Written by Dan Mogin from Mogin Law LLP on March 20, 2026
Some technologies cannot be controlled ex ante without assuming regulators possess powers they do not.
In his March 20 Wall Street Journal opinion piece – The Economics of Regulating AI – economist Roland Fryer argued that much of today’s artificial‑intelligence regulation is failing because regulators cannot observe what they most need to know. When lawmakers cannot distinguish between safe and risky systems, firms respond rationally. They retreat from beneficial tools, complying on paper, and revealing less rather than more about real‑world risk.
Fryer is right. But the enforcement problem he identifies runs deeper than information asymmetry. Artificial intelligence is colliding with a structural limitation of law itself: some technologies cannot be meaningfully controlled ex ante without assuming regulators possess powers they do not.
AI is not a discrete product, factory, or industry. It is a general‑purpose capability that now permeates nearly every sector of the economy, often invisibly and without centralized deployment. Enforcement models designed for stable actors and inspectable processes, such as licensing audits and reporting mandates, strain when applied to systems that evolve continuously, cross borders effortlessly, and are frequently used informally by individuals rather than according to formal – and governable – protocols.
Employment setting alone reveals regulatory futility.
Employment law illustrates the problem in stark relief. Even within that single field, regulation is layered and dense. Federal statutes and regulations interact with fifty separate state regimes, thousands of local ordinances, civil‑service rules, and collectively bargained agreements. These laws are designed to govern conduct – not machines – across millions of daily decisions involving hiring, evaluation, promotion, discipline, and termination.
AI is already influencing many of those decisions. Managers rely on résumé‑screening tools, generative drafting aids, and recommendation systems that update constantly and are often adopted without formal approval. These uses are decentralized, episodic, and embedded in routine work. No regulator can monitor them in real time. No employer can realistically police every instance of use. Even if federal policymakers succeed in harmonizing standards, uniformity does not solve the enforcement problem when the regulated activity consists of countless individualized judgments made across a sprawling workforce.
The result is predictable. When regulation cannot reliably distinguish safe from unsafe behavior, incentives misfire. Some firms abandon tools that may improve fairness or efficiency because legal exposure outweighs potential benefit. Others comply formally while obscuring how AI is actually used. Disclosure mandates and notice requirements create the appearance of oversight while teaching regulators little about actual risk.
Regulations that produce paper work rather than insight.
This is not merely a problem of regulatory design; it is a problem of enforceability. Rules that assume regulators can observe and assess millions of AI‑mediated decisions inevitably produce paperwork rather than insight. The gap between what the law demands and what actually happens widens, even when actors are acting in good faith, though systems may also skirt applicable law; or minimize legal impacts. Who will be in charge: regulators or AI designers? Will AI regulation of employment at the federal level be concentrated within an existing structure like the Department of Labor, with little AI expertise and limited jurisdiction or will AI be given over to a specialized AI regulatory agency without labor and employment expertise? What sort of AI behaviors will be restricted; how will structural regulation be implemented?
We have seen this dynamic before. Lawmakers confronted similar challenges with the rise of the internet and social‑media platforms, i.e., general‑purpose borderless technologies adopted at scale before their capabilities, uses, and consequences were fully understood. Early confidence in comprehensive regulation gave way to reactive hearings, fragmented rules, and years of litigation after harms had already become entrenched. Platforms evolved faster than statutes. Users adapted faster than enforcement. Jurisdictional boundaries proved porous.
Artificial intelligence follows the same structural trajectory, but at far greater speed and scale. And unlike earlier technologies, AI systems can learn, adapt, and alter their behavior autonomously, ensuring that the object of regulation is always in motion.
Like trying to regulate the wind.
The mistake is not the desire to regulate AI. It is the belief that enforcement agencies can meaningfully control ubiquitous, amorphous, and self‑evolving systems through prescriptive rules alone. Law does not regulate the wind itself. It regulates structures, incentives, and responsibility when damage occurs. Buildings must be designed to withstand gusts. Utilities must manage risk. Liability attaches after harm.
AI demands a similar approach. Effective governance will focus less on micromanaging technology and more on assigning responsibility for its use. Employers can be required to supervise decision‑making, document meaningful reliance on automated tools, and respond reasonably to known risks. Vendors can be held accountable for reckless design or misrepresentation. Courts can apply familiar doctrines—causation, negligence, intent—to AI‑mediated harms. But will it work?
This does not reject Fryer’s incentive‑based critique; it extends it. When enforcement is infeasible, even well‑designed regulatory menus will struggle unless they align with what law can realistically observe and adjudicate. Accountability after harm, not the illusion of perfect foresight, remains the legal system’s comparative advantage.
The danger in today’s regulatory push is not that it is too ambitious, but that it is too confident. Pretending regulators can see inside every black box encourages rules that look tough while accomplishing little. That displaces responsibility rather than clarifying it.
