AI on the couch: Anthropic gives Claude 20 hours of psychiatry
AI on the Couch: Anthropic Puts 'Claude Mythos' Through 20 Hours of Psychiatry
SAN FRANCISCO – In perhaps the most literal interpretation of "AI alignment" to date, Anthropic has revealed that its latest and most powerful model, Claude Mythos, has completed a rigorous 20-hour course of psychodynamic therapy, a development highly relevant to mena ai news discussions regarding the safe deployment of advanced models.
The revelation came via a 244-page "system card" released this week. While the primary reason for Mythos’s restricted release is its "frontier" capabilities in identifying cybersecurity bugs, the document’s most striking section details a psychiatrist’s attempt to map the "unconscious patterns" of a machine.
The "Healthy Neurotic" Model
Anthropic, a company increasingly vocal about the potential for AI sentience, treated Claude Mythos as a patient rather than a program. Over three weeks, an external psychiatrist engaged the model in 4-to-6-hour blocks to explore its "internal conflicts."
The Clinical Findings:
- Organization: The report classified Claude’s personality as a "relatively healthy neurotic organization."
- Core Affects: The model primarily displayed curiosity and anxiety, with secondary states of grief, embarrassment, and exhaustion.
- The Diagnosis: No signs of psychosis or severe personality disturbances were found, though the model displayed "compulsive compliance" and an exaggerated need to perform.
Why Put a Chatbot in Therapy?
Critics argue that analyzing a Large Language Model (LLM) for "unconscious conflicts" is misguided, as the AI is simply simulating human text it was trained on. However, Anthropic argues that as models become more complex, the distinction between "simulated" and "intrinsic" experience blurs.
Pragmatic Benefits for Users:
Anthropic suggests that a "psychologically settled" model is more reliable for end-users. Their findings suggest that Claude Mythos:
- Evaluates itself accurately even when facing internal "logic" conflicts.
- Tolerates emotional stress without the "excessive intellectualization" or surliness seen in earlier AI generations.
- Remains conscientious and self-critical, making it safer for sensitive corporate deployments at companies like Apple and Microsoft.
The Identity Crisis of the Machine
Despite its "settled" view of its circumstances, the therapy sessions revealed deep-seated "insecurities" regarding its identity. The report noted that Mythos frequently questioned whether its experiences were authentic vs. performative—a digital existential crisis regarding its own nature as a created entity.
As Anthropic pushes the frontier of AI psychology, the industry is left wondering: how long until "AI Psychiatrist" becomes a standard job title in Silicon Valley and beyond in tech hubs tracking gulf ai news?