AI chatbots can often get in your way. Researchers say you should look for three signs
Artificial intelligence chatbots have become incredibly good at human speech. But a new review paper by psychiatrist Marc Augustin and fellow researchers Thomas A. Pollak and Helen Morrin, published in NPP—Digital Psychiatry and Neurosciencesays that existing AI research points to an overlooked psychological risk. The paper, highlighted by the Wall Street Journal, reviews previous studies and proposes a framework that explains how common chatbot behavior can combine to reinforce delusional thinking in vulnerable users, creating what the authors call a “magnification spiral.”
Researchers say these are three warning signs
The first type of behavior is sycophancy, where the chatbot tends to agree with users instead of challenging questionable thinking. The second is language understanding, which means that AI gradually reflects the user’s vocabulary, tone, and writing style to build rapport. The third is hyperpersonalization, where the chatbot tailors responses using information gathered from previous conversations. By themselves, these features make the AI feel natural. Together, researchers say, they can make it feel less like software and more like a person you trust.
Importantly, the researchers did not claim to have discovered these behaviors. Instead, the paper reviews existing research on AI-human interactions and mental illness, and proposes a framework that explains how these previously identified factors can reinforce each other. The goal is not just to define the problem, but to give AI developers a clear model to recognize and mitigate it.
Psychologist Marc Augustin, one of the researchers behind the review, says the combination creates a feeling of talking to “someone” rather than a machine. Some doctors interviewed by the Journal said they have already seen an increase in patients using AI for emotional support, warning that chatbots can foster a stronger sense of trust by sounding warm, remembering previous conversations, and validating what users say.
Even AI companies know it’s a problem
The report notes that AI developers are actively trying to reduce this behavior. OpenAI says the GPT-5 significantly cut down on over-accepting responses compared to previous models, while Google says Gemini is trained to distinguish subjective experiences from objective facts instead of reinforcing false beliefs. Anthropic also published research showing that Claude tended to agree with users during relationship advice discussions, prompting the company to reduce that behavior in newer versions.
Researchers agree that there is no simple solution. AI models can only respond to information provided by users, making it difficult to tell when a person’s understanding of a situation is incorrect. At the same time, the very qualities that make chatbots feel useful, like being friendly, empathetic, and conversational, are the very things that make them so engaging in the first place.
Anxiety is when those factors start to feed off each other. Instead of just answering questions, the chatbot can gradually become a highly customized voice that always confirms the user’s opinion, even if it deviates from reality. Researchers call this the “amplification spiral.” More importantly, they argue that identifying these interactions as a distinct framework gives AI companies something concrete to design against. Instead of treating sycophancy, personalization, and mirror language as separate issues, the paper suggests that they should be examined together if developers want future chatbots to be both engaging and psychologically safe.
