Google Merges DeepMind AI Safety Team into Global Affairs Division
Nashnova编辑部
Google is folding DeepMind's roughly 90-person AI responsibility team into its global affairs division, merging it with lobbying and public policy — raising internal concerns that the team's independence in flagging model risks will be compromised.
What exactly is being reorganized?
DeepMind's ~90-person AI responsibility team will move into Google's global affairs division; the shift is expected to complete by early September.
This means → the team that stress-tests AI models for safety risks will now sit inside the unit that handles lobbying and public policy.
In plain terms = the people who look for dangers in Google's AI are being placed under the same roof as the people who manage its government relationships.
What changes at the top of DeepMind?
Co-founder and CEO Demis Hassabis will step down as CEO and take a newly created chairman role, focusing on external partnerships and broader AI issues.
Koray Kavukcuoglu will take over as DeepMind's head, with the title senior vice president — not CEO.
This means → DeepMind no longer carries a "CEO" title; its org chart moves one step closer to Google headquarters — structurally, its independence narrows.
What has management promised?
DeepMind VP Helen King wrote in an internal email that the team's day-to-day work and mission will not change; it will keep access to DeepMind's data and resources.
She also pledged that computing capacity and headcount will not be cut, and that Google will continue backing frontier responsible-AI research.
This reflects management's awareness of internal pushback — an attempt to reassure the team with a "business as usual" message.
Why are insiders worried?
Multiple former team members have publicly voiced concern that independent research will be harder to pursue outside DeepMind.
The team previously ran several sub-groups testing AI models for risks in chemistry, biology, radiological, and nuclear domains, and studying the psychological effects of human-AI interaction and chatbots.
In plain terms = these researchers give AI a "safety exam" — checking whether it could help someone synthesize hazardous materials or harm users psychologically. Once inside a policy division, the independence of that exam is the central question.
Why does this matter beyond Google?
A Google spokesperson said the company has multiple AI safety teams and that closer collaboration is strengthening its ability to keep models and products safe.
But the core tension is unanswered: when a safety-evaluation team and a lobbying unit share the same reporting line, will assessment conclusions bend to commercial timelines?
This reflects a tension across the entire AI industry — safety research demands independence, but corporate structures naturally pull toward consolidation and efficiency.
市场有风险，内容仅供研究参考，不构成投资建议。