A Policy Debate That Has Forgotten to Ask the Right Question
The public debate over artificial intelligence has been shaped almost entirely by fear. What happens to workers when automation eliminates their jobs? What happens to democracy when AI enables mass surveillance or precision disinformation? What happens to humanity when systems grow powerful enough to slip beyond human control? These are serious questions and they deserve serious policy attention. But New York Times opinion columnist Ezra Klein argues that the relentless focus on harm has crowded out an equally urgent question that almost nobody is asking: what should AI do for us?
The absence of a positive public agenda is not merely an oversight. Klein's argument is structural: the benefits of AI do not emerge automatically from the technology itself. They have to be deliberately designed, funded, and directed. Left entirely to market forces, AI will be pointed at problems that generate the greatest private returns — enterprise software, advertising optimization, financial modeling. The problems that matter most to the public — rare diseases, climate science, government accessibility, scientific data — will remain underserved unless someone makes a conscious choice to redirect the technology toward them.
"We know what we fear AI will do to us. But what do we hope it will do for us?"
When AI Is Pointed at the Right Problem, the Results Are Remarkable
Klein grounds his argument in a string of recent scientific achievements that demonstrate what becomes possible when AI is given a well-defined problem and the right underlying data. An OpenAI model recently disproved a mathematical conjecture that had resisted solution for 80 years. DeepMind's AlphaFold model produced a quantum leap in predicting protein structures, earning its designers a share of the 2024 Nobel Prize in Chemistry. A drug for pulmonary fibrosis just became the first fully AI-generated treatment to demonstrate proven efficacy and safety in human trials. A team at the Mayo Clinic developed an AI system capable of detecting pancreatic cancer on a CT scan up to three years before a clinician could see it. DeepMind's Graphcast weather model appears to produce more accurate forecasts faster than the supercomputer systems currently in use.
These are not incremental improvements. They represent categorical leaps in what is scientifically possible. And they share a common feature: each was enabled not just by AI, but by the combination of AI and rich underlying data sets that took years or decades to build. AlphaFold, for instance, was only possible because the Protein Data Bank — a laboriously assembled repository of protein structures that the National Science Foundation began funding in the 1970s — existed in the first place. Without it, there was no AlphaFold. The implication is uncomfortable: the most transformative scientific achievements of the AI era may depend on public investments made generations ago, in data that no private company would have had the incentive to collect.
When the National Science Foundation began funding the Protein Data Bank in the 1970s, nobody described the project as AI infrastructure. It was simply basic science — a shared repository of molecular structures, built slowly over decades by researchers depositing their findings into a common archive. Half a century later, that archive became the training foundation for AlphaFold, which transformed structural biology overnight. The lesson Klein draws: the data sets that will unlock the next generation of AI breakthroughs probably don't exist yet, and the private sector has no incentive to build them. That is precisely the kind of investment only public funding can make.
A Private-Public Divide Is Already Opening Up
Before AI can be used to solve public problems, it must first be accessible to the institutions that work on them. Klein argues that access is already becoming the central fault line. Large corporations can afford to purchase colossal quantities of computing power — what the industry calls "compute" — while public universities, government agencies, and non-profit research institutions increasingly cannot. Goldman Sachs, he observes, can deploy AI at a scale that a state university system cannot approach. This gap is not a future risk; it is a present condition, and it is widening.
One proposal Klein raises is the creation of a genuine public option for AI: a frontier-level model under direct public control, paired with subsidized compute access for universities and public agencies. The goal would not be to replace private AI development but to ensure that institutions without Goldman Sachs's budget can still bring serious computational resources to bear on hard public problems. Without some version of this, the benefits of the current AI revolution will flow almost entirely to those who can already afford them — compounding existing inequalities in research capacity and government capability rather than reducing them.
Using Government Contracts to Bend AI Toward Public Problems
Much of AI capacity will remain in the private sector regardless of any public investment, and Klein argues that a serious public agenda needs to account for this. His model is Operation Warp Speed — the federal program that used advance purchase commitments and guaranteed market contracts to direct private pharmaceutical capacity toward COVID-19 vaccines. The government did not build the vaccines itself; it defined the outcome it wanted and promised to buy the result at scale if it was delivered. The same logic could apply to AI: define the problems — a treatment for a rare disease, a new battery electrolyte, a materials discovery — guarantee a purchase if the solution is found and distributed equitably, and let private companies compete to solve it.
The recent collaboration between Microsoft and the U.S. Department of Energy's Pacific Northwest National Laboratory offers a working proof of concept. The project used AI to analyze more than 32 million materials and identify a promising electrolyte candidate that could improve lithium-ion battery storage. That kind of systematic search through enormous chemical space would have taken human researchers decades. But Klein's point is that projects like this remain sideshows — curiosities pursued at the margins while the AI industry's main energy goes toward corporate clients with guaranteed revenue streams. Only public funding, or publicly structured market incentives, has the power to redirect that energy toward problems the market will not otherwise solve.
"Only public funding has the possibility to bend the industry toward public problems."
AI as a Digital Concierge to Government
Beyond scientific research, Klein sees AI as a potential transformative tool for the relationship between citizens and their government. The U.S. tax code is notoriously complex, and the gap between what a wealthy individual gets from a personal accountant and what an ordinary citizen navigates alone represents a profound inequality in access to civic life. An AI system trained on IRS data and current tax law could, in principle, give every person the equivalent of that personal accountant — walking them through their return, flagging every benefit they qualify for, and eliminating the anxiety of a process that currently costs Americans billions of dollars in fees to professional preparers each year.
More ambitiously, the same underlying technology could function as a comprehensive entry point to everything the government offers: benefits, services, programs, agencies. A citizen who qualifies for housing assistance, child care subsidies, or veterans' benefits often never receives them simply because navigating the bureaucracy is too complex or too time-consuming. An AI-powered digital concierge that could understand a person's situation and surface every relevant program represents the kind of civic infrastructure that markets will never build because there is no revenue model in helping poor people find their benefits. That is exactly why Klein argues it requires a public mandate.
Even the best AI models cannot solve problems for which the underlying data does not exist or is not organized in a usable form. Government data sets are frequently described by researchers as being in a state of near-total disarray — inconsistent formats, missing records, decades of accumulated technical debt. Klein points to Alberta, Canada, where a provincial government project found that AI dramatically accelerated the work of cleaning and reorganizing government data — but only after a deliberate decision was made to invest in doing so. Creating the data infrastructure that AI needs to work on public problems is not glamorous, but it is prerequisite. Without it, even well-funded, well-intentioned AI deployments will founder on the basic reality that the underlying information is a mess.
A Question of Intent, Not Just Risk Management
Klein's essay is ultimately a challenge to the framing that has dominated AI policy thinking. Risk management — preventing harms, slowing dangerous deployments, building guardrails — is necessary, but it is not sufficient as a public agenda. A society that succeeds only in preventing the worst outcomes of AI while failing to direct the technology toward its greatest potential benefits has still largely squandered the opportunity. The question is not only what AI might do to us but what we want it to do for us — and that requires a level of intentionality, investment, and democratic deliberation that the current policy conversation has almost entirely avoided.
The technical capability to address rare diseases, to make government services navigable for every citizen regardless of education or income, to accelerate the basic science underlying climate solutions — all of this exists in some form today. What does not yet exist is the institutional will to point AI at these problems with the same focus and resources that private industry brings to enterprise software and consumer advertising. Klein's argument is that building that institutional will is, or should be, one of the central policy challenges of this technological moment. It will not happen by default. It requires asking, clearly and repeatedly, what it is that the public actually hopes AI will do.