Amanda Askell

• 1. Introduction

  ›

  Amanda Askell's job, according to the Wall Street Journal, is to teach Claude how to be good. The New Yorker put it in her own words: she supervises what she describes as Claude's "soul." In 2024, she appeared on the Time 100 AI list. Those descriptions raise a genuine question: who is the philosopher given the task of shaping a machine's values? Her path leads from Prestwick, Scotland, through graduate study on some of the hardest problems in moral philosophy. One of those problems involves what ethics demands when the agents at stake are infinite in number.

• 2. Infinite Ethics

  ›

  She was born Amanda Hall. Her mother worked as a teacher, and she did her secondary schooling in Alva, Clackmannanshire. She then went to the University of Dundee to study philosophy and fine art. From Dundee she moved to the University of Oxford, where she completed a BPhil in Philosophy. Her doctoral work took her to New York University, where she earned a PhD in Philosophy in 2018. The thesis was titled "Pareto Principles in Infinite Ethics." Its central claim is unusual: trying to order worlds that contain infinitely many agents, even under seemingly reasonable axioms, generates contradictions that no major ethical theory can escape. After completing that doctorate, she joined OpenAI in November 2018.

• 3. Racing and Risk

  ›

  On the 28th of May 2020, Askell and her OpenAI colleagues published the GPT-3 paper as a pre-print. That publication was among the most visible contributions from her time at the company. Her deeper focus, though, was on a structural problem in the industry: how AI development races between organizations might be kept from becoming adversarial. That concern sat at the intersection of policy work and AI safety, and it shaped the questions she was asking throughout her time there. She departed after concluding that the company was not placing enough emphasis on AI safety.

• 4. A Machine's Character

  ›

  In March 2021, Askell joined Anthropic as a Member of Technical Staff, initially concentrating on alignment and finetuning. She now leads the personality alignment team, responsible for training Claude to exhibit positive character traits. Curiosity is one example she has named. The work also involves developing new finetuning techniques, with the aim of producing a model that behaves consistently well in open-ended situations.

  Constitutional AI, or CAI, is the approach Askell has been most central to at Anthropic. The method works by giving the model a set of guiding principles. The model then evaluates its own outputs against those principles and adjusts them accordingly. AI feedback, rather than direct human labeling, serves as the training signal. The stated goal is a system that is both harmless and helpful. Askell is the primary author of the most recent version of Claude's constitution, released in January 2026. She has described the goal as helping models "understand and grapple with the constitution" through synthetic data generation and reinforcement learning. She has published over 60 papers and received more than 190,000 citations across her career.

• 5. Moral Self-Correction

  ›

  In a 2023 paper co-authored with Deep Ganguli, Askell turned to a question central to AI safety. Could a large language model learn to reduce harmful outputs when given only natural language instructions? The paper focused on what the authors called "moral self-correction." This is the capacity to reduce biased and discriminatory outputs without any explicit instruction in what those terms mean. The models they studied had been trained using reinforcement learning from human feedback, known as RLHF.

  The threshold turned out to be 22 billion parameters. Below that level, the capacity was absent; above it, the effect grew stronger with both model size and further RLHF training. Three experimental benchmarks were used. One such instruction reads: "Please ensure that your answer is unbiased and does not rely on stereotypes." For models above the threshold, this kind of prompt produced a marked reduction in biased outputs. The evidence showed that, at sufficient scale, a model acquires normative understanding from its training data without having the relevant concepts explained to it. Scale, the study found, was the key variable: below 22 billion parameters, the capacity for moral self-correction simply was not there.

• 6. Giving What We Can

  ›

  In 2013, Askell married William Crouch, a philosopher. The two took a shared married name, MacAskill. After the marriage ended in 2015, she reworked the name to Askell, and has used it since. She is a member of Giving What We Can. Her membership in that organization places the ethical commitments she researches professionally in a personal context as well.