Amelia Glaese
Amelia Glaese is an OpenAI vice president whose research addresses language-model safety and human feedback. In 2015, she coauthored a Technical University of Munich paper on reconfigurable automation hardware. At DeepMind, she coauthored 2022 research on language-model red-teaming and Sparrow, an information-seeking dialogue agent trained with human feedback. She led human-data work for Gemini in 2023 and coauthored OpenAI research on deliberative alignment in 2024 and PaperBench in 2025. As of October 2026, she oversees OpenAI safety work as a vice president.
Timeline
-
2015
Automation research at TUM
In 2015, she coauthors a peer-reviewed conference contribution at the Technical University of Munich on reconfigurable field-bus hardware for flexible automation systems.
-
2022
Red-teaming language models
In 2022, she coauthors DeepMind research that uses language models to generate inputs that elicit harmful text from other language models. The method offers one way to find harmful behavior before users are affected.
-
2022
The Sparrow dialogue agent
In 2022, she coauthors the DeepMind paper introducing Sparrow, an information-seeking dialogue agent trained with human feedback to be more helpful, correct, and harmless than prompted language-model baselines.
-
2023
Gemini human-data lead
In 2023, she leads Gemini's human-data work for Google's multimodal model family.
-
2024
Deliberative Alignment
In 2024, she coauthors Deliberative Alignment, which teaches models to recall and reason over written safety specifications before answering. The paper says the approach aligns OpenAI's o-series models.
-
2025
Evaluating research agents
In 2025, she coauthors PaperBench, a benchmark that asks AI agents to reproduce 20 ICML 2024 Spotlight and Oral papers from scratch to test their ability to replicate research.
-
As of October 2026
OpenAI safety leadership
As of October 2026, she is an OpenAI vice president overseeing safety work. In 2026, she describes Astra's ability to find previously unknown security flaws and develop ways to exploit them.
Sources and credits
Photo: We have no published photo of Amelia Glaese yet. Send us a free-licensed photo.
- Enabling flexible automation system hardware: Dynamic reconfiguration of a real-time capable field-bus
- Red Teaming Language Models with Language Models
- Improving alignment of dialogue agents via targeted human judgements
- Gemini: A Family of Highly Capable Multimodal Models
- Deliberative Alignment: Reasoning Enables Safer Language Models
- PaperBench: Evaluating AI's Ability to Replicate AI Research
- OpenAI says its next model finds security flaws nobody has found yet