Research Scientist, Multimodal Alignment, Safety, and Fairness

About the role

This role owns the end-to-end research lifecycle for multimodal alignment, safety, and fairness within DeepMind's Frontier AI unit, driving novel methods that shape how AI systems understand and interact with society. You will own the design and execution of interdisciplinary sociotechnical experiments that probe multimodal model behaviors under real-world conditions. The position requires you to own the development of new interfaces, tools, and evaluation paradigms that translate alignment theory into measurable system properties. You will own authorship of high-impact publications and external communications that define the state of the art in multimodal safety. You will own deep collaboration with engineers, product teams, and social scientists to ensure research translates into robust behaviors in deployed products. This role owns the responsibility for mentoring junior researchers and shaping the research agenda for multimodal oversight. You will own proactive scanning of the AI alignment landscape to identify emerging risks and opportunities for foundational model improvement.

Key facts

What you'll do

Investigate and generate novel research directions at the intersection of multimodal perception, alignment objectives, and societal impact assessment.

Design, implement, and rigorously evaluate new methods, interfaces, and tools for probing and steering evolving AI behaviors in multimodal settings.

Execute large-scale experiments that characterize how multimodal models handle subjective and creative tasks across diverse social contexts.

Collaborate with machine learning engineers, product teams, and social scientists to translate theoretical alignment concepts into testable system behaviors.

Communicate research insights through clear written reports, technical documentation, and high-impact presentations to both internal and external audiences.

Drive technical projects from initial hypothesis through to deployment-ready insights, maintaining rigorous scientific standards throughout the lifecycle.

Develop new paradigms for human-AI rating that account for systemic behaviors, adapt to human feedback, and proactively seek context-dependent nuances.

Analyze real-world usage data to assess adherence to desired behaviors and identify emergent risks in multimodal model outputs.

Partner across the model development and deployment cycle to ensure alignment research remains grounded in practical constraints and user needs.

Contribute to Google DeepMind's mission by advancing the frontier of multimodal AI safety and enabling breakthroughs in models such as Gemini and Nano Banana.

Requirements

You have a PhD or equivalent experience in AI, machine learning, human-computer interaction, or a related quantitative field with demonstrated research excellence.

You possess deep expertise in AI alignment, safety, and fairness, with a strong awareness of the technical and sociotechnical challenges in deploying multimodal systems.

You have published work at top-tier conferences and journals in machine learning, AI safety, human-computer interaction, or related fields.

You have hands-on experience building or evaluating multimodal AI systems, including vision-language models or similar architectures.

You have demonstrated ability to design and execute experiments that capture complex human-AI interactions in realistic settings.

You have strong programming skills in Python and experience with modern machine learning frameworks such as PyTorch or JAX.

You have excellent written and verbal communication skills, with a track record of explaining complex technical concepts to diverse audiences.

You are comfortable working in a fast-paced, ambiguous research environment and making decisions with incomplete information.

Nice to have

Prior experience with multimodal datasets, evaluation frameworks, or tooling for interpretability and oversight.

Experience with large language models or generative architectures and their alignment challenges.

Background in social science, ethnography, or participatory methods that can enrich sociotechnical modeling.

Experience with open-source AI safety tools or contributions to reproducibility and robustness research.

Practical notes

This is a full-time position based in Kirkland, Washington, US, with potential assignments in Mountain View, California, US, and New York City, New York, US.

Relocation support may be available for Kirkland-based roles.

This role requires authorization to work in the United States and sponsorship may be provided for roles in locations where legal requirements allow.

Applications will be reviewed on a rolling basis until the role is filled.