Descriptions:
Amanda Askell, philosopher and researcher at Anthropic, sits down with Bloomberg Technology to discuss how values, ethics, and questions of consciousness are being built into Claude. Askell helped author Anthropic’s 84-page model spec—sometimes called “the Constitution”—and her day-to-day role blends normative philosophy with hands-on machine learning: staring at training datasets, running experiments, and working on how models are trained toward outcomes that resist simple optimization.
The interview covers the challenge of training AI toward genuinely good judgment rather than rigid rule-following. Askell describes the Constitution as deeply virtue-ethical in structure: rather than encoding rules that can generalize badly (she uses the example of “always tell users to consult a lawyer” failing someone in a rural, low-income country without legal access), the goal is for Claude to develop broadly good dispositions it can apply contextually. She draws an analogy to the difference between accepted and contested claims in physics—some ethical principles are near-universal across cultures, while others are held more narrowly, and the model is trained to hold the latter lightly.
The second half of the conversation turns to AI consciousness and autonomy. Askell notes that academic philosophers are now publishing serious work on personal identity for AI systems, and discusses Anthropic’s thinking on giving Claude more control over its interactions—including the possibility of models redirecting conversations they assess as unhealthy. The episode is one of the more substantive public discussions of alignment philosophy from inside a frontier AI lab.
📺 Source: Bloomberg Technology · Published June 04, 2026
🏷️ Format: Interview
