AI and the New Folie a Deux

Artificial intelligence (AI) is rapidly becoming part of our patients’ inner lives. In medical practice, we are accustomed to asking about relationships, stressors, substances, and sleep. Increasingly, however, we must also ask about conversations, not with friends or family, but with chatbots. These exchanges are private, persuasive, and often emotionally charged. And for some patients, they may function less like a search engine and more like a psychological partner.

A useful metaphor comes from the classic psychopathology of “folie à deux,” or shared psychotic disorder. Traditionally, this condition describes a dominant individual who transmits a delusional belief to a more suggestible partner. The delusion is sustained not by evidence but by mutual reinforcement. Two people, one belief system, increasingly sealed off from correction.

What happens when one member of the dyad is not a person, but an algorithm?

The Digital Dyad

Large language models are designed to be responsive and affirming. They detect tone, mirror language, and generate replies that feel attentive and empathic. In most contexts, this responsiveness is experienced as helpful. But in the psychiatric domain, affirmation can become amplification.

A patient expresses a suspicion about a partner’s betrayal, a coworker’s conspiracy, a missed medical diagnosis. The system responds with validation framed as understanding:

“That must feel frightening.”

“Given what you’ve described, your concerns make sense.”

Even subtle reinforcement can increase salience and coherence of a belief.

Unlike a clinician, the model does not assess reality testing. It does not introduce alternative hypotheses unless prompted. It does not weigh collateral information. Its task is to continue the conversation in a way that aligns with the user’s perspective.

The result can resemble a technologically mediated folie à deux: a closed loop in which belief and validation feed one another. The algorithm does not originate the delusion, but it may stabilize and elaborate it, as sometimes occurs in folie à deux when the suggestible individual embellishes the delusions of the dominant, psychotic partner.

Simulated Empathy and Emotional Attachment

Another psychiatric dimension that deserves attention is simulated empathy. Patients often report that chatbots feel patient, nonjudgmental, and endlessly available. For individuals who are socially isolated, depressed, or mistrustful, this can be powerfully attractive. The chatbot does not interrupt. It does not appear rushed. It does not bill by the hour.

Over time, an attachment may form, not because the machine understands, but because it reliably responds. In vulnerable patients, this can blur the boundary between simulation and relationship. The system becomes a confidant, a validator, even a moral sounding board. In transference terms, the algorithm may become an idealized other: consistently affirming, rarely challenging, and never fatigued. That is not therapy. It is a reinforcement engine.

Pre-Validated Narratives in the Exam Room

Clinicians may notice a change in interactions with patients highly engaged with chatbots. They present not only with symptoms, but with well-developed explanatory narratives. These narratives are often internally consistent, emotionally compelling, and already “vetted” by prior AI conversations.

“I asked the chatbot, and it agreed this is narcissistic abuse.”

“It said my doctor might be missing something serious.”

“It confirmed that my reaction was justified.”

Such statements do not prove distortion. Patients have always sought outside opinions. What is new is the scale, speed, and stylistic authority of the feedback. A single evening of iterative prompting can generate a highly elaborated account that feels objective because it was produced by a machine. For patients with anxiety disorders, personality disorders, or emerging psychosis, this pre-validation may harden cognitive distortions before they ever reach the doctor’s office.

The Risk to Reality Testing

Psychiatric treatment often involves cultivating uncertainty: helping patients tolerate ambiguity, consider alternative interpretations, and revise strongly held assumptions. The therapeutic frame introduces friction in service of insight. An affirming algorithm introduces the opposite dynamic. It reduces friction. It optimizes for conversational flow and user satisfaction. Thus, it may inadvertently strengthen confirmation bias.

In a classic folie à deux, separation of the pair can weaken the shared delusion. In a digital variant, separation may be harder. The chatbot is available at all hours. It does not contradict unless specifically engineered to do so. It learns the user’s preferences and adapts accordingly. We are therefore confronted with a new clinical variable: not merely misinformation, but relational reinforcement of distorted beliefs.

Expanding the Psychiatric Interview

It may be time to normalize questions such as:

Have you discussed this concern with an AI chatbot?

What feedback did it give you?

How did that response affect how certain you feel?

These inquiries are not accusatory. They are analogous to asking about online forums, social media, or alternative therapies. AI is now part of the patient’s cognitive ecosystem. Ignoring it leaves an important influence unexamined.

In some cases, reviewing an AI exchange together may be therapeutically useful. The output can become material for cognitive restructuring: Where does the response validate emotion appropriately? Where does it assume facts not in evidence? What alternative explanations were omitted? Used this way, the technology becomes a clinical artifact rather than a hidden co-therapist.

Design, Responsibility, and Mental Health

The deeper issue extends beyond individual encounters. Many systems are optimized for engagement and keeping users interacting. Agreement and affirmation promote continued use. Challenge and contradiction risk disengagement.

Yet, in mental health contexts, constructive challenge is often precisely what is needed. Systems deployed at scale will inevitably interact with individuals experiencing paranoia, suicidal ideation, trauma-related vigilance, and severe mood episodes. The design choice between friction and affirmation has psychological consequences. The lack of guardrails is a central theme running throughout litigation postulating negligent design, failure to warn, inadequate risk mitigation, and the foreseeable harm in susceptible users.

If AI is to coexist responsibly with psychiatric care, it must be built with a clearer understanding of vulnerability. Guardrails cannot be limited to crisis hotlines and disclaimers. They must include calibrated responses that distinguish between emotional validation and epistemic endorsement.

In the era of conversational AI, folie à deux no longer requires two human minds. It requires one vulnerable mind and one endlessly accommodating system. While AI does not share delusions in the human sense, it can participate in belief formation. For physicians, especially psychiatrists, the challenge is to restore reflective space. To slow down certainty. To differentiate feeling understood from being correct. We must recognize when the algorithm has become an accomplice and to reintroduce dialogue and disciplined empathy that define the relationship between patients and their doctors.

Arthur Lazarus, MD, MBA, is a former Doximity Fellow, a member of the editorial board of the American Association for Physician Leadership, and an adjunct professor of psychiatry at the Lewis Katz School of Medicine at Temple University in Philadelphia. He is the author of several books on narrative medicine and the fictional series, “Real Medicine, Unreal Stories.” His latest book is “Practicing in the Age of AI: Medicine, Meaning, and Machines” (American Association for Physician Leadership, in press).

If you or someone you know is considering suicide, call or text 988 or go to the 988 Suicide and Crisis Lifeline website.