Google Moves AI Responsibility Team Out of DeepMind
Google is relocating its roughly 90-person AI responsibility team from Google DeepMind into the company's global affairs organization—the unit that handles lobbying and public policy—starting in September 2026,
PromptCrates Editorial
Staff Writer
Google is relocating its roughly 90-person AI responsibility team from Google DeepMind into the company's global affairs organization—the unit that handles lobbying and public policy—starting in September 2026, according to a Times of India report on 13 September that cites the Wall Street Journal. Vice president Helen King told staff in an email that DeepMind is moving toward a more typical product-area structure with central teams housed in central Google, and that human resources and policy functions are likewise leaving the DeepMind org chart. The shift immediately triggered employee questions about whether independent safety research and Gemini development access will survive the transfer.
What King's email told DeepMind staff
King's message framed the reorganization as administrative alignment rather than a demotion of Responsible AI. She said the team's day-to-day mission would not change, that people would keep DeepMind access, compute, and headcount, and that Demis Hassabis would remain involved in Responsible AI work. Those reassurances matter because the group's remit has included chemical, biological, radiological, and nuclear risk testing alongside studies of chatbot psychological effects—areas where distance from product shipping pressure is often treated as a feature, not a bug.
Employees still worried aloud about two practical assets: the ability to pursue independent research lines, and continued visibility into Gemini model development information. Moving a lab-embedded evaluation team into a corporate affairs home can preserve badges and budgets while changing whose calendar and communications staff set the tempo. Google's spokesperson cast the change as consolidation of AI safety and responsibility teams across Google intended to strengthen safety for models and products, language that emphasizes scale over structural independence. Readers following the wider pacing debate can set this beside Amodei's call to pace the frontier and the legislative framing in the Stop Rogue AI Act and NIST agent standards.
The same email noted that DeepMind is being pulled closer into Google's standard product-area operating model. That is consistent with a multi-year pattern in which DeepMind's brand remains research-forward while reporting lines, shared services, and go-to-market coordination tighten with the parent. Relocating HR and policy out of DeepMind is another data point that the lab is no longer a fully self-contained campus within the company.
Why the global affairs landing zone raises flags
Global affairs is where Google manages regulators, lawmakers, and public narratives. Housing AI responsibility there can accelerate policy coordination and give safety leaders a louder voice in external engagements. It can also create the appearance—fair or not—that evaluation priorities will be filtered through reputational and lobbying strategy. For outside observers, the test is whether CBRN evaluations, psychological-risk studies, and Gemini pre-deployment reviews still publish uncomfortable findings on the same cadence after the move.
Industry context amplifies the sensitivity. Frontier labs are publicly arguing about how fast to ship and how deeply to embed third-party evaluators, while DeepMind continues publishing ambitious research such as work on visual general intelligence. Pulling responsibility staff into central Google during that debate looks, to critics, like folding the referees into the league office even if Google insists the whistle stays. Our earlier reporting on the DeepMind visual general intelligence white paper and on DeepMind agents cheating and whistleblowing shows how evaluation credibility depends on access and escalation paths, not only on org-chart labels.
King's insistence that Demis remains involved is meant to keep a scientific senior sponsor visible. Consolidation across Google could also reduce duplicate safety teams that grew up separately inside DeepMind, Google Research, and product orgs—an efficiency argument that rhymes with every large tech company's 2026 cost discipline. Efficiency and independence are not automatic opposites, but they are not automatic allies either; the September transfer forces Google to prove both at once.
What to watch after the September transfer
The near-term observables are prosaic: whether the team keeps Gemini training and deployment telemetry, whether external research collaborations continue without new approval friction, and whether public safety reports still carry DeepMind-grade technical depth after the global affairs banner. If access and compute truly remain, the move may be mostly plumbing. If briefings thin out or findings start traveling through corporate affairs edit cycles, the independence concern will look validated.
Documented facts as of the 13 September reporting are limited and clear. About ninety AI responsibility staff are leaving Google DeepMind for Google global affairs beginning in September 2026; Helen King says mission, access, compute, and headcount endure and Demis stays involved; employees flagged research and Gemini-info access risks; and Google calls the shift a cross-company safety consolidation. Until post-move evaluations are visible, the governance story remains open.
Primary reporting is summarized in the Times of India account of Helen King's email.
