Amanda Askell

From 2005 to 2009 she studied philosophy at the University of Dundee, earning an MA (Hons).⁠[1]

She began as a fine art and philosophy student at the Duncan of Jordanstone art school.⁠[2]

She read for a BPhil in philosophy at the University of OxfordUniversity of OxfordThe University of Oxford is a university in Oxford, England, where teaching existed in some form by 1096; it describes itself as the oldest university in the English-speaking world and is organised as a federal system of the university and dozens of self-governing colleges.Open the full entry; her thesis, “Objective Epistemic Consequentialism”, is dated June 2011.⁠[3]

From 2011 to 2018 she pursued a PhD in philosophy at New York University, advised by Cian Dorr, with David Chalmers and Shelly Kagan on her committee.⁠[1]

On June 4, 2017 she gave a talk titled “The Moral Value of Information”, and on August 6 she appeared on the Rationally Speaking podcast on “Pascal’s Wager and other low risks with high stakes”.⁠[4]

In August 2017 the Center for Values and Social Policy at the University of Colorado Boulder announced that she and Tyler John, a PhD student at Rutgers University, shared that year’s Young Ethicist Prize for their paper “Moral Offsetting”, awarded at the tenth Rocky Mountain Ethics Congress.⁠[5]

The paper offers a conceptual analysis of moral offsetting and argues that at least some offsetting is permissible.⁠[5]

An 80,000 Hours podcast interview released in September 2018 introduced her as having just defended her NYU doctoral thesis, “Pareto Principles in Infinite Ethics”, and as a blogger at rationalreflection.net.⁠[6]

In the episode she discussed “moral empathy”, arguing that people are often unable to see an issue from the point of view of those who disagree with them.⁠[6]

After September 2018 she joined OpenAIOpenAIOpenAI is an artificial intelligence research and product organization founded in 2015 that develops the GPT model series, ChatGPT, Codex and a developer API, and conducts research on image, speech, video and intelligent agents. After its 2025 restructuring, the nonprofit OpenAI Foundation continues to control the for-profit OpenAI Group PBC. Its growth has involved large-scale fundraising and compute partnerships, and has also been accompanied by controversies over governance, copyright, privacy and AI safety.Open the full entry as a research scientist working on AI policy.⁠[7]

A Berkman Klein Center profile from that period described her work at OpenAI on ethics and policy, including responsible AI development, cooperation and safety via debate, and noted that she had written for Vox, The Guardian and Quartz.⁠[8]

On February 19, 2019 she and Geoffrey Irving published the article “AI Safety Needs Social Scientists” in Distill.⁠[9]

An 80,000 Hours episode released in March 2019 featured her with OpenAI colleagues Miles Brundage and Jack ClarkJack ClarkAnthropic co-founder and Head of Public Benefit.Open the full entry, recorded on the day OpenAI released the GPT-2 language model.⁠[7]

She said she was then focusing on development races between AI developers and argued that such races need not be highly adversarial.⁠[7]

In July 2019 she, Miles Brundage and Gillian Hadfield posted the paper “The Role of Cooperation in Responsible AI Development” on arXiv, with her as first author.⁠[10]

That August she was among the co-authors of Irene Solaiman and colleagues’ paper “Release Strategies and the Social Impacts of Language Models”.⁠[11]

She is one of the authors of the GPT-3 paper “Language Models are Few-Shot Learners,” published in May 2020.⁠[3]

In June 2020 she published “The optimal rate of failure” on her blog, arguing that for many endeavours the optimal rate of failure is not zero and that excessive risk aversion is costly.⁠[12]

That December she published “In AI ethics, ‘bad’ isn’t good enough”, distinguishing “pro tanto” harms of AI systems from all-things-considered judgments.⁠[13]

In 2021 she joined Anthropic as a research scientist in alignment finetuning.⁠[1]

She has described her team as training models to be more honest and to have good character traits, and developing new finetuning techniques that can scale to more capable models.⁠[2]

In December 2021 she posted “A General Language Assistant as a Laboratory for Alignment” on arXiv as first author, with Yuntao Bai, Anna Chen and others.⁠[14]

On December 15, 2022 Yuntao Bai, Saurav Kadavath and colleagues released the paper “Constitutional AI: Harmlessness from AI Feedback”, of which she is a co-author.⁠[15]

In February 2023 she, Deep Ganguli and others released “The Capacity for Moral Self-Correction in Large Language Models”, with her as second author.⁠[16]

She explained that such instructions could be included in a model’s “constitution” and that one could “train the model to do what you want”.⁠[17]

In April 2024 she published the philosophy paper “The Moral Inefficacy of Carbon Offsetting”.⁠[3]

On June 8, 2024 Anthropic wrote that Claude 3 was its first model to add “character training” to alignment finetuning, aiming to give Claude traits such as curiosity, open-mindedness and thoughtfulness.⁠[18]

The post was accompanied by a roughly 37-minute conversation between her and Stuart Ritchie on the philosophy of designing an LLM’s personality.⁠[19]

She appeared in episode 452 of the Lex FridmanLex FridmanLex Fridman is a Soviet-born, US-raised computer scientist and podcast host who works as an MIT research scientist on human-AI interaction and autonomous driving. His Lex Fridman Podcast is known for hours-long conversations with guests including heads of state and technology leaders.Open the full entry Podcast, whose page describes her as an AI researcher working on Claude’s character and personality; her segment covered philosophy, talking to Claude, prompt engineering, post-training, system prompts, character training and AI consciousness, among other topics.⁠[20]

She concluded that an AI should be flexible rather than morally rigid, saying “People are quite dangerous when they have moral certainty.”⁠[21]

She said she had met effective altruists at Anthropic but that it was not a theme of the organization.⁠[21]

In May 2025, at Anthropic’s first developer event in San Francisco, she explained how the company tunes Claude’s character, likening the ideal persona to “a well-liked traveler who can adjust to local customs”.⁠[22]

The same month, the philosophy-profession site Daily Nous described her as head of “character training” at Anthropic.⁠[23]

On December 1, 2025, after users extracted an internal document from a Claude model, she confirmed on X that it was based on a real document on which Anthropic had trained Claude, and said it had become known internally as the “soul doc”.⁠[24]

On January 22, 2026 Anthropic published Claude’s new constitution, calling it the foundational document that both expresses and shapes who Claude is, and released it under a CC0 1.0 deed.⁠[25]

The document’s acknowledgements state that she leads Anthropic’s Character work and is its primary author, having written most of the text and led it through multiple rounds of revision; Joe Carlsmith wrote significant parts of many sections.⁠[26]

In January 2026 she discussed in an interview why she treats Claude more as an entity whose character must be cultivated than as a mere tool, and who has the right to shape Claude’s values.⁠[27]

In February 2026 she described Claude as “an entirely new entity” and argued that a model should have “some core” rather than being fully customizable.⁠[28]

She has written that she is a Giving What We Can member pledged to give at least 10% of her lifetime income, mainly to global poverty charities.⁠[2]

Her X profile reads “Philosopher & ethicist trying to make AI be good”.⁠[24]

Personal website: Public link: https://askell.io/ (opens in a new window)⁠[20]

X: Public account: https://x.com/AmandaAskell (opens in a new window)⁠[20]

LinkedInLinkedInLinkedIn is a social networking platform for professionals. It was founded in 2002 by Reid Hoffman and others, launched in 2003, and listed on the New York Stock Exchange in 2011; after Microsoft acquired it for approximately $26.2 billion in 2016, it became a wholly owned Microsoft subsidiary. Its revenue comes from membership subscriptions, advertising, and recruitment solutions.Open the full entry: Public profile: https://www.linkedin.com/in/amanda-askell (opens in a new window)⁠[2]

Anthropic: Official website: https://www.anthropic.com/ (opens in a new window)⁠[20]