Matt Mittelsteadt
In recent weeks, the specifics of Congress’s proposed state AI regulatory moratorium have dominated AI policy discussions. Because it’s unclear if this specific approach can pass congressional muster, it’s essential to keep focused on the underlying “why”—regulatory harmonization.
This year, a steady drumbeat of state legislatures has passed AI regulations. What was once a small regulatory club has rapidly expanded to include economic heavy hitters like Texas, California, and soon, New York. Elsewhere, similar regulations are on the way. By one estimate, the count of pending state AI bills numbers in the thousands. Today, the United States is rapidly sleepwalking into a fragmented state patchwork.
Such division is a major problem.
Unlike a unified national approach, unharmonized state regulations would incur significant added costs divorced from any potential value the regulations may offer. The extra cost most cited is the strain on innovation and productivity. With each overlapping law, would-be innovators will be forced to divert ever-growing sums from R&D toward compliance, customization, and expensive legal counsel—resources only large firms can typically afford. Without harmonization, we risk stagnating innovation, slowing productivity growth, and concentrating benefits.
These economic risks only tell part of the patchwork-cost story. Less emphasized, yet perhaps more important, are the added safety harms we could incur if policy is fragmented. As safety promotion is unquestionably the aim of most AI regulations, policymakers must contend with the no-benefit costs that an unharmonized state-patchwork would bring.
To understand potential risks, let’s consider two significant ways the coming patchwork may undermine the very safety legislators hope to promote.
Transparency Confusion
The first added risk is transparency. Today, algorithmic transparency rules are perhaps the most common denominator across state AI regulatory proposals. To the credit of legislators, transparency can indeed help minimize safety concerns. With solid data, consumer choice can be better informed and risks appropriately managed. Such benefits, however, depend on data being clear, simple, and ideally aggregated. A patchwork nurtures the opposite. From a multitude of transparency regulations will naturally spring a confusing collage of differing standards, measures, and conclusions. Counterintuitively, more transparency rules could yield less transparency.
Given the current AI reality, such unharmonized transparency rules are likely. In industry, there is little consensus on measuring “AI ground truth.” A first challenge is definitional: what even is AI? Because AI is not a specific technology but more a general notion or goal, there are hundreds of possible definitions and little consensus. That opens a wide door to policy diversity and challenges a consistent approach to regulatory scope.
A second difficulty is measurement. Evaluation obsolescence is a persistent industry challenge: almost as soon as evaluation criteria are introduced, they are rendered moot by shifts in the technical landscape. As a result, gold standard metrics are in constant flux and ballooning in number as experts introduce countless would-be replacements to attempt to fill the void. Such a unique swirl means various state transparency regulations are almost certain to measure and report inconsistently.
These realities are a breeding ground for confusion and perhaps an opening for consumer harm. If definitions of AI are inconsistent, for instance, it’s easy to imagine a consumer in state-straddling Kansas City seeing a service labeled “AI” on one block and not AI a few streets over. Likewise, if states create a mess of uneven evaluations, consumers are sure to misinterpret safety data, or worse, tune out evaluations altogether.
Unlike a unified national approach, fragmented transparency regulation naturally invites conflict and confusion. While it’s hard to predict what future harms transparency efforts might mitigate, if there are risks, a clash of regulatory data will do little to help.
Denial of Safety-Enhancing Technologies
A second, more significant added cost is the denial of safety-enhancing AI technology. While AI is often narrowly pigeonholed as an efficiency driver, the most critical emerging use cases involve automating tasks humans have demonstrably failed to manage safely.
A great example is cybersecurity. In 2024, the number of discovered software vulnerabilities surged 38 percent. In 2025, meanwhile, the number of cyberattacks grew a remarkable 47 percent. As the volume of risks rapidly balloons, human defenders have failed to keep pace. The result has been a litany of real, physical harm. In 2024, a cyberattack on Change Healthcare left thousands of hospitals unable to process transactions. This forced delays in medically necessary care and direct patient harm.
Where humans have failed, however, defensive AI tools offer a glimmer of cyber hope. Early evidence suggests countless just-emerging tools can spot novel insecurities, write programming fixes, update flawed legacy systems, and autonomously detect attackers. In a few short years—if not months—AI could drive a digital safety revolution and prevent further harm.
Driverless vehicles offer an even more compelling AI safety story. It’s no exaggeration to claim human drivers are a safety liability. In 2022, there were 44,000 motor vehicle fatalities on American roadways and another 2.6 million crash-related emergency department visits. Against this safety crisis, AI provides hope. According to a recent study from Swiss Re, an insurer, Waymo’s driverless cabs yielded a remarkable “88 percent reduction in property damage claims and a 92 percent reduction in bodily injury claims” compared to humans. With such staggering figures, driverless cars could be the single biggest safety innovation in our lifetimes. In a matter of years, AI may all but eliminate this leading cause of death.
These specific examples are worth highlighting because their singular potential hinges on regulatory harmonization. In the case of cybersecurity, digital systems are often deeply integrated across jurisdictions, and therefore, safety success demands consistent tooling across state lines. If even one state denies or limits essential AI security tools, it could create an unsecured weak point and easily spread attacks to all others. Interstate consistency is more essential in the case of driverless vehicles. If consumers or firms can’t legally drive across states due to a patchwork, they simply won’t use the technology. It’s hard to imagine the market demand for a state-limited car.
In both cases, lives are on the line. If a convoluted regulatory patchwork emerges, it could cost both substantial safety gains and preventable deaths.
Conclusion
These safety costs are significant but hardly a panacea. As state frameworks grow more fragmented, new unintended safety consequences will emerge. While states will always play a policy role, policymakers must recognize that benefits can be best maximized with a consistent, simple, national approach. If we truly wish to ensure the noble goal of safety, harmonization must be an imperative.