Google Moves Its ~90-Person AI Responsibility Team Out of DeepMind

Just weeks after the leadership overhaul that saw Demis Hassabis step aside as Google DeepMind CEO, another shoe has dropped — this time landing squarely on the team charged with keeping Google’s AI honest.

According to a Wall Street Journal exclusive published late on August 26, Google is moving its roughly 90-person AI responsibility unit out of Google DeepMind and into the company’s global affairs organization. The shift relocates the researchers and policy specialists who evaluate the risks and societal impact of Google’s most advanced models out of the lab that builds those models, and places them under the corporate umbrella that handles government relations and public policy.

The move has already prompted some members of the team to depart, the Journal reported, and researchers both inside and outside the unit have raised concerns about what the new reporting structure means for their independence — and for their ability to detect threats emerging from Google’s newest frontier models.

What Actually Changed

The details matter here, because this is not a case of a team being disbanded. The unit — led by Helen King, DeepMind’s Senior Director of Responsibility — survives intact, with the same headcount of around 90 people and, according to the Journal’s reporting, the same stated mission. King has reportedly sought to reassure employees that the changes won’t affect the team’s day-to-day work, and that its focus will remain on assessing the risks and societal impacts of AI systems.

What changes is who they answer to. Previously embedded inside Google DeepMind, the team now sits within Google’s global affairs organization — the corner of the company run by President of Global Affairs Kent Walker, which oversees policy, legal, regulatory, and government engagement. That org has chaired the company’s AI Responsibility Council since Google published its AI Principles in 2018, so the new alignment is not without precedent. But it does mean the people evaluating frontier-model risks no longer report through the same chain of command as the people building frontier models — and no longer sit inside the lab where the raw research access lives.

For a company that has spent years telling regulators its safety work is woven directly into model development, the optics are delicate: the evaluation function moves closer to the policy and communications apparatus, while DeepMind itself becomes a more purely product-and-research engine.

The Latest Chapter in a Brutal Month for DeepMind

The relocation cannot be separated from the broader unraveling at Google DeepMind over the past three weeks. The timeline is worth revisiting, because the responsibility-team move is best understood as the third act of a reorganization that began at the top:

• August 5: Sundar Pichai announced that Demis Hassabis would transition out of the Google DeepMind CEO role, becoming chairman of the lab and Alphabet’s chief scientist. Chief Technology Officer Koray Kavukcuoglu — DeepMind’s Chief AI Architect — stepped up as SVP of Google DeepMind, taking over day-to-day operations and responsibility for all AI model development.
• Same week: Jeff Dean, Google’s chief scientist and one of its most storied engineers, announced his departure after 27 years, alongside Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, to found a startup called Discovery Loop.
• August 12: The Journal reported that Hassabis had pitched the creation of an independent AI oversight entity before the shake-up was announced — a detail that reads very differently in hindsight.
• August 26: The AI responsibility unit is moved out of the lab entirely.

Add Fortune’s reporting on stalled model efforts, missed deadlines, staff burnout, and an internal revolt over defense work, and the picture is of an organization being aggressively re-centered from a research-first lab into a product-delivery engine — with Silicon Valley accountability structures replacing London-lab autonomy. Reuters reported that Hassabis and Kavukcuoglu told staff several nontechnical teams were being moved out of DeepMind and into the corporate reporting structure. The responsibility team is by far the most consequential of those.

Why Independence Is the Whole Ballgame

The core tension in AI safety governance is structural: the people best positioned to find problems with a frontier model work for the organization with the strongest incentive not to find them. Every credible framework for responsible AI — from internal review boards to the frontier-safety commitments signed by major labs — rests on giving evaluators genuine independence from delivery pressure.

Embedding the responsibility team inside DeepMind was one way to buy proximity: researchers sitting next to the developers, with access to internal systems, capable of raising flags before a release rather than after. Moving the unit into global affairs buys a different asset — distance from product pressure, plus a direct line into the policy conversations that increasingly determine what Google can ship and where. Both models have theoretical merit. The concern researchers raised to the Journal is that the transition risks delivering neither: losing the deep access that comes from being inside the lab, without gaining the hard independence of a truly external reviewer.

It is also impossible to ignore the sequence. An oversight body proposed, then a shake-up, then the oversight function relocated — each step individually defensible, collectively suggestive of a safety apparatus being steadily softened as Google races OpenAI and Anthropic on shipping velocity. DeepMind lost its CEO and chief scientist in the same week; the responsibility team is now the third major shoe to drop.

What Google Would Say

In fairness, the company’s position is coherent. Google published its AI Principles in 2018 — earlier than most peers — and has released annual responsible-AI progress reports since, with its February 2026 edition focused specifically on agentic and frontier risks. Kent Walker’s organization has run the AI Responsibility Council from the beginning. Arguably, consolidating responsibility work under the same roof as regulatory engagement reflects where AI governance actually happens now: in Brussels, Washington, and state capitols, not just in model cards.

And the alternative reading of the Hassabis oversight-body pitch cuts the other way too — Google’s chief scientist wanted an external watchdog with teeth, which is more ambitious than what most labs have proposed.

What to Watch

Three signals will tell us whether this is a governance upgrade or a quiet demotion. First, headcount: does the 90-person unit hold its size and budget through the transition, or does it thin out as departing members aren’t backfilled? Second, access: does the team retain pre-deployment evaluation rights over Gemini-family frontier models, or does its review become retrospective? Third, output: do the annual responsibility reports keep publishing unflattering findings at the 2026 cadence and depth?

Google is making a bet that responsibility work scales better as a corporate function than a lab function. The next year of Gemini releases will show whether the bet pays off — and whether the researchers who stayed can still catch what the ones who left saw coming.

Sources are listed below, auto-generated from this post’s frontmatter.