Nobody feels great about AI right now. The people running the frontier labs say in public that they’re worried about what they’re building. The incumbents say the same. Young people are worried about jobs, the public is worried about safety and governments aren’t sure what to do.
Markets are trying to figure it out, too. On October 26, Anthropic filed to go public, and reports put its valuation between $1.5 and $2 trillion. Three days later, Alphabet lost nearly 13% of its value in a single day, about $500 billion, on worries about whether its AI spending will pay off.
A Third Way to Govern AI
Right now, there are two approaches to govern AI. One is the status quo: patchy rules like the EU AI Act and California’s SB 53, safety left to a few firms policing themselves and creators paid one lawsuit at a time. It isn’t going great. The other is heavy government involvement: licensing, a new regulatory agency, steep taxes or a government stake. That model is common in industries the public loves to hate, like healthcare, airlines and telecom. It hasn’t come to AI yet, but if AI starts eliminating jobs faster than the economy can absorb them, heavy regulation could come quickly. We think there’s a third way that’s simpler than either and much better tested.
What Good AI Rules Should Do
The fear with regulation is that it creates an industry of disputes and injunctions. Complex rules slow everyone down. They most hurt small firms that can least afford compliance. So we start with a narrower question: What are the big things AI rules have to handle?
We think there are three.
Safety: In a survey of 2,778 AI researchers, the median respondent put the chance that AI leads to human extinction, or something similarly permanent, at 5%. Between 38% and 51% put the chance of a generally bad outcome at 10% or more. For most of us, that’s not close enough to zero. Even if the worst never happens, AI amplifies our ability to do harm at least as much as it amplifies our ability to do something good. Whether it’s a catastrophic cyberattack or something more conventional, AI is likely to be at the center of whatever is the next global event we wish hadn’t happened.
Fairness: The most capable models were built from a public inheritance: the open web, public research, and decades of books and articles. We don’t think that should become a permanent private claim. Fairness sits underneath a lot of our politics right now. In March, a record 59% of voters told NBC News that the economic and political systems are stacked against people like them. People who feel that way, on the left and the right, are open to almost any alternative that promises to change it. Technology, more broadly, was already widening the wealth gap in the U.S. Economists Daron Acemoglu and Pascual Restrepo estimate that automation accounts for 50–70% of the change in the U.S. wage structure over the last four decades. AI is that same technology with the dial turned up, and it’s headed for job-heavy sectors like trucking and for white-collar jobs that used to feel safe.
Competition: The big IPOs are net positive. They broaden ownership and require greater transparency from the firms. But the early indications are that we’re headed for a tech oligopoly as opposed to a public good that, for example, internet infrastructure has mostly become. Search went one way. In 2024, a federal judge ruled that Google holds an illegal monopoly in general search. The internet’s plumbing went the other. Nobody owns the protocols that move data across the internet, or Linux, which runs most of the world’s websites. Anyone can build on them, and much of the tech industry, Google included, did. Concentration is expensive. Since 1980, average price markups in the U.S. have risen from 18% above marginal cost to 67% now. When more firms can build on last year’s frontier the way they build on internet protocols and Linux, more of the public participates:
-
As customers, through lower prices and thinner margins
-
As builders, founding and working at the next hundred firms
-
As workers, with more employers competing for their talent
Competition also keeps a healthier distance between industry and government. A few megafirms that outspend the public interest on lobbying is exactly the arrangement antitrust expert Tim Wu warns about in “The Curse of Bigness.”
Our Proposal: Commons in, Commons out
We propose a simple rule. Any firm that trains AI models on material it hasn’t licensed, or builds on a model that used unlicensed data, must release the weights of those models to the public once a year. Firms keep a lead of six to 18 months, or about a year on average. The IPOs still happen. The value just spreads out.
Consider how this one rule impacts the three key AI industry variables that need to be managed.
Safety: There’s no way to regulate AI the way we regulate nuclear weapons or human cloning. Those bans work because they depend on rare materials, specialized facilities and a small number of people with know-how. AI depends on chips and files. Models get copied, leak and get rebuilt by others. Whether the threat is catastrophic, like a malevolent AGI or a bad actor with bad intentions, or the everyday work of cybersecurity, the safest arrangement is one where the good guys have at least what the bad guys have, and ideally a little before.
We’ve spent the last 20 years learning how this works in computer security. In the early 2000s, software vendors sat on reported vulnerabilities for a year or more, sometimes never disclosing them. The theory was that what attackers didn’t know couldn’t hurt anyone. It didn’t hold up. Today, the standard is coordinated disclosure. The vendor gets a fixed window to fix the problem, and then the details go public. Google’s Project Zero settled on a 90-day deadline. By 2019, it reported that 97.7% of the issues it found were fixed inside that window. A deadline everyone knows about works better than hopes and secrets.
A fixed annual release date is the same bargain. The firm, its customers and the national security community get about a year with a model before everyone else does, and they know exactly when that year ends. That’s time to find the dangerous capabilities and build the defenses against them. And like disclosure deadlines, the lag can tighten as the field matures.
Competition: We already run a system like this, and it works, if imperfectly. A patent gives an inventor about 20 years of exclusivity in exchange for publishing how the invention works. The public grants a head start and gets the knowledge back on a schedule.
AI might be different. But how? The best argument is that too much disclosure removes the incentive to invest at the frontier. We don’t find that persuasive. First, a year’s lead is still worth a great deal, as the IPO valuations show. Second, moving faster on frontier AI is the one thing almost nobody wants. In five national polls this year, between 55% and 77% of Americans said AI is moving too fast or should slow down. In the AP-NORC poll, just 8% said it’s moving too slowly.
Fairness: What was built from the commons goes back to the commons, once a year.
The Window Is Closing
None of this rules out more regulation. To the public, we’d say: Try this first, and if it isn’t enough, ask for more. To insiders, we’d say: This is your best bet for avoiding something heavier.
Here’s why timing matters. Right now, the people with a direct financial stake in how AI is governed are a fairly small group of founders, employees and investors. Most of them are going to do extraordinarily well.
Under Nasdaq’s new fast-entry rules, a newly public company of significant size can join the Nasdaq 100 after 15 trading days. From there, it lands in index funds, pensions and the retirement accounts of the 62% of Americans who own stock. Once that happens, every rule that touches AI’s value has tens of millions of people on the other side of it. Acting then will be a lot harder than acting now.
We don’t think that’s a reason to stop the IPOs. Under what we propose, they still happen. But it is a reason to settle the rules first.
Visualizing the 3 Paths for AI Rules
Here’s our scorecard, including where our proposal is open to argument.