So far, the policy conversation about A.I. has focused on preventing harms to the public. What if A.I. leads to mass job loss? What if it’s used for surveillance? For bioweapons? What if it concentrates wealth and power in the hands of the few while acting as an agent of cognitive and social atrophy for the many? What if future systems slip out of our control or beyond our understanding?
These dangers are real, or at least plausible. Policies to prevent or respond to them are essential. But there is another policy conversation that is strangely absent: How does A.I. benefit the public?
The public goods of A.I. are, I recognize, not the most politically popular topic at the moment. A.I.’s polling is dismal, even if its usage is exploding. Data centers are hideously unpopular and, in some jurisdictions, banned. Pope Leo XIV devoted his first encyclical to the technology’s dangers. Many see the use of A.I. at all as borderline immoral — it adds demand to an overhyped, dangerous technology whose rollout needs to be slowed or perhaps even stopped.
I start from a different perspective. A.I. is here. It will be used. But how it is used, for what and by whom are meaningful questions. A.I.’s benefits will not emerge automatically or inevitably. It will take work to identify the problems A.I. can help the public solve and then create the data, financing and “compute” — the raw processing power that A.I. systems use when queried, supplied by data centers full of specialized chips — needed to actually solve them.
When A.I. is pointed at the right problem, in the right way, the results can be remarkable. An OpenAI model just disproved a conjecture that had eluded mathematicians for 80 years. A drug for pulmonary fibrosis just became the first fully A.I.-generated treatment with proven efficacy and safety in humans. A Mayo Clinic team developed an A.I. system that can detect pancreatic cancers on a CT scan up to three years before clinicians can see them. DeepMind’s Graphcast model appears to generate weather predictions both faster and more accurately than the supercomputer systems that are used now. The 2024 Nobel Prize in Chemistry went in part to the designers of the AlphaFold model for the quantum leap it offered in predicting protein structures.
The corporations buying A.I. access for themselves are not deluded. But they are finding that you cannot solve problems just by throwing A.I. at them. It takes work to structure a problem in a way that allows A.I. to be useful, just as it took work, in previous generations, to integrate I.T. into a company or redesign a factory to take advantage of electricity.
The AlphaFold advance was possible only because of the Protein Data Bank, a laboriously created database of protein structures that the National Science Foundation began funding in the 1970s. No Protein Data Bank, no AlphaFold.
So a public agenda for A.I. needs to be more than a vague intention to toss A.I. at public problems. It starts with access, but it does not end there.
Still, start with access. It might begin with an actual public option for A.I.: Perhaps there should be at least one frontier-level model under direct public control. It would at least begin with the recognition that there is already more demand for compute than there is supply. That might mean buying compute for public purposes and making certain it is affordable for universities, public agencies and others.
The digital divide I mentioned above is already opening up into a private-public divide, where Goldman Sachs can afford colossal quantities of compute and public universities cannot. I want the public to have access to enough compute that it can unleash A.I. on hard problems.
But much of the A.I. capacity will remain in the private sector. So a public agenda for A.I. should also give the private sector reason to work on public problems. Like in Operation Warp Speed, the government could define the outcomes it wants — a drug, a solution — and guarantee a market if it’s found and distributed equitably.
What kinds of public problems can A.I. solve? Look closely at the math advance OpenAI just released, and what you’ll see is that the A.I. system was able to combine two of A.I.’s strengths to achieve the result: One was a vast knowledge of existing mathematics, such that it could apply a different branch of the discipline to the problem. The other was the ability to attempt possibilities that would have been too laborious for human mathematicians to try.
There are a lot of problems that could benefit from the sustained attention of tireless researchers with a vast knowledge of all relevant subject areas. Rare diseases, for instance, where the suffering is immense but the market for a cure is often not large enough to attract much private interest. Or hunting out new uses for existing drugs. Or the search for new materials to extend long-term battery storage.
Right now, we’re seeing promising pilot projects. Microsoft and the U.S. Department of Energy’s Pacific Northwest National Laboratory worked together on a project that used A.I. to analyze more than 32 million materials to discover an electrolyte that might improve lithium-ion batteries. But projects like these are a sideshow right now as the A.I. giants chase the guaranteed revenues of corporate clients. Only public funding has the possibility to bend the industry toward public problems.
Then there’s the way A.I. could change the relationship between citizens and a government that is often maddeningly difficult to navigate. President Trump killed the I.R.S.’s Direct File program, but the technology exists now to go far beyond it: There could be an A.I. system that works through your taxes with you, grounded in both the I.R.S.’s data on your income and the most up-to-date information on the tax code. Every person could have the equivalent of a personal accountant. More ambitiously, you could imagine that same L.L.M. as an entry point to government services: a kind of digital concierge to everything the government has that might help you.
When we want A.I. to solve real problems for real people, we need to make sure the data exists. That means cleaning up government data sets that are currently in a shambles (a project that the province of Alberta’s government found A.I. could make much faster and easier). It may also mean funding the creation of novel data sets that could eventually give A.I. systems traction on scientific problems that are currently beyond our capability to solve. Those data sets — like the Protein Data Bank — would be public goods, and so would need to be funded by the public.
What I’ve offered here is only the barest sketch toward an agenda. But perhaps it’s enough to make a broader point: If we want an A.I. that serves the public good, we need to define the public goods that A.I. can serve and create the conditions under which A.I. can be useful.
That means answering a question that’s been somewhat ignored. We know what we fear A.I. will do to us. But what do we hope it will do for us?
