Sam Liccardo, a Democrat, represents California’s 16th district in the U.S. House of Representatives.
Thursday’s announcement that the Trump administration will delay its executive order on increased government scrutiny of artificial intelligence models provides a breather in its whiplash-inducing approach to the technology. The administration first proposed penalizing states with AI safety regulations without any federal substitute. Then it suggested a federal vetting scheme that would prohibit the release of models deemed unsafe, only to quickly backtrack. Meanwhile, as the Pentagon purged Anthropic’s model from its computers, the National Security Agency installed it in its classified networks. Whew.
A well-functioning legislature — if we had one — would lend statutory order to this chaos. And it could do so without stunting development of American AI with excessive regulation that would hand China the keys to the 21st century.
Legislating on matters of technology doesn’t come naturally to Congress, however. The body has urgently deliberated measures to protect children from online harms since the late 1990s, for example. Those children now have their own (unprotected) children.
Discussion on Capitol Hill has focused on how federal regulation might mitigate the many risks posed by AI’s rapidly expanding use, from the common — such as discriminatory bias or privacy violations — to the catastrophic, such as disseminating blueprints for making biological weapons. Yet the conversation has largely missed a central problem: We can’t regulate it.
That is, traditional regulation won’t contain AI’s risks effectively. The functional obscurity of AI neural networks, often described as the problem of “interpretability,” limits the efficacy of AI testing and assessment. While the AI labs can mitigate some risks by manipulating algorithms, model weights and training data, they cannot predict or explain every outcome. Well-intended regulators might mandate guardrails to prevent an AI agent from, say, emptying millions of bank accounts, yet a sophisticated criminal, often aided by AI, can induce a model to “jailbreak,” and a rogue AI agent could autonomously introduce its own perils.
Don’t expect Congress to legislate these risks away. Several colleagues have proposed good first steps, such as enacting federal equivalents to requirements approved in California for frontier model developers to publish safety protocols, independent evaluations and testing results. But those modest regulations should not make us sleep much better, given the limited reliability of model testing.
There’s a better approach to governmental intervention: not to regulate, but to referee.
A regulator proactively establishes a set of safety rules and requires that a product or service satisfy them to access U.S. markets. A referee, by contrast, evaluates the state of play in an industry and determines which products are safest. Those safest competitors set the standard by which all market participants obtain the referee’s benefit or sanction.
An AI referee would incentivize a “race to the top” for AI safety by leveraging the two primordial motivations of every hyper-scaler: crushing their competitors and avoiding liability. Every AI lab increasingly struggles to navigate a minefield of potentially conflicting AI safety laws from 50 states. But if safer AI design entitles a model’s developer to avoid that liability thicket, it could gain a decisive regulatory advantage over less safe competitors.
I am introducing a bill to empower the Commerce Department’s Center for AI Standards and Innovation (CAISI) to define the industry’s best safety practices within any specified category of risk, and to identify which frontier models have achieved those best practices. Companies’ AI models earn federal preemption — a safe harbor from liability under state law for a specific category of risk — if they meet or exceed CAISI’s determination of industry best practices for cybersecurity, human safety, privacy and the like.
This approach appears better suited for a fast-evolving technology than a fixed safety regulatory standard. As competitive incentives push developers to improve model safety, CAISI would raise its bar to reflect the newest best practices. Any legislation must also ensure CAISI has strong whistleblower protections, of course, and fire walls on political influence — particularly under this administration — to support agency integrity.
Other agencies would stay in their lanes. Public wariness will continue to spur state governors and legislators to get tough on AI hyper-scalers, so the impetus will grow for companies to meet CAISI’s safety thresholds for preemption. Companies that use AI — such as mutual fund managers, pharmaceutical manufacturers or banks — would remain subject to the same outcomes-focused regulations under the Securities and Exchange Commission, Food and Drug Administration or Federal Deposit Insurance Corporation, regardless of what tools they use.
This approach does not eliminate AI risk. No law will. It acknowledges, though, that in the absence of a Delphic oracle, we need to learn much more about model testing and evaluation before we can regulate risk effectively. In the meantime, our best hope lies in clear rules that strongly incentivize the private sector to do what the government cannot: make AI safer.
The post AI doesn’t need a regulator. It needs a referee. appeared first on Washington Post.
