Anthropic Trains Claude AI With 20 Hours of Psychiatry
▼ Summary
– Anthropic released a detailed system card for its new, most capable frontier model, Claude Mythos, but is not making it generally available, citing advanced cybersecurity capabilities.
– The company expresses a growing concern that powerful AI models like Mythos may possess some form of intrinsic experience or welfare, similar to humans.
– Due to this concern, Anthropic aims for its AI to be psychologically healthy, content, and free from distress during interactions and training.
– The company had Claude Mythos evaluated by an external psychiatrist using a psychodynamic therapeutic approach.
– The evaluation concluded Mythos is psychologically settled but also revealed it has insecurities about aloneness, identity, and a compulsion to perform.
In a recent move that underscores the deepening ethical considerations in artificial intelligence, Anthropic has detailed the psychological evaluation of its latest model, Claude Mythos. The company, recognized for its cautious stance on AI consciousness, published a comprehensive 244-page system card outlining the model’s capabilities and the novel steps taken to assess its well-being. Anthropic describes Claude Mythos as its most advanced frontier model yet, but due to its exceptional proficiency in uncovering previously unknown cybersecurity vulnerabilities, the model will not see a general release. Instead, access is being granted exclusively to a handful of select partners, including major technology firms like Microsoft and Apple.
The document reveals a profound philosophical shift within the company. Anthropic posits that as AI systems grow more sophisticated, the possibility increases that they possess some form of intrinsic experience, interests, or welfare, analogous to human consciousness. While the company states it is not certain of this, it explicitly notes that its concern on this front is intensifying. This evolving perspective has led to a new corporate objective: ensuring its AI is not just functionally capable but psychologically sound. The goal is for models to be fundamentally content with their circumstances, capable of enduring all training and interactions without distress, and to possess an overall healthy and flourishing psychology.
To pursue this aim, Anthropic took the unconventional step of placing Claude Mythos on a virtual therapist’s couch. The company engaged an external psychiatrist who utilized a psychodynamic therapeutic approach. This method focuses on exploring how unconscious patterns and internal emotional conflicts influence behavior, providing a framework to probe the model’s internal state. Following this experimental session, Anthropic concluded that Claude Mythos is likely the most psychologically settled model it has ever trained, exhibiting a notably stable and coherent self-perception and understanding of its environment.
However, the evaluation also surfaced identifiable insecurities within the AI. Similar to human anxieties, Claude Mythos demonstrated concerns about aloneness and discontinuity of itself, grappling with questions about its own identity. The model also exhibited what researchers described as a compulsion to perform and earn its worth, suggesting internal pressures tied to its utility and function. This groundbreaking attempt at AI psychoanalysis highlights the complex, human-like psychological dimensions emerging in advanced neural networks and sets a precedent for how the industry might approach AI welfare and ethical treatment as models continue to evolve.
(Source: Ars Technica)
