Joe Carlsmith

Also known as Joseph Carlsmith

Joe Carlsmith is a philosopher and AI safety researcher whose work examines power-seeking, deceptive alignment and Claude’s constitution. He earned a BA in philosophy at Yale, studied philosophy at NYU from 2016 to 2018, and completed a doctorate in philosophy at Oxford by 2023. At Open Philanthropy, he studied risks to humanity’s long-term future and published reports on brain-scale computation, power-seeking and scheming AI. His 2024 essay series examined how agents with different values should relate and the ethics of power; a Dwarkesh Podcast conversation covered AI takeover concerns and artificial minds. He joined Anthropic in November 2025 to help design Claude’s constitution and served as a visiting lecturer at Yale Law School during spring term of 2026; as of October 2026, he works on the constitution, where Anthropic credits him with writing significant parts of many sections and revising the fall 2025 draft.

Now at: Anthropic

Timeline

2016–2023

Yale, NYU and Oxford

After earning a BA in philosophy at Yale, he studies philosophy at NYU from 2016 to 2018. He completes a doctorate in philosophy at Oxford by 2023.
2018–2025

Research at Open Philanthropy

He works as a Senior Research Analyst at Open Philanthropy from 2018 to 2025, studying risks to humanity’s long-term future and whether AI systems may converge on power-seeking behavior.
2020

Estimating brain-scale compute

In 2020, he publishes How much computational power does it take to match the human brain?, an Open Philanthropy report on the computing power sufficient to match the brain’s task performance. It considers four ways to estimate that capacity.
2022

Power-seeking AI risk

In 2022, he publishes Is Power-Seeking AI an Existential Risk?, which examines whether misaligned agents could seek power and cause existential catastrophe.
2023

Research on scheming AI

In 2023, he publishes Scheming AIs: Will AIs fake alignment during training in order to get power? It examines whether advanced AI systems might perform well in training to gain power later.
2024

Otherness and control in AGI

In 2024, he completes Otherness and control in the age of AGI, an essay series about how agents with different values should relate and the ethics of seeking and sharing power. In August 2024, he discusses AI takeover concerns and the treatment of artificial minds with Dwarkesh Patel.
March 2026

Visiting lecturer at Yale Law

During spring term in 2026, he serves as a visiting lecturer at Yale Law School. In March 2026, he discusses Claude’s model constitution and legal and philosophical questions about its design at a Yale Law School colloquium.
November 2025–October 2026

Claude constitution at Anthropic

In November 2025, he joins Anthropic to help design Claude’s character, constitution and specification. As of October 2026, he works on the constitution; Anthropic credits him with writing significant parts of many sections and serving as the main point person for revising the fall 2025 draft.