Google Moves AI Responsibility Team From DeepMind to Lobbying Division

Google is moving its team responsible for evaluating AI model safety from the DeepMind lab to the Global Affairs division, which handles lobbying and regulator relations. The change, announced to employees in an email in late August, takes effect on September 1 and has raised concerns among some team members about the independence of their research.

What exactly is changing

The group, known internally as the AI responsibility unit, had until now operated within Google DeepMind, the research lab responsible for developing Gemini models. Its job is to test models for the most serious risks, including the potential to help build chemical, biological, radiological, or nuclear weapons, as well as to analyze patterns in how users interact with chatbots and the psychological effects of those interactions.

After the change, the team falls under Global Affairs, the Google and Alphabet division led by Kent Walker, which oversees content policy, government and regulatory relations, and legal and compliance matters. This is the same division responsible for the company's lobbying of regulators worldwide, including on matters concerning AI regulation.

Employee concerns

Some team members worry that the new reporting structure will limit their research independence and make it harder to access the teams developing Gemini models inside DeepMind. Several employees tried, before September 1, to transfer to research groups that remain within the lab's structure - but those transfer requests were denied.

Helen King, the Google DeepMind vice president who leads the team, sought to reassure employees in internal correspondence, saying research priorities would not change and that publishing findings externally would continue to be supported. She also said the team would retain access to DeepMind's computing infrastructure and its current staffing levels.

Publishing research externally is still supported and important - Helen King, Google DeepMind vice president

Many teams across Google work on AI safety and responsibility, and by bringing our AI responsibility teams closer together, we're strengthening their ability to inform the safety of our models and products - Google spokesperson

Broader reorganization context

The team's move is part of a deeper overhaul of DeepMind that has been underway since August 2026. That's when co-founder and then-CEO Demis Hassabis moved to the role of chairman, while day-to-day management of the lab passed to Koray Kavukcuoglu, DeepMind's former chief technology officer and Alphabet's chief AI architect. Google's longtime chief scientist, Jeff Dean, also departed after 27 years at the company.

The reorganization is meant to bring DeepMind closer to Google's standard divisional structure, similar to the one used for Search, after more than a decade of operating as a largely autonomous unit since Google's 2014 acquisition of the lab. The changes coincided with reports of delays in developing subsequent versions of the Gemini models and growing competitive pressure from OpenAI and Anthropic.

Why it matters

Moving the team that assesses the most serious AI-related risks closer to the company's lobbying structures raises questions about whether model safety evaluations will remain independent of Google's business and regulatory interests. That question carries added weight given that Hassabis publicly called in July 2026 for the creation of an independent, US-led body to test advanced AI models for national security risks before they reach the market.

For companies and institutions tracking the regulatory policies of major AI firms, it's a sign that the line between safety research and political engagement with regulators is blurring at the biggest AI labs. Google defends the decision, arguing that bringing its responsibility teams closer together will strengthen, not weaken, oversight of model safety.

The matter directly affects the Gemini models used by millions of users and businesses worldwide, including in Poland, where Google is expanding access to its AI tools through business and educational services. The independence of risk assessments for these models has practical implications for trust in their commercial and public use.