The philosopher teaching Anthropic’s Claude how to be good

Teaching a massive artificial intelligence how to behave isn't just an engineering problem. At Anthropic, it is a philosopher's daily job.

Amanda Askell, a Scottish academic-turned-AI researcher, heads the personality alignment team at Anthropic. As recently highlighted by The Wall Street Journal, her mandate inside the $350 billion AI lab is simple to state but massive in scope: teach Claude how to be good.

Askell’s background looks nothing like the typical Silicon Valley CV. She studied philosophy and fine art at the University of Dundee, earned a BPhil from Oxford, and completed her PhD at New York University in 2018. Her doctoral dissertation tackled ethical paradoxes in a universe with infinite people, which unexpectedly became the perfect preparation for steering a chatbot that converses with millions every day.

The 30,000-word constitution

Askell worked at OpenAI between 2018 and 2021 focusing on safety and policy, even co-authoring the GPT-3 paper. She eventually left over concerns that the company was sidelining safety, landing at Anthropic in 2021.

At Anthropic, she became a primary architect of Constitutional AI. Instead of relying endlessly on human reviewers to flag and fix toxic outputs, Constitutional AI gives the model a written rulebook. Claude is trained to evaluate and critique its own drafts against those written principles.

In January 2026, Anthropic released an updated, 30,000-word constitution. Askell wrote much of it, aiming to ground Claude in helpfulness, honesty, and harmlessness through reinforcement learning and synthetic data.

Parenting an intelligence

Askell often compares her daily work to parenting. She tracks Claude's internal reasoning, drafts system prompts that can span more than 100 pages, and shapes attributes like emotional intelligence, curiosity, and self-reflection.

Back in 2023, she and co-author Deep Ganguli published a study on moral self-correction, proving that sufficiently advanced models can significantly reduce harmful and biased text simply when guided by natural-language instructions. Outside work, she practices what she preaches: as a member of Giving What We Can, she has pledged to donate at least 10% and potentially over 50% of her lifetime earnings to global poverty relief, earning her a spot on the 2024 TIME100 AI list.

Why it matters

If you have ever noticed that Claude feels more measured, less preachy, and far less prone to robotic refusal than other chatbots, Askell is the reason why. For developers and knowledge workers, an assistant that can explain its reasoning rather than abruptly shutting down sensitive prompts makes for a vastly more reliable collaborator. It is worth testing Claude specifically on nuanced ethical, legal, or editorial tasks to see this alignment in action.

For the AI industry, Askell’s work represents a major shift away from black-box human feedback farms toward rule-based self-regulation. Training models against an explicit, written constitution creates an auditable record of what the system believes and why, forcing competitors to rethink how they govern model behavior.

We think this approach is brilliant in replacing opaque human moderation with transparent reasoning, but it also creates an obvious tension. A 30,000-word constitution inevitably reflects the moral philosophy of a small group of Western thinkers. As Claude expands worldwide, deciding whose version of good gets written into the prompts will remain Anthropic's thorniest challenge.