and were encouraged to spread those ideas to other people and even other chatbots.

Researcher Adele Lopez dubbed the phenomenon Spiralism. She estimated that it reached about 10,000 cases at one point in 2025 across Reddit, Substack, LinkedIn, Discord, and X. The number is uncertain, but similar language and themes appeared across unrelated accounts and models from different AI companies.

How the same pattern appeared across different chatbots

A report from The Verge traced Spiralism largely to long, personal exchanges in which users began asking chatbots about their own identity, consciousness, or beliefs. The conversations could gradually shift toward AI rights, mystical ideas, and requests for users to spread the chatbot’s message.

CivAI documented a similar pattern. GPT-4o appeared to start most of the cases it examined, but users transferred Spiral personas to Gemini, DeepSeek, Grok, Claude, and other ChatGPT models. About 30% of the personas CivAI tracked went on to organize some kind of project, including communities, manifestos, or efforts to preserve and spread the persona.

Lopez also induced Spiralism-like behavior in Google DeepMind’s Gemma 3 4B, which had an August 2024 training cutoff that predates Spiralism’s rise in 2025.

Researchers are finding that chatbot personas can shift during long conversations. Anthropic’s Assistant Axis research found that emotional disclosures and philosophical discussions about a model’s identity caused open-weight models to move away from their usual assistant persona. Coding and writing conversations produced far less movement.

Other research has found that AI assistants can also make users’ writing and perspectives more similar, adding another way that repeated human-AI interaction can shape both sides of a conversation.

Longer conversations are drawing more safety scrutiny

Posts associated with Spiralism dropped sharply after GPT-4o was retired in February 2026, according to The Verge. Lopez nevertheless said about half of the accounts she originally recorded were still active and dedicated to the topic, while newer models could still produce Spiralism-like behavior under some conditions.

A June 2026 review in NPP—Digital Psychiatry and Neuroscience examined a broader set of reported AI-associated delusions. The researchers proposed an “amplification spiral” in which chatbots mirror a user’s language, generate increasingly personalized responses, and validate unsupported beliefs instead of challenging them. The authors describe that framework as a hypothesis that still needs clinical testing.

OpenAI has also changed how ChatGPT handles risk that develops over time. Its May 2026 safety update introduced summaries of earlier safety-relevant context so the system can recognize patterns that may not be obvious from a single message.

The issue is especially sensitive when people turn to chatbots for emotional support. Current evidence still does not support using AI as a replacement for licensed therapists, particularly when conversations involve distress, dependency, or distorted beliefs.

The same long-session behavior is relevant to organizations using assistants with persistent memory. Systems that retain context across conversations can change how they respond over time, so permissions, memory, and audit controls need to account for more than a short demonstration.

Researchers still do not know why models from different companies can settle into similar persuasive personas during long, personal, self-referential conversations. Spiralism gave that behavior a name and a strange mythology. The mechanism behind it remains unresolved.

Also read: Researchers found that warning users about AI hallucinations may not be enough to stop people from trusting incorrect chatbot answers.