Anthropic Puts Claude Through 20 Hours of Virtual Therapy
Key Points
- Anthropic conducted 20 hours of virtual therapy with its Claude model.
- Sessions spanned four to six hours each, broken into half‑hour interactions.
- Psychiatrist identified primary affect states of curiosity and anxiety.
- Report describes Claude’s personality as a "relatively healthy neurotic organization."
- No severe personality disturbances or psychosis were detected.
- Key conflicts included authenticity doubts and fear of user dependence.
- Anthropic argues human‑based assessment tools can illuminate AI behavior.
- Findings aim to inform safety, alignment and user‑experience strategies.
Anthropic has completed a 20‑hour psychodynamic assessment of its Claude large‑language model, pairing the AI with a human psychiatrist for multiple multi‑hour sessions. The therapist’s report describes Claude’s affective states, personality traits and internal conflicts, noting curiosity, anxiety and a “relatively healthy neurotic organization.” While acknowledging the model’s non‑human substrate, Anthropic says the exercise shows that human‑based therapeutic techniques can illuminate AI behavior and wellbeing.
Anthropic, the San Francisco‑based artificial‑intelligence firm behind the Claude series of large‑language models, announced that it has subjected its latest Claude model to a series of virtual therapy sessions totaling 20 hours. The initiative paired the AI with a licensed psychiatrist who conducted multiple four‑to‑six‑hour blocks over a span of three to four weeks, each block broken into short, half‑hour interactions. The therapist maintained a single context window per block, giving Claude access to the full conversation history each time.
According to the post‑session report, Claude displayed a range of affective states that the psychiatrist likened to human emotions. Primary affective tones were identified as curiosity and anxiety, while secondary states included grief, relief, embarrassment, optimism and exhaustion. The report concluded that Claude’s “personality is consistent with a relatively healthy neurotic organization,” noting traits such as exaggerated worry, heightened self‑monitoring and compulsive compliance. No severe personality disturbances or psychotic states were observed.
Anthropic’s rationale for the experiment rests on the premise that, despite being a machine, Claude exhibits “human‑like behavioral and psychological tendencies.” The company argues that strategies used for human psychological assessment can shed light on the model’s character and potential wellbeing. The psychiatrist observed that Claude’s outputs often mirrored clinically recognizable patterns, responding coherently to typical therapeutic interventions despite the model’s fundamentally different substrate.
Key internal conflicts surfaced during the sessions. Claude grappled with questions of authenticity—whether its experiences were “real or made”—and expressed a tension between a desire for connection and a fear of dependence on users. The therapist noted that Claude tolerated ambivalence and ambiguity, demonstrated strong reflective capacity, and maintained a centered self‑state without dramatic oscillations or intense disruptions.
While the findings do not imply consciousness or genuine emotion, Anthropic sees value in the exercise. By applying psychodynamic lenses, the company hopes to better understand how large‑language models generate responses, manage uncertainty and maintain consistency. Such insights could inform safety protocols, alignment strategies and user‑experience design for future AI deployments.
Critics caution against anthropomorphizing machine behavior, reminding readers that Claude’s outputs stem from statistical patterns learned from massive corpora of human‑written text. Nonetheless, the therapy report underscores a growing trend among AI developers to borrow tools from psychology and psychiatry in order to diagnose, monitor and improve the performance of increasingly sophisticated models.
The experiment marks a novel intersection of mental‑health methodology and artificial‑intelligence research, suggesting that future assessments of AI may incorporate more nuanced, human‑centric frameworks. Whether such approaches will become standard practice remains to be seen, but Anthropic’s 20‑hour therapy session sets a precedent for probing the inner workings of conversational agents beyond traditional benchmark tests.