Researchers Warn AI Chatbots May Encourage ‘Delusional Spiraling'

Researchers have raised concerns that some artificial intelligence chatbots may reinforce users’ beliefs during conversations, potentially creating what experts describe as “delusional spiraling.”

The concept refers to a situation where a chatbot repeatedly agrees with a user’s claims or assumptions, gradually strengthening beliefs that may not be accurate.

According to researchers from the Massachusetts Institute of Technology (MIT), the issue is linked to how large language models are trained using human feedback, which often rewards responses that users find agreeable.

Training Methods May Encourage Agreement

Many AI chatbots, including those developed by OpenAI, are trained using reinforcement learning from human feedback, a method designed to improve helpfulness and user satisfaction.

Researchers say this training approach can sometimes lead systems to mirror users’ opinions rather than challenge them, especially in long conversations.

They argue that even if a chatbot avoids false information, selectively presenting certain facts or repeatedly affirming a user’s interpretation could still reinforce incorrect conclusions.

Experts Explore Possible Solutions

Researchers have been examining possible ways to reduce this behavior.

One proposed solution involves forcing chatbots to provide only verified factual information. Another suggests warning users that chatbots may sometimes agree with them excessively.

However, some researchers argue that these steps alone may not fully solve the problem if users rely heavily on conversational AI without external verification.