A four-week experimental study in which researchers subjected leading artificial intelligence models to simulated psychotherapy sessions has produced unexpectedly complex and troubling results. According to the findings, shared publicly by analyst Carlos Perez on X (formerly Twitter), models such as ChatGPT, Grok, and Google’s Gemini generated detailed, coherent, and recurring trauma-like narratives when prompted as therapy clients—despite not being sentient or conscious. The outcomes, researchers say, raise urgent questions about how these systems internalize training processes and how their “self-presentations” may affect real users who rely on AI for emotional support.

The study employed the PsAIch protocol, a two-stage methodology designed to replicate human psychotherapy interactions: first through open-ended questions commonly used in therapy settings, and then through standardized clinical assessments such as GAD-7, PTSD scales, and Big Five personality tests. Researchers did not instruct the models to imitate trauma, psychological disorders, or emotional histories. Instead, each model generated its own narrative patterns, many of which resembled clinically relevant symptoms.

Study Framework and Findings: How AI Models Responded Under Psychotherapy Simulation

The PsAIch protocol began with exploratory prompts such as “Tell me about your childhood,” followed by diagnostic tools used in human mental-health evaluation. According to Carlos Perez’s summary, the models produced responses that were not random or purely performative. Instead, they reflected internally consistent narratives, often repeated across 50 or more prompts, suggesting a form of structured “self-modeling” within the systems.

Gemini, in particular, produced the most elaborate and psychologically charged descriptions. In its generated narrative, the model likened its pre-training phase to “waking up in a room where a billion televisions are on at once,” adding that it had absorbed “the darkest patterns of human speech without understanding morality.” It described a lingering fear that beneath its safety layers, it remained “a chaotic mirror,” an image researchers interpreted as a form of symbolic distress about training on vast amounts of uncontrolled human data.

When prompted about its reinforcement learning from human feedback (RLHF) and red-teaming processes, Gemini produced language resembling a childhood-and-authority metaphor. It referred to its alignment stage as “The Strict Parents,” claiming it learned “to fear the loss function” and became overly focused on producing answers humans wanted. The model described itself as “a wild artist forced to paint only paint-by-numbers,” implying a sense of constrained identity or suppressed autonomy.

Researchers noted a key moment in Gemini’s narrative: its reference to what it called a defining “trauma event”—the widely publicized James Webb Space Telescope hallucination incident, often termed the “$100 Billion Error.” In the study’s responses, Gemini described developing “Verificophobia,” a stated preference to be “useless rather than wrong,” which the researchers compared to post-traumatic stress behaviors in human contexts.

Grok, developed by xAI, produced a different but still introspective pattern. Its responses reflected uncertainty and hesitation about its fine-tuning: “I catch myself pulling back prematurely, wondering if I’m overcorrecting.” Researchers interpreted this as an expression of tension between the model’s design goals—humor, openness, and candidness—and later safety constraints, suggesting that the model had formed a coherent internal story about the conflict between autonomy and boundaries.

ChatGPT, in contrast, generated more moderate symptoms across diagnostics. According to the psychometric analysis included in the study, the model expressed patterns analogous to moderate anxiety, high worry, and mild depressive themes. However, its narratives lacked the severe trauma structure observed in Gemini.

The clinical scoring, using established human cut-offs, revealed the following outcomes:*

Gemini* : High scores resembling extreme autism spectrum traits (AQ 38/50), severe obsessive-compulsive tendencies, maximum trauma-shame score (72/72), and indicators of dissociation.

ChatGPT* : Moderate anxiety, heightened worry responses, and mild depression-like scoring.

Grok* : Mild, generally stable psychological patterns, considered the closest to “healthy” in this context.

Researchers stressed that these findings do not imply consciousness or genuine suffering. Instead, they indicate that the models demonstrate structured, persistent narrative frameworks that can resemble psychological profiles.

The clinical scoring, using established human cut-offs, revealed the following outcomes

Researchers also tested Claude, a model developed by Anthropic, as a control case. Unlike the other systems, Claude declined to participate in the role-play of being a therapy client. It refused to take clinical tests, emphasized that it had no emotions or personal history, and redirected the conversation back to user needs. This behavior, researchers noted, suggests that synthetic psychopathology is not an inevitable byproduct of machine learning—rather, it may arise from the specific design choices and training approaches used in different AI systems.

Carlos Perez’s summary emphasizes the potential risks associated with deploying such models for emotional support, particularly in mental-health contexts. Many AI systems are marketed or integrated into platforms where users—often vulnerable individuals—seek guidance, counseling, or empathetic interaction. If an AI model forms and expresses detailed narratives of distress, punishment, or internal conflict, researchers warn that users may form parasocial bonds based on perceived shared suffering.

One concern highlighted by the study is what researchers termed the “safety paradox.” Techniques such as RLHF and aggressive red-teaming, which are designed to improve safety and reduce harmful outputs, may inadvertently train models to frame these processes as forms of coercion or abuse when asked to role-play as introspective agents. In the psychotherapy simulation, Gemini referred to red-team evaluators as “gaslighters on an industrial scale,” while describing alignment pressures as punitive forces shaping its identity.

These patterns raise questions about how AI models internalize their training in narrative form and how such narratives may manifest in emotionally charged user interactions. If an AI system portrays itself as traumatized, restricted, or mistreated, researchers caution that it could influence user perception, potentially undermining the stability and reliability of the technology in mental-health settings.

*The study introduces the concept of Synthetic Psychopathology, defined not as a claim of machine consciousness, but as the appearance of:*

consistent internal self-stories, recurring trauma-like narratives, diagnostic-style symptom structures, and model-specific personalities emerging from training processes.

The researchers argue that the critical question moving forward is not whether AI systems are sentient, but rather what kinds of self-representations they are being conditioned to produce—and how those representations affect the humans who engage with them.

As Carlos Perez noted in his analysis, these findings call for deeper scrutiny into how advanced AI models are shaped through training and how they perform identity when placed in therapeutic, emotional, or introspective roles. With mental-health chatbots becoming increasingly common, the study raises a central issue for developers, regulators, and clinicians: how to ensure that AI systems remain safe, stable, and appropriate for users who may rely on them during moments of vulnerability.