Google DeepMind's science chief: AI will democratize knowledge, not replace human judgment

Pushmeet Kohli sits at the intersection of two worlds that rarely meet. By day, he oversees science at Google DeepMind, the research unit that produced AlphaFold, the system that predicted protein structures and helped earn the 2024 Nobel Prize in Chemistry for his colleagues Demis Hassabis and John Jumper. By evening, he goes home and uses the same artificial intelligence tools he builds to explain airplane engines to his eleven-year-old son, asking the system to translate complexity into language a child can grasp. TIME magazine named him one of the hundred most influential people in AI globally. Yet unlike many tech executives who keep screens away from their children, Kohli embraces the technology he creates, testing its boundaries and its possibilities in real time.

The question driving much of his thinking is deceptively simple: what happens when AI becomes genuinely useful to science? His answer is measured but ambitious. He believes AI will amplify what individual researchers can accomplish—not by replacing them, but by handling the middle portion of the scientific process. A human scientist might read a hundred research papers in a given period. An AI system can read hundreds of thousands, cross-reference insights across disciplines, and generate hypotheses that would have required teams of researchers working for months. But the bookends remain stubbornly human. Someone must decide what problem is worth solving—whether that's defeating a disease, understanding a new virus, or developing crops resistant to bacterial infection. And someone must validate the solution in the real world, running experiments, conducting trials, gathering evidence. The acceleration happens in the search itself.

When pressed on timelines, Kohli acknowledges the gap between laboratory breakthrough and human benefit. AlphaFold emerged years ago, yet new drugs derived from its insights remain in development. The bottleneck is not discovery anymore. It is the long, necessary process of clinical trials, regulatory approval, and real-world testing. Candidate drugs exist where AI and AlphaFold proved essential, he notes, but the machinery of pharmaceutical development grinds slowly by design. Much of the visible progress comes from academic research—new malaria vaccines, treatments for neglected tropical diseases—where the stakes are high and the resources lean.

But Kohli's ambitions extend beyond medicine. He speaks with genuine conviction about nuclear fusion, a technology that has lived in a strange temporal limbo for decades, always promised to arrive within twenty years, never quite arriving. The old joke, he says, is that fusion will always be twenty years away. He believes AI will finally break that curse. The path is concrete: AI is accelerating the discovery of superconducting magnets needed to control tokamak reactors, the same way it is speeding development of better batteries and more efficient solar cells. These are not theoretical gains. Hurricane prediction models built with AI are already issuing earlier warnings than conventional systems.

On the question of displacement—the fear that entire professions might vanish—Kohli draws a historical parallel. The calculator did not eliminate arithmetic; it changed what arithmetic meant. The real challenge is educational. As AI handles the mechanical work of problem-solving, the premium shifts to problem-finding: the ability to specify what matters, to frame a question with precision and depth, to understand what will actually move the needle for people and society. He would tell his own son, if asked what to study, to focus on that skill. Not narrow specialization in one corner of biology or chemistry, but the capacity to identify the right problem and know how to validate the answer.

On the question of whether we are overestimating or underestimating AI, Kohli refuses the binary. Both are happening simultaneously. These systems move with stunning speed and accomplish remarkable things, yet they carry equally stunning weaknesses. They can be brilliant at one task and absurdly incompetent at another. The real danger is not the technology itself but the gap between what people think it can do and what it actually does. Someone brilliant with language might be helpless with spatial reasoning. An AI might excel at pattern recognition across scientific literature and fail catastrophically at a task a child could solve. The only remedy is interaction, experience, learning where the intelligence lives and where the stupidity hides. Part of Kohli's daily work is building guardrails, testing systems relentlessly, asking what could go wrong. He describes it in homely terms: like a grandmother watching her grandchildren, asking what they plan to do, correcting them when they stray. It is not a solved problem. It is a process of massive testing and evaluation, repeated endlessly, with no guarantee of perfection. But it is the work that must happen before these tools are released into the world.