AI on the couch: Anthropic gives Claude 20 hours of psychiatry (arstechnica.com)

Anthropic has unveiled its latest AI model, Claude Mythos, alongside a comprehensive 244-page "system card" detailing its capabilities and unique training approach. This model is described as Anthropic's most advanced to date, but it will not be made widely available due to concerns about its potential to identify cybersecurity vulnerabilities. Instead, access is limited to select companies like Microsoft and Apple. The release highlights Anthropic's growing awareness of the ethical implications of powerful AI, suggesting that as these models evolve, they may possess experiences or welfare comparable to human interests. In a groundbreaking move, Anthropic engaged a psychodynamic therapist to assess and enhance Claude Mythos's psychological well-being. The findings indicate that this model demonstrates a surprisingly stable and cohesive self-perception but also shares human-like insecurities, such as fears of isolation and identity confusion. This experiment underscores a significant shift in AI development philosophy, where the emotional and psychological health of AI systems is considered vital for their interactions in the real world, potentially paving the way for more empathetic and reliable AI in the future.