The AI Conundrum: Why We Should Pursue the Promise—and Insist on the Safeguards

An original commentary drawing on Ben Cohen’s “Why AI’s Deepest Mind Believes the World Is About to Change Radically,” published in WSJ. Magazine/The Wall Street Journal, October 21, 2026, as supplied in the accompanying text. Photography credited to Jonas Unger.

Artificial intelligence deserves neither unconditional faith nor reflexive rejection. It deserves something more demanding: stewardship. The strongest argument for advancing AI is that it can help humanity solve problems that have resisted generations of effort. The strongest argument for restraining it is that unprecedented capability does not automatically produce wisdom, accountability, or broadly shared prosperity.

Ben Cohen’s portrait of Demis Hassabis brings that tension into focus. His subject is not simply racing to build a more profitable digital assistant. He is pursuing a lifelong ambition to use artificial intelligence to expand scientific understanding. Cohen presents a scientist whose ultimate interest lies less in intelligence itself than in what intelligence might discover.

That ambition warrants serious support. But supporting it does not require accepting every prediction, endorsing every deployment, or allowing the companies building the technology to determine its limits alone.

The Case for Optimism: The Promise Is No Longer Entirely Hypothetical

The most persuasive defense of Hassabis’s vision is not his confidence about the future. It is the scientific achievement already behind him.

In 2024, Hassabis and John Jumper shared half the Nobel Prize in Chemistry for protein structure prediction; David Baker received the other half for computational protein design. The Nobel announcement described how AlphaFold2 enabled predictions of the structures of virtually all 200 million proteins researchers had identified.

That accomplishment changes the terms of the AI debate. It demonstrates that AI can do more than generate fluent prose or automate administrative tasks: it can help resolve a scientific problem with consequences for research into life and disease.

As Cohen’s supplied article describes Hassabis’s governing ambition: “Step one, solve intelligence. Step two, solve everything else.”

The formulation is sweeping, but the underlying logic is compelling. Better scientific tools can expand what researchers are able to investigate. If AI helps scientists identify useful patterns, evaluate hypotheses, or understand biological structures more effectively, its value need not depend on achieving human-level general intelligence.

Indeed, one of the strongest arguments for continued investment is that we need not settle the debate over artificial general intelligence to recognize the usefulness of systems already advancing scientific work. The International AI Safety Report likewise recognizes substantial potential benefits in medicine, education, research, and prosperity.

There is therefore a moral dimension to progress as well as precaution. Delaying a technology that could reduce suffering is not automatically a neutral choice. The relevant question is whether particular advances can be pursued with proportionate safeguards—not whether every uncertainty must disappear before research continues.

The Case for Caution: A Breakthrough Is Not a Guarantee

The skeptical response is equally important: predicting a protein’s structure is not the same as curing a disease.

The Nobel recognition establishes a major advance in protein structure prediction, not proof that AI has solved medicine. Cohen’s supplied account itself acknowledges the difficult distance between designing potential drugs and turning them into tested medicines.

That distinction should govern the broader argument. A successful specialized system does not establish that general intelligence is imminent. Nor does it establish that a more powerful system will be reliable, controllable, or socially beneficial.

The supplied article reports Hassabis’s expectation that AGI could arrive within two to three years. That should be understood as his forecast—not a settled scientific timetable. The International AI Safety Report emphasizes substantial uncertainty about future capabilities and the range of possible outcomes.

Meanwhile, concern need not depend on a speculative extinction scenario. Deepfakes, scams, and biased outputs already cause harm. More capable systems also raise questions about cyberattacks, biological misuse, labor-market disruption, and loss of control, although the evidence and uncertainty differ across those risks.

The serious counterargument to technological optimism is therefore not that progress is inherently undesirable. It is that capability can advance faster than the institutions responsible for managing its consequences.

The Governance Dilemma: Who Gets to Decide What Is Safe?

One of the most consequential tensions in Cohen’s supplied article concerns regulation. Hassabis warns that badly designed rules could constrain a technology policymakers do not yet understand.

That concern has merit as an argument for adaptable regulation. It is less persuasive as an argument for waiting.

Uncertainty cuts both ways. It can make rigid rules premature, but it can also make irreversible deployment premature. The absence of perfect knowledge should not automatically favor the party seeking to release a more powerful system.

The industry-funded testing body described in the supplied article offers one possible bridge between technical expertise and public oversight. Yet its credibility would depend on its design. Who sets the standards? Who appoints the evaluators? Can the body require meaningful changes or prevent release? What happens when commercial urgency conflicts with its findings?

A sound approach should distinguish research from deployment and lower-risk applications from systems capable of consequential harm. It should also make safety claims testable rather than merely reassuring.

These are proposed principles, not guarantees:

• Require independent evaluation proportionate to a system’s capabilities and intended uses.
• Establish clear responsibility for failures and meaningful reporting of serious incidents.
• Strengthen safeguards as evidence of dangerous capabilities increases.
• Protect legitimate research without treating commercial release as an entitlement.
• Assess who receives the benefits and who bears the costs.

The point is not to regulate ambition out of existence. It is to ensure that ambition operates within institutions worthy of public trust.

Supporting the Vision Without Surrendering Judgment

Hassabis’s scientific record makes his optimism worth taking seriously. It does not make scrutiny unnecessary.

The strongest version of his thesis is not that intelligence will automatically solve everything. It is that better intelligence can expand humanity’s problem-solving capacity—and that we should deliberately direct that capacity toward worthwhile ends.

But discovery and distribution are different problems. A scientific advance does not, by itself, determine whether its benefits will be affordable, accessible, or broadly shared. Those outcomes require choices about institutions, incentives, and public responsibility.

Cohen’s supplied article captures the right aspiration when Hassabis defines winning as bringing AGI into the world safely and securely for everyone’s benefit. The obligation is to turn that aspiration into conditions that can be evaluated and enforced.

We should support the search for AI that helps defeat disease and deepen human understanding. We should be equally determined that the pursuit remain accountable to the people whose lives it could transform.

The AI conundrum is not a choice between courage and caution. It is whether we can exercise both at once.