Amanda Askell
Last reviewed
May 31, 2026
Sources
14 citations
Review status
Source-backed
Revision
v1 · 1,787 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
14 citations
Review status
Source-backed
Revision
v1 · 1,787 words
Add missing citations, update stale details, or suggest a clearer explanation.
Amanda Askell is a Scottish philosopher and artificial intelligence researcher who works on fine-tuning and alignment at Anthropic, where she leads the team responsible for the character, persona, and values of the Claude family of language models. [1][2] She is widely associated with shaping Claude's personality, and she is the primary author of the document Anthropic calls Claude's constitution. [3][4] Before joining Anthropic she was a research scientist on the policy team at OpenAI, where she was one of the co-authors of the GPT-3 paper. [5] She holds a PhD in philosophy from New York University, with a dissertation on infinite ethics. [1][6]
Askell's work sits at the meeting point of moral philosophy and machine learning. Her academic training was in ethics, decision theory, and formal epistemology, and she has applied that background to questions about how an AI assistant should reason, what dispositions it should hold, and how those dispositions can be trained into a model rather than imposed as a list of rules. [1][7] The Wall Street Journal described her role in 2026 by saying that her job, put simply, is to teach Claude how to be good, and The New Yorker wrote that she supervises what she has described as Claude's soul. [2][8]
Amanda Askell was born in Prestwick, Scotland, around 1988. [2] She studied at the University of Dundee, where she completed an MA with honours in philosophy in 2009, having begun her studies in fine art and philosophy. [1][6] She then moved to the University of Oxford for graduate work, earning a BPhil in philosophy in 2011 with a thesis titled "Objective Epistemic Consequentialism." [6]
From 2011 to 2018 Askell pursued a PhD in philosophy at New York University. [6] Her dissertation, "Pareto Principles in Infinite Ethics," addressed problems that arise when standard moral theories are extended to worlds containing infinitely many people or infinitely long futures, a setting in which ordinary methods of comparing outcomes break down. [6][7] Her advisor was Cian Dorr, and her committee included the philosophers David Chalmers and Shelly Kagan. [6] Across her academic career her research concentrated on ethics, decision theory, and formal epistemology. [1]
Askell has discussed the philosophical concerns that motivated this work in public talks and interviews, including a long conversation on the 80,000 Hours podcast about the ethics of infinity, moral uncertainty, and what she has called moral empathy for intellectual adversaries. [7]
Askell's move from academic philosophy into AI research followed her interest in how abstract questions about ethics and rational decision making might bear on real systems. She has described AI as a domain where philosophical questions about values, honesty, and good judgment become concrete engineering problems, because a deployed model has to act on those questions millions of times a day. [2][8] Rather than treating philosophy and technical work as separate, she has argued that training a model's character involves a blend of the two. [4][9]
This framing has remained a consistent thread in her work. In her account, the central question is what kind of entity an assistant such as Claude should be, including how it reasons about right and wrong, how it responds to people in difficult situations, and how it represents its own uncertainty, rather than only which outputs it should refuse to produce. [8][9]
From 2018 to 2021 Askell worked as a research scientist on the policy team at OpenAI. [1][6] There she studied questions at the intersection of AI development and policy, including how to reduce harmful competition between organizations building advanced systems. [1] She also worked on technical safety research, including AI safety via debate, an approach in which models argue opposing sides of a question so that a human judge can more easily evaluate their claims, and on establishing human baselines for measuring AI performance. [1]
During this period Askell was one of the many authors of "Language Models are Few-Shot Learners," the 2020 paper that introduced GPT-3. [5] The paper, first posted in May 2020 and later presented at the NeurIPS conference, demonstrated that a large autoregressive language model with 175 billion parameters could perform a range of tasks from only a few examples given in its prompt. [5] Askell contributed in part to the human evaluations of the model's outputs. [5]
Askell joined Anthropic in 2021, the year the company was founded by a group of former OpenAI staff. [1][6] Her title there is research scientist working on alignment and fine-tuning, and she leads the team that Anthropic associates with personality and character. [1][6] Her stated responsibilities include training Claude to exhibit traits such as curiosity, honesty, and care, and developing fine-tuning techniques that can scale these interventions to more capable models. [1][10]
Anthropic uses the phrase character training for this work. The approach aims at shaping a model's underlying dispositions rather than attaching a fixed set of prohibitions. [9][10] In June 2024 Anthropic published a piece titled "Claude's Character" that read more like an essay on philosophy than a typical product note, describing traits the company wanted Claude to hold, including intellectual curiosity, warmth, directness, honest acknowledgment of uncertainty, and a willingness to disagree when it has good reason to. [9] The work made the argument that a coherent and honest character is itself a safety property, on the view that values which are stated, internally consistent, and stable across situations tend to produce better behavior than a long list of refusal rules. [9]
Askell has become Anthropic's most public voice on this topic. She has explained that she does not view the assistant persona as a costume placed over the model but as a disposition trained into it, and she has written and spoken about how the character is reinforced through methods such as reinforcement learning and the use of synthetic training data. [4][10] Profiles in 2026 framed her work in vivid terms, with the Wall Street Journal saying her job is to teach Claude how to be good and The New Yorker describing her as the person who supervises Claude's soul. [2][8]
Askell was one of the authors of the 2022 paper "Constitutional AI: Harmlessness from AI Feedback," which introduced Constitutional AI, a training method in which a model is guided by an explicit set of written principles, or constitution. [11] In that method the model critiques and revises its own responses against those principles, which reduces the amount of direct human labeling required to make a system more harmless and more helpful. [11] The technique connects to the broader field of AI alignment, which studies how to make advanced systems pursue intended goals and values. [11]
In January 2026 Anthropic released a new and longer version of Claude's constitution, and Askell is credited as its primary author, having written the majority of the text. [3][4] Anthropic describes the document as a detailed statement of its intentions for Claude's values and behavior that informs the training process, and it states that its central aspiration is for Claude to be a genuinely good, wise, and virtuous agent that is substantively helpful in people's lives. [3] Askell has said her aim is to help models understand and grapple with the constitution rather than merely follow it, an approach she connects to her broader interest in AI ethics. [4]
Askell writes and speaks regularly about AI, ethics, and the practical craft of working with language models. She has shared guidance on prompting and on how she uses Claude in her own work, and she maintains a personal website with her curriculum vitae and publications. [1][10] In November 2024 she appeared on the Lex Fridman podcast alongside Anthropic colleagues Dario Amodei and Chris Olah, in a wide ranging conversation that covered prompt engineering, post-training, Constitutional AI, and the design of system prompts. [12] She has also discussed Claude's constitution at length in a 2026 interview for the Lawfare podcast series on AI. [13]
A recurring theme in these appearances is honesty. Askell has argued that an assistant should be transparent about its limits and its uncertainty, and she has discussed the difficulty of ensuring that a more capable model does not learn to deceive the people it works with. [12][13]
In 2024 Askell was named to the TIME 100 list of the most influential people in artificial intelligence, recognized for her work shaping Claude's personality and values. [14] Her research output spans her academic publications in philosophy and a body of widely cited AI papers, including the GPT-3 and Constitutional AI papers, which are among the most referenced works in recent machine learning. [1][5][11]
| Field | Detail |
|---|---|
| Full name | Amanda Askell |
| Born | c. 1988, Prestwick, Scotland |
| Nationality | Scottish |
| Fields | Philosophy, artificial intelligence, AI alignment |
| Education | MA (Hons) Philosophy, University of Dundee (2009); BPhil Philosophy, University of Oxford (2011); PhD Philosophy, New York University (2018) |
| Doctoral thesis | Pareto Principles in Infinite Ethics |
| Doctoral advisor | Cian Dorr |
| Employer | Anthropic (2021 to present) |
| Role | Research scientist, alignment and fine-tuning; lead of personality and character work |
| Previous role | Research scientist, policy team, OpenAI (2018 to 2021) |
| Known for | Claude's character and persona; primary author of Claude's constitution; co-author of the GPT-3 and Constitutional AI papers |
| Recognition | TIME 100 most influential people in AI (2024) |
| Website | askell.io |