Ethan Perez

Research

Several of Ethan Perez's papers concern finding and measuring failures in language models. Perez was first author of "Red Teaming Language Models with Language Models" (2022), which uses one model to generate test cases that surface harmful behavior in another, and of "Discovering Language Model Behaviors with Model-Written Evaluations" (2022), which automates the writing of behavioral evaluations with language models.³⁶

Perez is the last-listed author of "Inverse Scaling: When Bigger Isn't Better" (2023), which presents tasks where larger models perform worse, and a co-author of "Sleeper Agents" (2024), which studies whether deceptive behaviors trained into models persist through safety training. Earlier, Perez co-authored the 2020 retrieval-augmented generation (RAG) paper.⁷⁸⁴

Safety work at Anthropic

Perez is the last-listed author of the January 2025 Anthropic paper introducing Constitutional Classifiers, safeguards trained on synthetic data generated from a natural-language constitution to defend against universal jailbreaks. In August 2025 Perez co-authored Anthropic's posts on filtering harmful chemical, biological, radiological and nuclear content from pretraining data and on a pilot exercise in which Anthropic and OpenAI ran alignment evaluations on each other's public models.²⁹⁵

Perez is listed with an Anthropic affiliation on the August 2026 Alignment Science post introducing the Conceptual Reasoning Index, a set of benchmarks for the kinds of argument used in philosophy and AI futurism, written with researchers at Redwood Research.¹

Work and contributions

• Constitutional Classifiers Published research

  Last-listed author of the 2025 Anthropic paper introducing Constitutional Classifiers.²

• Red Teaming Language Models with Language Models Published research

  First author of the 2022 paper, which automatically generates test cases with one language model to find cases where a target model behaves harmfully, instead of relying only on hand-written tests.³

• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks Published research

  Co-author of the 2020 paper that combines a pre-trained language model with a retrievable non-parametric memory for knowledge-intensive tasks.⁴

Timeline

1. May 22, 2020Co-authored the retrieval-augmented generation (RAG) paper for knowledge-intensive NLP tasks.⁴
2. Feb 7, 2022First author of "Red Teaming Language Models with Language Models".³
3. Dec 19, 2022First author of "Discovering Language Model Behaviors with Model-Written Evaluations".⁶
4. Jun 15, 2023Last-listed author of "Inverse Scaling: When Bigger Isn't Better".⁷
5. Jan 10, 2024Co-authored "Sleeper Agents", on deceptive behaviors that persist through safety training.⁸
6. Jan 31, 2025Last-listed author of the Anthropic paper introducing Constitutional Classifiers.²
7. Aug 27, 2025Co-authored findings from a pilot Anthropic and OpenAI alignment evaluation exercise.⁵
8. Aug 12, 2026Co-authored the Alignment Science post introducing the Conceptual Reasoning Index.¹

Where they work

• Anthropic 34 people in the wiki

Related people

• Nicholas CarliniAlso at Anthropic
• Samuel R. BowmanAlso at Anthropic
• Collin BurnsAlso at Anthropic
• Evan HubingerAlso at Anthropic
• Samuel MarksAlso at Anthropic
• Ian R. McKenzieSafety and alignment · Evaluation and benchmarks
• Margaret MitchellSafety and alignment · Evaluation and benchmarks
• Yulia TsvetkovLanguage models · Safety and alignment
• Aylin CaliskanSafety and alignment · Evaluation and benchmarks
• Su Lin BlodgettEvaluation and benchmarks · Safety and alignment