Google Parks DeepMind Safety Team Under Lobbying Arm

Google is moving its roughly 90-person AI responsibility team out of Google DeepMind and into the company’s Global Affairs division by early September, according to an internal email reviewed by The Wall Street Journal. The unit that tests frontier models for chemical, biological, radiological and nuclear hazards will now report through the same arm that runs lobbying and public policy.

The timing lands while Gemini has crossed one billion monthly active users and Google Cloud just posted an 82 percent revenue jump. Researchers inside the team already fear the change will distance them from the engineers shipping the next Gemini versions.

The move is administrative on paper. In practice it rewires how findings on the heaviest model risks travel from the lab floor to the groups that decide what ships and what gets said about it in public.

The 90-Person Unit Heads to Global Affairs

The AI responsibility group evaluates the heaviest risks that accompany advanced models. Its work covers three main tracks that sit at the heart of how Google says it keeps Gemini in bounds.

• Testing models against chemical, biological, radiological and nuclear (CBRN) hazards
• Studying how people actually interact with chatbots day to day
• Examining the psychological effects those systems can have on users

Those tracks are not interchangeable. CBRN testing probes whether a frontier model can help a bad actor with weapons-related knowledge. Interaction studies watch real user behavior for failure modes that lab benchmarks miss. Psychological work asks whether prolonged chatbot use changes mood, dependency, or judgment. Together they form the hands-on layer of Google’s stated safety process.

Helen King, the Google DeepMind vice president who leads the team, told staff in the email that the core research mission, computing resources, headcount and the ability to publish externally would stay intact. A Google spokesperson added that pulling safety groups closer together across the company would strengthen how findings shape models and products.

Still, the destination is Global Affairs, the corporate home for policy work and government relations. Other support functions such as HR and policy teams are making the same move out of DeepMind as the lab is folded into a more standard product-area structure.

King’s assurances address resources and publication rights. They do not rewrite the org chart. Once the team sits under Global Affairs, its managers answer through a chain built for lobbying and external messaging rather than for pure research iteration. That is the structural change staff are weighing against the promises of unchanged headcount and compute.

Leadership Changes That Made Room for the Shift

The relocation sits inside a wider reshuffle that began earlier in August.

1. Early August 2026: Demis Hassabis steps back from day-to-day CEO duties at Google DeepMind to become chairman of the lab and chief scientist at Alphabet, focusing on external engagements.
2. Same period: Koray Kavukcuoglu takes operational control as senior vice president overseeing Gemini development.
3. Late August: Barret Zoph, a former OpenAI research executive and Thinking Machines Lab co-founder who earlier spent six years at Google Brain, joins as vice president of research to bring reinforcement-learning and post-training expertise.

Google has been pulling DeepMind, acquired in 2014 and long run with relative autonomy, into tighter corporate alignment. The lab’s London research culture is giving way to the product-area model used by Search and other core divisions. Managers have privately acknowledged the risk that some talent will walk.

Hassabis moving to chairman and chief scientist roles frees operational bandwidth for Kavukcuoglu to run Gemini as a product line. Zoph’s arrival adds post-training depth at the same moment safety evaluation is being reassigned. The sequence reads as a single redesign: research leadership refreshed for speed, support functions standardized, and the responsibility unit placed where policy and government relations already live.

DeepMind’s long autonomy made sense when the lab was primarily a research bet. With Gemini past the billion-user threshold and Cloud revenue surging, the company is treating the lab more like Search or Cloud. That product-area template has less room for a semi-independent safety shop reporting only inside the lab.

Researchers Worry About Distance From Gemini Builders

Two current DeepMind employees told the Journal that some members of the responsibility team fear the new reporting line will blunt their independence and reduce day-to-day insight into the newest models. Several applied to transfer into research groups that are remaining inside DeepMind. Those requests were turned down.

The team’s focus on emerging and frontier topics in responsible AI would remain unchanged and publishing research externally continues to be supported and important.

Helen King, Vice President, Google DeepMind, internal email

King also wrote that access to DeepMind itself would be preserved and that Hassabis still cares about responsible AI and plans to stay involved. On X, reaction was swift and mostly skeptical: users repeatedly noted that the same people who probe bioweapon risks and chatbot psychology will now sit next to the lobbying operation. That crowd reading matches the internal fretting that technical safety work could slowly bend toward public-affairs priorities.

The denied transfers are concrete. Staff who wanted to stay beside the model builders were told no. That outcome feeds the worry that early, informal access to Gemini checkpoints will shrink even if formal compute allocations hold. Safety findings that once moved through hallway conversations and shared lab channels may now travel as scheduled reports into a policy organization.

Public reaction on X focused on the optics of CBRN and psychology researchers sharing a division with lobbyists. Internal concern is more mechanical: whether the team still sees models early enough to shape them, or only late enough to describe them.

What Stays Inside DeepMind and What Leaves

Not every safety function is leaving. Specialized technical teams focused on model behavior, privacy and security remain inside the lab. The Responsibility and Safety Council, long co-chaired by King and the lab’s COO, continues to review high-impact work against Google’s AI Principles. The lab still maintains its Frontier Safety Framework protocols for identifying critical capability levels and preparing mitigations before models ship.

The broader DeepMind responsibility and safety approach keeps an AGI Safety Council led by co-founder Shane Legg focused on extreme long-term risks. Those structures stay. The 90-person responsibility unit that does the hands-on CBRN and psychological testing is the piece moving under Global Affairs.

That split is the practical irony. The people closest to the models keep some safety work; the group charged with the most visible societal-risk evaluations reports through the policy and lobbying channel.

The retained councils and frameworks still set principles and review high-impact work. The departing unit is the one that runs the tests those reviews often rely on. Keeping governance inside the lab while relocating the testing bench creates a handoff. How cleanly findings cross that handoff will decide whether the split is tidy or leaky.

Commercial Payoff Meets the Reorg

Alphabet’s AI bets are delivering hard numbers at the same moment the org chart is being redrawn.

Sundar Pichai told investors that AI infrastructure and solutions powered the Cloud surge and that nearly 90 percent of the Fortune 100 now use Gemini Enterprise. The official remarks confirm Cloud revenue grew 82 percent while the consumer Gemini app closed in on the billion-user mark. Those results create pressure to ship faster. They also raise the cost of any slowdown if safety researchers leave or if external regulators question whether oversight remains sufficiently independent.

An 82 percent Cloud jump and a consumer product crossing one billion users are not abstract context. They are the reason the product-area model is winning inside DeepMind. Faster shipping cycles reward tighter coupling between research, product, and policy. They also make any friction from safety review more expensive in the short run.

Token throughput rising from 16 billion to 22 billion per minute in a single quarter shows the same acceleration on the developer side. The responsibility team’s CBRN and psychology work scales against that volume. Placing the team under Global Affairs may speed the translation of findings into external commitments. It may also lengthen the path those findings take before they reach the engineers adding the next capability.

Global Affairs Pulls Safety Toward Policy Work

Global Affairs already houses lobbying and public policy. Adding the 90-person responsibility unit means the same division will now own both the external narrative on AI risk and the internal group that measures CBRN exposure, chatbot behavior, and psychological effects.

King and the spokesperson framed the shift as bringing safety groups closer together so findings shape models and products more effectively. The mechanism is consolidation: fewer separate safety silos, one path into the corporate functions that brief lawmakers and draft public positions.

Consolidation has a tradeoff visible in the existing facts. Technical teams on model behavior, privacy, and security stay inside DeepMind beside the builders. The unit that produces the most externally legible risk evaluations moves beside the people who explain Google to governments. Independence, in the researchers’ telling, was partly a function of proximity to Gemini development. Global Affairs proximity optimizes for a different audience.

• Computing resources and headcount: unchanged per King’s email
• Right to publish externally: preserved
• Day-to-day access to newest model versions: the open concern
• Reporting line: now through policy and government relations

Hassabis, now focused on external engagements as chairman and chief scientist, is described as still caring about responsible AI and planning to stay involved. His new brief already points outward. The responsibility team’s new home points the same direction. Whether that alignment helps safety findings travel farther, or simply reframes them for policy first, is the question the reorg leaves unresolved.

The September Deadline Tests Early Access

The transition is set to finish in early September. By then the reporting line will be live, the denied internal transfers will be settled fact, and the team will be working from inside Global Affairs while Gemini development continues under Kavukcuoglu.

Preserved compute and headcount mean the unit can still run CBRN evaluations and interaction studies at prior scale. Preserved publishing rights mean external papers can still appear. The missing guarantee is the informal one: seeing frontier checkpoints early, while mitigations are still cheap to design.

The Responsibility and Safety Council and the Frontier Safety Framework remain inside DeepMind as review layers. They will still need test results. If those results now arrive on a Global Affairs calendar rather than a lab calendar, review cadence could shift even when principles stay the same.

Commercial momentum makes the test sharper. Cloud at $24.8 billion for the quarter, Gemini past a billion users, and API tokens at 22 billion per minute all argue for speed. Safety work that slows a release becomes more visible on that backdrop. Safety work that arrives too late to change a release becomes invisible. The September cutover decides which pattern the 90-person team inherits.

Managers Already Flag Talent Risk

Internal discussions have already surfaced the obvious downside. Managers conceded the corporate shifts carry a real risk of top talent moving to rival labs. The denied transfer requests inside the responsibility team are an early signal that some staff prefer to stay closer to the model builders. High-profile exits earlier this year, including Noam Shazeer to OpenAI and John Jumper to Anthropic, already rattled investors. Adding Barret Zoph is one attempt to restock research leadership.

Google’s own recent product tweaks show the company is sensitive to user-facing AI friction. The decision to give people a clear Google Health AI coach off-switch after backlash over clutter is a small example of reacting when responsibility issues surface publicly. Scaling that responsiveness while the core risk-evaluation team sits under Global Affairs is the test ahead.

Shazeer and Jumper left for competitors before this reorg. Managers now flag the same outflow risk across the broader DeepMind alignment push. The responsibility team’s blocked transfers show the pressure is already local: people tried to move closer to research and were refused. Retention will depend on whether King’s guarantees on mission, compute, and publishing feel sufficient once the Global Affairs line is real.

The transition is set to finish in early September. Computing access and headcount stay the same on paper. Whether the day-to-day independence and early insight into Gemini development survive the new reporting line is the open question the researchers themselves have already put on the table.