When Nobel laureate Demis Hassabis sat down with science communicator Derek Muller recently, the conversation started where it usually does: celebrating AlphaFold's spectacular success. The numbers are indeed staggering—200 million protein structures predicted in a few years, compared to 150,000 painstakingly discovered over five decades. It's the kind of breakthrough that makes for great headlines and inspiring documentary footage of a world map lighting up as researchers globally access these structures.
But strip away the celebratory tone, and what emerges is a far more complex—and uncomfortable—conversation about the future of artificial intelligence. One that Hassabis himself seems acutely aware of, even if he doesn't have all the answers.
The AlphaFold Story: Innovation Built on Foundations
Let's be clear about what AlphaFold represents. It's a genuine scientific breakthrough, no question. The ability to predict protein structures at scale has already accelerated drug discovery, environmental research, and our fundamental understanding of biology. Over 2.5 million researchers from nearly every country have used it. That's transformative.
But here's what often gets lost in the triumphalism: AlphaFold couldn't exist without those first 150,000 structures that structural biologists spent half a century discovering. As Hassabis himself acknowledges, "we couldn't have done it without the first 150,000." The system learned from decades of painstaking human work, then scaled that knowledge through machine learning techniques.
This isn't just academic acknowledgment—it's a fundamental pattern we need to understand about AI development. These systems don't create knowledge from nothing. They consolidate, pattern-match, and extend existing human expertise at unprecedented scale. The question that should concern us: who gets to control the systems that do this consolidation?
The Open Science Paradox
Hassabis presents DeepMind's decision to openly release AlphaFold's predictions as obviously beneficial: "We knew we could only think of a tiny fraction of what the entire scientific community might do with it." And he's right about the benefits. Open science accelerates progress. Distributed innovation works.
But then comes the catch he's also candid about: "At the same time you want to restrict access to that same technology to would-be bad actors whether that's individuals or even rogue nations."
This is the core paradox of powerful AI systems. Open them up, and you enable beneficial applications but also potential misuse. Lock them down, and you concentrate power in the hands of whoever controls the technology—likely large corporations or state actors with the resources to develop these systems.
"It's very hard balance to get right," Hassabis admits. "There's no one's yet got a good answer for how you do both of those things."
At least he's honest about it. But honesty doesn't solve the problem.
The Democratization That Worries Everyone
When Muller pressed Hassabis on the recent emergence of more accessible, "thrifty" AI models from companies like DeepSeek and Alibaba, the concern became explicit. Initially, developing cutting-edge AI seemed to require resources comparable to the Manhattan Project—state-level backing or massive corporate infrastructure. That concentration of resources was, in a perverse way, a safety feature.
But that's changing. "It's sort of available to everyone and it is worrying," Hassabis concedes. More people accessing these technologies means more potential for innovation from unexpected places—"kids like I was back when I was tinkering around with theme park can now work on some really interesting AI systems."
It also means the barrier to misuse keeps dropping.
Hassabis floats the idea of creating economic incentive structures that favor responsible actors—having "the players that have the right intentions, you know, backed by government society" become more successful and powerful. But think about what that actually means. It means creating systems where governments and markets decide which AI development is "responsible" and deserves support.
Who defines "right intentions"? What happens when those definitions shift? And most fundamentally: aren't we just recreating the same concentration of power we're supposedly worried about, just with a different justification?
The Race to the Bottom No One Can Stop
Perhaps the most revealing moment in the conversation comes when Hassabis discusses "race dynamics"—the competitive pressure that drives companies to cut corners, move faster, prioritize deployment over safety.
"Even if all the actors are good in that environment, let alone if you have some bad actors, that can drag everyone to rush too quickly, to cut corners," he explains. "For any individual actor it sort of makes sense but as an aggregate it doesn't."
This is the tragedy of the commons playing out in real-time with the most powerful technology humans have ever developed. And Hassabis knows it. He's been advocating for international cooperation, celebrating summits at Bletchley Park and in Paris. But cooperation requires trust, shared values, and the willingness to sacrifice competitive advantage for collective safety.
Look around at the current geopolitical environment. Now ask yourself: how realistic is that?
Hassabis himself admits he'd "much rather there be a calm CERN-like effort towards AGI, these final few steps, but given the geopolitical framework we're in, maybe that's not possible."
Translation: We know what we should do. We probably won't do it. So now we need to "be more pragmatic."
The Questions We Should Be Asking
When Muller asks about the future and what to tell his four kids about school and preparation, Hassabis offers reassuring but generic advice: embrace new technologies, learn math and computer science, "learn to learn" and adapt quickly.
But the question beneath the question—what kind of world are we building, and who gets to decide?—remains largely unanswered.
Here's what we should be interrogating:
On concentration of power: When Hassabis says AI needs "more voices and stakeholders" beyond "100 square miles of California," but DeepMind is owned by Google/Alphabet, what does that actually mean in practice? How much independence does DeepMind really have from one of the world's largest corporations?
On democratic oversight: International summits are good. But who has a seat at those tables? How do the billions of people whose lives will be transformed by these technologies participate in decisions about their development and deployment?
On the "new renaissance": Hassabis envisions a "golden age" of scientific breakthroughs in the next decade—curing diseases, solving energy crises, addressing climate change. That would be extraordinary. But who owns the AI systems that make these breakthroughs? Who profits? And what happens to the researchers, institutions, and communities whose knowledge these systems learned from?
On safety: If the people building these systems admit they don't have good answers for fundamental safety and control questions, why is development accelerating rather than slowing down?
What This Means
I don't doubt Hassabis's sincerity when he talks about responsibility and the need for careful development. His concerns about race dynamics and misuse appear genuine. The decision to open-source AlphaFold's predictions has created enormous scientific value.
But sincerity and good intentions aren't governance structures. They're not accountability mechanisms. And they certainly aren't sufficient guardrails for technology that Hassabis himself describes as "one of the most important things ever invented."
The uncomfortable truth is that we're conducting a massive, irreversible experiment with AI development, and the people running the experiment keep telling us they don't have all the answers about how to keep it safe or fair.
Maybe it's time to ask whether "move fast and figure it out later" is really the approach we want to take with artificial general intelligence.
References
- Source interview: Demis Hassabis and Veritasium's Derek Muller discuss AI, AlphaFold and human intelligence
- Original URL: https://www.youtube.com/watch?v=Fe2adi-OWV0
- Published: May 21, 2025
- AlphaFold Protein Structure Database: https://alphafold.ebi.ac.uk/
- DeepMind's AI for Science initiatives
- International AI Safety Summit proceedings
About the Author
Léa Rousseau is a tech reporter for Digiall's official blog, specializing in artificial intelligence ethics, technology governance, and the social impact of innovation. She brings a critical perspective to technology coverage, questioning corporate narratives and examining power structures in the tech industry.
What do you think?
Have thoughts on AI governance and democratization? We want to hear your perspective. Follow Digiall for more critical analysis of technology trends shaping our future.
#ArtificialIntelligence #AlphaFold #AIEthics #TechnologyGovernance #DeepMind #OpenScience
Beyond the Nobel: What Demis Hassabis Won't Tell You About AI's Democratization Problem