Develop classifiers, data pipelines, automated evaluations, and cross-context monitoring systems using model activations and chains-of-thought. The Safety Oversight team at Google DeepMind uses large-scale production traffic to detect model misbehavior and user misuse. This new team within the GenAI safety organization ensures real-world safety and alignment of deployed Generative AI models.

DeepMind★5.0 (2)· London

Builds classifiers, data pipelines, and automated evaluation methods using parallelised data pipelines, statistical modeling, and model activations/chains-of-thought. The team monitors deployed generative AI models' safety and alignment by analyzing large-scale production traffic and developing cross-context systems to detect coordinated harms and misuse.

Build classifiers and data pipelines to detect model misbehavior and misuse, and develop cross-context monitoring systems for coordinated harms using Gemini and GenMedia models. You will research novel monitoring methods (e.g., model activations, chain-of-thought) and collaborate with infrastructure teams. The Safety Oversight team monitors the safety and alignment of deployed AI models using large-scale production traffic and automated evaluations.

You will build classifiers, data pipelines, and automated evaluation systems to monitor deployed generative AI and LLM models for safety and alignment issues. As part of Google DeepMind's Safety Oversight team, you will detect production safety failures, model misuse, and coordinated harms at scale.