The Free Enterprise Nation All articles
Entrepreneurship & Free Markets

A Permission Slip for Every Model: Why Preemptive AI Licensing Would Freeze the Field

The Free Enterprise Nation
A Permission Slip for Every Model: Why Preemptive AI Licensing Would Freeze the Field

Photo: DFID - UK Department for International Development, CC BY 2.0, via Wikimedia Commons

Every generation of American technology has produced the same legislative reflex: something new and powerful appears, and Washington concludes that the responsible course is to require a federal blessing before anyone is permitted to build it. Railroads, radio, aviation, and pharmaceuticals all passed through that gate. Sometimes the result was sensible. Often it was a moat.

Artificial intelligence now stands at the same threshold, and the proposals on offer — licensing regimes, pre-deployment certification, mandatory liability insurance scaled to model capability, and agency sign-off before training runs — deserve to be examined for what they would actually do rather than what their sponsors say they would do.

Who Can Afford the Waiting Room?

A licensing regime is not a neutral safety filter. It is a fixed cost, and fixed costs are regressive. A large laboratory with a compliance department, outside counsel, and a government affairs office can absorb an eighteen-month review without breaking stride; the review is simply another line item. A two-person startup training a specialized model on a rented cluster cannot. Neither can an academic group, a hobbyist, or the open-source community that has done so much of the foundational work in this field.

The predictable outcome of pre-clearance is therefore not fewer models. It is fewer competitors. The firms that write the largest checks in Washington would also be the firms best positioned to survive the process they helped design — a coincidence that history suggests we should not treat as one.

The Open-Source Blind Spot

Regulators tend to write rules for the world they can see: a named company, a filed application, a known model. Open-source development does not fit that template. Weights are published, downloaded, fine-tuned, and repurposed by people no agency has ever heard of. A statute that makes it unlawful to distribute an unlicensed model does not stop misuse; it stops distribution. The result is that the beneficial uses — medical imaging tools built in a university lab, translation models for low-resource languages, fraud detection for small credit unions — disappear, while bad actors simply operate outside the registry.

Europe's experience with its 2024 AI Act offers a cautionary preview. The compliance obligations are layered and real, and the early consequence has been a great deal of lawyerly activity and a noticeable chill among smaller developers. American policymakers should be able to learn from that without repeating it.

Liability, Not Permission

The alternative to licensing is not a vacuum. It is law — the ordinary, ex post kind that governs nearly every other product in America. If a company deploys a system that defrauds consumers, discriminates in violation of existing civil rights statutes, or causes measurable harm, it should be sued, fined, and held to account. Clear rules of liability, applied after the fact, punish actual misconduct rather than speculative risk. They also leave room for the countless beneficial applications nobody in a hearing room anticipated.

A workable framework would rest on four pillars:

  1. Transparency where it matters. Disclosure of training data provenance, known limitations, and evaluation results for systems used in consequential decisions — housing, credit, employment, medicine.
  2. Sectoral enforcement. Existing regulators already have authority over lending, insurance, securities, and medical devices. They should use it rather than cede the field to a new super-agency.
  3. Contract and tort, left intact. Users should be able to negotiate terms, and courts should be able to award damages. Preemption of state tort remedies in exchange for a federal permission slip is a terrible trade for consumers.
  4. No speech licensing. A model that generates text is, among other things, a printing press. Prior restraint of expression, however dressed up as safety certification, runs headlong into the First Amendment.

The Cost of Getting This Wrong

The United States holds its lead in this industry not because of federal guidance but because of an open labor market, deep capital markets, cheap compute, and a culture that tolerates people building things before regulators understand them. Each of those advantages is a policy choice, and each can be undone.

There is no shame in admitting that legislators do not yet know enough about this technology to write a durable rulebook. That is an argument for humility, for targeted enforcement against demonstrated harm, and for letting a thousand experiments run under the ordinary watch of courts and consumers. It is not an argument for handing the incumbents a licensing regime and calling it consumer protection.

Guardrails belong on roads. In markets, the better answer has always been competition, disclosure, and the rule of law — applied evenly, and applied after the fact.

All Articles

Related Articles

Where the Next Apple Goes to Die: Municipal Bureaucracy and the Death of the Garage Startup

Where the Next Apple Goes to Die: Municipal Bureaucracy and the Death of the Garage Startup

The Toll Booth Before the Front Door: How Regulatory Costs Bury Small Businesses Before They Open

The Toll Booth Before the Front Door: How Regulatory Costs Bury Small Businesses Before They Open

Hooked on Handouts: How Federal Subsidies Rig the Market and Punish the Independent

Hooked on Handouts: How Federal Subsidies Rig the Market and Punish the Independent