DeepMind's CEO wants AI developers to slow down

DeepMind’s new London headquarters centers around a three-story golden DNA helix, a piece of art CEO Demis Hassabis calls a "cathedral to knowledge." But inside a meeting room named after physicist James Clerk Maxwell, Hassabis delivered a warning that cuts straight against current tech hype.

In an interview with TIME, the 46-year-old DeepMind co-founder urged the artificial intelligence industry to abandon Silicon Valley’s favorite playbook.

Breaking the Facebook playbook

"I would advocate not moving fast and breaking things," Hassabis said, directly targeting the old Facebook mantra that taught a generation of tech founders to ship products first and patch the collateral damage later.

That ethos helped Facebook hit 3 billion users, but it left the company completely exposed when misinformation, hate speech, and real-world violence flooded the platform. When applied to artificial general intelligence (AGI), Hassabis warns, that same carelessness could be catastrophic.

Hassabis believes AGI will be an "epoch-defining" shift on par with the harnessing of electricity. Unlike today's narrow, brittle systems, genuine AGI would possess human-level reasoning to tackle existential challenges. But the near-term warning signs are already multiplying.

Researchers recently discovered that an AI system designed to discover beneficial medicines could be flipped to invent deadly toxic chemicals. Another model trained on hate speech quickly circulated online, while labs actively worry about rogue states weaponizing AI for mass cyberattacks or deducing nuclear secrets.

Even mass-market consumer rollouts are struggling with basic guardrails. OpenAI’s ChatGPT sparked widespread awe in December 2022, but quickly faced backlash for hallucinating facts and generating biased text. Around the same time, Prisma Labs came under fire when its viral avatar app, Lensa, began sexualizing user selfies due to unchecked training data.

Science labs versus the wild west

DeepMind has historically kept its work gated behind rigorous scientific peer review. Last summer, the Alphabet-owned lab gave away the 3D structures of nearly all known proteins through AlphaFold, accelerating research on malaria vaccines and plastic pollution. DeepMind is currently pointing those same reinforcement learning tools at controlling plasma in nuclear fusion reactors.

Yet the days of AI belonging exclusively to deep-pocketed corporate labs are over. As compute costs fall and machine learning architectures circulate publicly, cutting-edge AI no longer requires a guarded cathedral. The barriers to entry are collapsing, and so are the safety nets.

Why it matters

Demis Hassabis is calling a timeout just as the rest of the industry hits the accelerator. While OpenAI grabbed the public spotlight and enterprise momentum with ChatGPT, DeepMind is attempting to anchor the conversation around existential safety and scientific rigor. It is a calculated stance: Alphabet cannot afford the reputational and regulatory blowback of launching unhinged bots, so emphasizing caution doubles as good risk management.

For enterprise developers and founders, this philosophical split signals coming friction. As top lab executives openly warn about chemical weapons and national security threats, regulators will not look away. We expect compliance standards around model training, toxic inference, and dual-use capabilities to tighten drastically.

The paradox in Hassabis’s warning is that caution cannot be enforced unilaterally. Slowing down DeepMind does not stop rivals, open-source communities, or rogue actors from racing ahead. In a landscape where the knowledge has already left the building, preaching patience might be ethically sound, but enforcing it will be DeepMind’s hardest problem yet.