The internal email went out on a Monday morning with little ceremony: the team that studies the risks and social impact of artificial intelligence would leave DeepMind, the laboratory that builds Google’s frontier models, and report to a new home. Starting next week, the roughly 90-person unit will sit inside Google’s global affairs division, the department that manages the company’s lobbying, public policy, and communications, according to two current DeepMind employees.
The move rearranges one of the most closely watched reporting lines in AI research. For years, the responsibility team worked alongside the researchers who developed Gemini and the models that preceded it, part of a laboratory founded on the principle that safety research should live close to capability work. The change shifts the team out of the lab entirely and into the corporate apparatus that manages Google’s standing with regulators and the public.
The reorganization comes weeks after Demis Hassabis, DeepMind’s co-founder and chief executive, stepped down and moved into a newly created chairman role. His departure from the top job had already prompted questions about how the laboratory’s leadership would change. The current employees said the responsibility team’s move was part of the broader realignment under the new structure, and that some team members worry the change will weaken the work.
The concerns are practical as well as symbolic, the employees said. Teams that report into policy and communications divisions can face pressure to frame findings in ways that serve the company’s interests, they noted, and researchers who study harms, from model bias to the societal effects of automation, need independence to publish candid results. Researchers at other companies have made the same argument for years, and several prominent safety researchers have left leading laboratories after their teams were dissolved or moved away from the core research effort.
Google’s own history with such teams is uneven. The company dissolved its AI ethics council in 2019 after an internal outcry, and it has repeatedly reorganized the groups that study AI’s social implications. The current unit, which grew out of those earlier efforts, has focused on risk assessment, red-teaming, and the study of how models behave in the real world, according to the employees. Its reports have fed both internal safety reviews and Google’s public statements on responsible AI.
The change also reflects the shifting politics of AI safety inside large technology companies. With regulators in the United States, Europe, and Asia pressing for rules on foundation models, the function that translates research findings into policy positions has grown more important and more visible. Moving the team into global affairs could be read as Google consolidating that translation function under one roof, or as the company pulling safety research closer to its messaging operation. The employees who spoke with this newsroom said the distinction matters, and that the team’s ability to publish independently will be the test.
The laboratory’s public posture has been unchanged since the announcement, the employees said. DeepMind told the team that the move would not affect access to models and data, and that researchers would continue to work with the laboratory’s engineering groups on evaluations and audits. Whether that holds in practice will become clear in the coming months, as the team settles into its new reporting line and its first rounds of research travel through the new chain of command.
The unit’s history explains why the reporting line matters to its members. It traces its roots to the AI principles Google drew up in 2018, after a staff revolt over the company’s work on a military drone-analysis program. The principles pledged that Google would not build AI for weapons or surveillance, and the company created the responsibility team to hold it to those promises, the employees said. For seven years the group has operated inside the laboratory, publishing risk assessments and pushing product teams to test models before release. Moving that function into the policy division raises a question the employees said they have already heard from colleagues: who watches the watchers when the watchers report to the people whose job is persuasion?
The stakes are rising as governments write the rules the team studies. The European Union’s AI Act requires risk assessments for foundation models, and regulators in Washington and Beijing have pressed companies to document how they test frontier systems. A team that can point to independent evaluation work gives a company credibility in those conversations, which is why its placement is being watched. Google’s competitors have kept their oversight functions inside their research organizations, and several advertise those lines publicly. The 90 people moving next week will be the visible test of whether Google’s version of that commitment survives the move.
For the field, the shuffle is another data point in a two-year pattern. Across the industry, laboratories have moved safety teams around as competition has intensified, sometimes to keep safety researchers close to product groups, sometimes to consolidate oversight functions into policy arms. Anthropic, built around a safety mission, has expanded its responsible-scaling work even as it pushed models to market quickly. OpenAI has reorganized its safety functions several times since 2023. Each reshuffle draws the same question: whether the researchers keep the independence that made their warnings credible.
In Google’s case, the answer will show up in the research itself. If the team’s next evaluations appear in the open with the same candor as before, the reporting line will matter less than the behavior it enables. If the work goes quiet, or begins to sound like policy positioning, the move will look like what some employees fear it is. Either way, the 90 researchers who moved desks this month are now operating under a structure that has rarely been tested with a team of this size and this close to the models it studies.
