A modest proposal for licensing, liability, and accountability
Because powerful AI empowers everybody to do a bit of everything, from law to medicine to refrigerator repair, it seems almost obvious that policymakers should reduce or discard many licensing barriers. But on closer inspection, this might make the underlying problems worse, a bit like throwing water on a grease fire. Perhaps we should instead recommit to licensing schemes in principle—in order to preserve human stewardship of important cultural, economic, and political activities—but also embark on a project to reimagine those schemes in light of a new technological reality. That project should cover not only who provides AI-powered professional services, but also who gets to use the most powerful AI tools.
Let us first steelman the idea of shredding lots of red tape. It would certainly be a good thing to create new opportunities for talented people who lack formal qualifications, and to drive down the price of unaffordable services. We must not lose sight of these goals.
But if professional qualifications are eliminated wholesale, leading AI providers will become the standard-bearers for all those services. The human contribution to legal, medical, educational, and even executive services may become increasingly marginal and low-value. This points toward at least three worrisome outcomes:
- Power concentration at the AI firms
- Human disempowerment and deskilling
- The transformation of important human practices
The first two are familiar; the third needs explanation. Law, medicine, and other venerable professions are not simply static functions or services provided to society. They are evolving human practices, with enormous political and cultural import. The professional cultures and associations that steward these practices get some things right and many things wrong. But however much they are in need of reform, the power these professions wield is almost certainly not something we want to either automate or turn over to AI companies.
Are we quite sure we’d be happy, for example, if medical ethics norms were set by Anthropic, or the government, or the market? Human practitioners play an important part in stewarding medical ethics, and the same is true in law, education, and countless other licensed fields. (It is worth noticing that one prestigious field that has effectively no licensing requirements—software engineering—has in recent decades not covered itself in ethical glory.)
Well-designed new licensing systems could keep humans in charge of important practices, incentivize the continued maintenance of human expertise, and preserve meaningful human roles in the economy. One possible paradigm is to organize the professions around “accountability rents” as opposed to “complexity rents,” as the excellent “bull” case for work and expertise put it. In the twentieth century, the professions thrived because people needed specialized high-skill services to navigate the complexity of modern life. But in the AI era, our problem is not an excess of complexity, but a dearth of accountability. When a treatment fails, or a bridge collapses, or a legal strategy proves ruinous, it is becoming almost impossible to find anyone who is actually responsible—only a mess of highly automated processes. This is, in a sense, our new problem. Let us therefore put people on the job. Critical work should require an accountable human, and that accountable human will need new skills: not only an understanding of, say, medicine or civil engineering, but the technical skills to oversee mostly automated work occurring at vast scale. In short, embracing accountability rents could kill three birds with one stone: addressing the accountability crisis, keeping our institutions human, and creating human jobs.
It is easy to imagine this new licensing regime being perfectly compatible with accelerating productivity and AI use. For example, AI tools already perform tasks that would count as unlicensed practice if a human did them. So what if, instead of drafting a legal contract for you directly, Claude smoothly handed you off to an AI-powered legal provider of your choice, who would do the same? This new ecosystem of legal service provision would still be absolutely tech-forward and user-friendly; it would be competitive and nowhere near as expensive as hiring a lawyer today. But the end-service provider would be a partnership of licensed professionals, as law firms already are, instead of a general-purpose corporation. This provider would assume the liability connected with its services. Certainly, this would require rethinking the whole pipeline of legal certification, from law school to bar exams. But that needs to happen anyway. In addition to its other benefits, this would prevent a few technology companies from becoming the de facto owners of our entire legal, medical, and professional culture.
Finally, licensing schemes can settle not only who gets to provide certain AI services, but also who gets to use them. Increasingly, the AI companies are acting as de facto regulators by banning and monitoring uses of their products that they deem unsafe, illegal, or threatening to their competitive advantage. Blurring these motives together lets them abuse their power. Why not, then, treat certain types of AI use as privileges licensed by public authorities rather than private ones—analogous to operating heavy machinery? It seems appropriate for states, not corporations, to adjudicate who has the privilege to use AI for advanced machine learning research, or biomedical inquiry, and under what circumstances. In both cases, the aim is the same: to keep important practices under human and public stewardship rather than ceding them to a handful of AI firms.

