Google shifts DeepMind's AI-responsibility team to global affairs
Google is folding DeepMind's AI-safety work into its corporate policy arm, the clearest sign yet that the lab's responsible-AI function is being absorbed into the company's wider machine. It's the kind of reorganization safety researchers have long worried about — and now they're saying so publicly.
Google is moving DeepMind's roughly 90-person "AI responsibility" team — the group focused on the risks and societal impact of AI — into the company's global affairs unit, according to an internal email reported by the Wall Street Journal. The Journal said researchers inside the team have raised concerns about how the change affects their independence and their ability to do the work. Google has not commented publicly on the move, which follows the broader DeepMind reshuffle earlier this month when Demis Hassabis stepped back from day-to-day leadership.
Why it matters: relocating safety oversight out of the research lab and into global affairs is a structural statement, not a cosmetic one. Global affairs sits closer to regulation, lobbying and corporate positioning than to the research floor, and the people whose job is to flag risk now report into that orbit. That is precisely the dynamic that safety teams have historically warned about — independence erodes fastest when the function that raises hard questions gets folded into the parts of the company responsible for managing how the outside world sees it. The concern from researchers that their independence is at risk is the tell: the substance is being kept, but its institutional footing is changing. If Google wants the credibility that comes from having a real safeguard, moving the team out of the lab is a strange way to signal it.
Apple researchers published Luce, a new approach to single-image-to-3D generation that unifies geometry and physically based rendering (PBR) materials — albedo, metallic-roughness and surface normals — in a single voxelized Gaussian cloud. On the Toys4K benchmark it claims state-of-the-art results, improving the FID by 28 percent over the strongest baseline, and on a new benchmark of AI-generated images it raises the CLIP image-alignment score to 0.8519 versus 0.8299. The model is designed to drop into standard rendering pipelines, producing assets that preserve fine details like text, logos and inscriptions.
The practical payoff is that a generated 3D object can now be relit and composited like a real one, which is the difference between a research demo and something a game, film or e-commerce team can actually use. Apple keeps quietly advancing generative-3D research while the attention is on larger labs, and Luce is another step toward making generated assets production-ready rather than just sharable.
What to watch: which lab's governance structure breaks first under the pressure of an AGI timeline — and whether any of them fight to keep safety independent from the corporate machine.
When safety research gets folded into public affairs, does it lose the independence that made it worth having in the first place? Tell us in the comments.
Sources: Techmeme · MSN / Wall Street Journal · Apple Machine Learning Research · arXiv
