Amanda Askell

Amanda Askell's work connects questions about values to the behavior of AI assistants. Her 2018 philosophy doctorate at New York University examined how to compare worlds containing infinitely many people. The dissertation, Pareto Principles in Infinite Ethics, starts with a familiar intuition: making some people better off without making anyone worse off should improve a world. Askell shows how that principle, combined with other assumptions, leaves many infinite worlds impossible to rank. The work concerns the limits of ethical comparison rather than a recipe for choosing a single correct set of values.[3]

Askell studied philosophy at the University of Dundee and Oxford before her doctorate at NYU. Her personal biography describes earlier research at OpenAI on its policy team, including AI safety through debate and human baselines for evaluating AI performance. At Anthropic, her team works on fine-tuning methods intended to improve honesty and other character traits as models become more capable. This connects philosophical questions about judgment to training methods that can change how an assistant responds in practice.[6]

Training helpful assistants

Askell is the first-listed author of A General Language Assistant as a Laboratory for Alignment, published in 2021. The team studied how prompting and training could make general-purpose assistants more helpful, honest, and harmless. Their experiments compared imitation learning with models trained to distinguish or rank responses. Ranked preferences performed better than imitation in the reported experiments and often improved more favorably with model size. This work made assistant behavior an empirical research subject, with interventions and evaluations that could be compared rather than only described as aspirations.[2]

Constitutional AI

She also co-authored the 2022 Constitutional AI paper. Its training process asks a model to criticize and revise responses using written principles, then uses AI judgments to train a preference model. Reinforcement learning uses that preference model as a reward signal. The experiment reduced reliance on human labels of harmful answers and aimed for assistants that could explain objections without simply evading questions.[5]

Claude's character

At Anthropic, Askell leads Character work and is the primary author of Claude's constitution. The document addresses how Claude should act when helpfulness, honesty, safety, and competing instructions come into conflict. Anthropic says Askell wrote most of the text and credits Joe Carlsmith and other collaborators with substantial contributions. Written largely for the model itself, the constitution informs training and states the behavior Anthropic wants to produce. It is useful for understanding the intended reasoning behind Claude's answers, while the company acknowledges that actual behavior can fall short of those intentions.[1]

Before Claude, Askell studied the incentives facing organizations that develop AI. In a 2019 paper with Miles Brundage and Gillian Hadfield, she argued that competition can push companies to spend too little on safety and security. The paper treats responsible development as a collective-action problem. A company may want safer practices while fearing that competitors will gain an advantage by avoiding their cost. Its proposed research agenda asks how cooperation could make responsible behavior easier to sustain across the industry.[4]