Catalytic Regulation: Incentivizing Safety During a Regulatory Drought
Apr 20, 2026
In 1959, a midsize Swedish car company did something its competitors thought was myopic, if not reckless. It effectively open-sourced the three-point seat belt, the greatest safety innovation in automotive history. The prevailing industry wisdom at the time was blunt: “Safety doesn't sell.” Just three years prior, Ford had offered seat belts for a $9 surcharge, as part of its 1956 Lifeguard campaign; despite Robert McNamara's championing of the program, the safety push failed to give Ford a competitive edge. Henry Ford II reportedly grumbled, as he was dialing back its campaign, “McNamara is selling safety, but Chevrolet is selling cars.” But Volvo was neither myopic nor reckless; in fact, it saw further than any of its competitors. While they competed fiercely for dominance in a race for horsepower, engine efficiency, and design, Volvo could see that consumers cared about safety and reliability, too. The bet paid off: Volvo became one of the most recognized automotive brands in the world. According to Volvo, seat belts have since saved over 1 million lives.
American AI needs its Volvo moment.
The aim of catalytic regulation is to enable this moment. It is a family of positive incentives designed to channel market forces toward safety. Where traditional regulation works through mandates and penalties (“do this or else”), catalytic regulation works through rewards. Think tax credits for safety R&D, procurement incentives for verified-safe systems, and prestige mechanisms that make safety a competitive Schelling point. The goal is not just to subsidize safety at the AI industry’s margins, although that alone would be worthwhile. It is to catalyze a deeper shift in the culture that animates American AI innovation, marking safe and powerful AI as the very thing that American labs can do better than any competitor.
Let us now explore how catalytic regulation can meaningfully improve AI safety both today and in the future, before describing some specific example regulations.
Why should we turn to catalytic regulation? For one thing, positive incentives are more realistic in today’s political environment, but they can also lay the groundwork for stronger regulations in future, while promoting better general safety culture within organizations.
Current deregulatory trends favor catalytic regulation. Any discussion of AI regulation must start by acknowledging political reality. In the United States, policymakers increasingly view AI through the lens of strategic competition with China. That outlook has produced a strongly deregulatory federal stance, grounded in the belief that limiting regulation is necessary to preserve innovation and US competitiveness relative to its international rivals. America’s AI Action Plan, which the White House released in July 2025, frames AI policy almost exclusively as a reaction to an AI “arms race,” calling for the curtailment of state-level regulation and the removal of barriers to innovation. Even relatively light-touch proposals face organized resistance from a tech lobby that has successfully cast any regulatory measure as saddling American developers with European-style burdens. For better or worse, there is currently relatively little organized political appetite for — and strong industry pressures against — traditional command-and-control regulatory measures such as mandates, prohibitions, and prescriptive safety requirements.
Catalytic regulation evades these barriers. Tax credits and procurement incentives neither impose compliance costs nor slow development. They do not ask senators to stand up and argue for restricting US companies. The pitch is simpler and politically commonsense: we are investing in US leadership toward safer AI. This strategy works with the grain of deregulatory attitudes. Indeed, as the AI race intensifies, the case for positive incentives only gets stronger, while the case for traditional regulation gets weaker.
Catalytic regulation complements other safety efforts. Catalytic regulation is not merely a second-best substitute for the regulation we wish we had. Suppose there was the political will to legislate AI safety: what standards or norms should we require labs to satisfy? There are already several reasonable ideas: it is not wise to release models that can contribute to the development of chemical and biological weapons; red-teaming can expose latent development gaps; and using models to perform behavior audits on other models can often improve robustness. But there is still much we do not know about how to make today’s models safe. For instance, studies have found that safety benchmarks are heavily correlated with model capabilities, meaning that capabilities progress can be misleadingly portrayed as safety improvements and making it harder for us to articulate clearer standards. We do not yet have the three-point seat belt for AI.
Here, the utility of catalytic regulation is that it can spur advances in our understanding of safety that are desperately needed to define any broader regulatory regime. If catalytic tools advance safety practices, benchmarking infrastructure, and institutional capacity today, they make future mandates more effective should the political winds shift. In this sense, catalytic regulation is also complemented by traditional regulation; while positive incentives alone may not ensure that all market participants adopt high safety standards, they can establish best practices as a standard of “reasonable care,” a powerful anchor in negligence law that regulators can point to when they do act. This pro-safety infrastructure, if built during the current regulatory drought, could become a foundation from which to manage the flood of AI disruption.
Catalytic regulation influences organizational culture. A broader benefit of catalytic regulation is the way it can influence “organizational culture” — the general attitudes that govern what people do “when no one is looking.” Organizational culture fosters values, illuminating the blind spots that regulators do not observe. It is what determines whether the chef washes their hands when visiting the bathroom, whether a NASA engineer stays overtime to check the numbers once more, and whether a nurse reports a surgery accident or closes ranks with the attending surgeon. Given the countless pockets of local knowledge that exist within AI labs (while remaining unknown to Washington regulators), inculcating safety as a value within organizational culture is key to minimizing future AI harms.
We can draw lessons from previous reliable organizations. The paradigmatic examples of organizations suffused by safety culture are High Reliability Organizations (HROs), which have long been studied by organizational sociologists. Familiar examples of HROs include nuclear facilities, air traffic controllers, and aircraft carriers. What ties them together is what physicist-turned-political-scientist G.I. Rochlin, in a prologue to New Challenges to Understanding Organizations (1993), calls their “effective management of innately risky technologies through organizational control of both hazard and probability.”
One surprising lesson from studying HROs is that their safety culture can precede safety regulation, rather than following it. After the Three Mile Island accident, in 1979, the nuclear industry created a self-regulatory body, the Institute of Nuclear Power Operations, to instill professional safety norms even before regulators codified them. Volvo made the three-point seat belt standard in 1959, nine years before federal law required seat belts in new passenger cars and decades before using three-point seat belts became a ubiquitous norm.
There is no master recipe for instilling safety culture, but catalytic regulation has some key ingredients. It builds on, and contributes to, an attitude of collaboration between society and the organization. It is designed not to punish deviation but to reward attention and care. All American labs have their internal “safety factions”; catalytic regulation aims to reward and amplify their work.
Catalytic regulation has positive expected value, with little downside. Even if it does not solve every safety problem, and it will not, the marginal question is straightforward: are we better off with these tools than without them? The measures are low-cost, politically feasible, and compatible with every other regulatory strategy on the table. They foreclose nothing while building the institutional scaffolding that any future approach will need.
If policymakers pursue catalytic regulations for AI, which specific interventions are most promising? We will discuss four potential mechanisms: corporate incentives, demand-side incentives, market guarantees, and prestige.
The most direct form of catalytic regulation targets labs, the hubs of AI innovation, by offering tax subsidies for investments in AI safety research. In “Racing to Safety: Tax Policy for AI Safety-by-Design,” my co-author (tax expert Mirit Eyal) and I propose such a program.
The 1983 Orphan Drug Act provides a blueprint. The model here is the research and development tax credit, which offers firms a sizable incentive for investment in more basic forms of innovation. This credit has a long and successful history. One notable example is the Orphan Drug Act, which incentivized pharmaceutical companies to invest in finding life-saving interventions for rare diseases, left “orphan” by normal market forces. In the decades following the act’s passage, hundreds of orphan drugs received approval, creating an entire therapeutic category that the market had effectively abandoned.
The orphan drug credit did not solve a financing problem; pharmaceutical companies are already well resourced. Rather, it solved an allocation problem, redirecting where the next marginal dollar went. AI safety faces the same structural problem (with similar potential solutions). For instance, despite being flush with capital and publicly committing to devoting “20% of the compute we've secured to date” to its safety team, OpenAI later left the team with only 1% to 2%. Just as the orphan drug credit succeeded in directing the attention of highly capitalized pharma companies toward solving rare diseases, earmarked subsidies could incentivize the work of internal safety teams, especially given how low the baseline currently is. A safety R&D credit wouldn’t compete with a lab's total budget; it would shift the calculus when a research director decided how to allocate the next team of engineers or the next cluster of GPUs.
These kinds of research subsidies also have the advantage of scalability. Policymakers could limit support if AI capabilities development stalls, or increase it if the pace of progress continues unabated. The most serious objection to safety research subsidies is the difficulty in distinguishing genuine safety investments from general capability development; as mentioned earlier, many safety benchmarks appear to track general model capabilities without labs focusing specifically on safety. This concern is real, and we design attestation and audit mechanisms around it in the companion paper. But labs’ incentives to invest in capabilities are already saturated; the possibility that companies might design elaborate schemes to redirect modest safety subsidies toward capability research would be economically irrational, especially in light of potential reputational blowback.
While government regulations can shape the behavior of domestic corporations, AI safety is a global problem.
Even if the US were to implement regulations, jurisdictional and practical constraints would limit the direct effects on foreign developers, cloud providers, and other AI stakeholders. However, another form of catalytic regulations can help to bridge this gap; tax incentives directed at consumers, which offer rebates or credits for purchasing safe AI systems, can shape which products enter the US market, regardless of where those products originate.
US energy-efficiency standards for consumer products shaped global manufacturing. The energy-efficiency credit provides a model for a consumer-focused approach. When the federal government offered tax advantages for energy-efficient appliances, vehicles, and building systems, the immediate effect was domestic: American consumers shifted their buying habits toward qualifying products.
But the secondary effect was global. Foreign manufacturers seeking access to the American market adapted their product lines to meet American certification standards. Korean appliance manufacturers, Japanese automakers, and Chinese solar-panel producers all redesigned products to qualify. In this way, a domestic certification requirement effectively exported American energy standards without requiring any international agreement or diplomatic leverage.
The same model can foster AI safety adoption. The example of energy-efficiency incentives is directly applicable to AI. An “AI Safety Usage Credit,” providing a modest tax reduction for businesses and individuals subscribing to AI services that meet rigorous safety certification, would create a systematic price advantage for certified systems. A business choosing between two model providers would find that only one came with, say, a 10% tax credit per token. The certification would focus on properties that markets do not necessarily or fully price in: for example, robustness to adversarial inputs, resistance to deceptive alignment, interpretability of core reasoning, and safeguards against autonomous misuse. We may not yet have a comprehensive set of criteria for AI systems to meet, but we could start with common sense certification requirements, and update them on a semi-annual or annual basis, as safety innovation flourished.
The international implications are significant, because consumption subsidies give us leverage over labs that reside outside of our jurisdiction, potentially catalyzing a much broader cultural change in AI development.
Tax incentives can mitigate the cost to developers of investing in safe AI. However, they do not fully eliminate the “appropriability problem”: firms that share breakthrough safety techniques bear the development costs while competitors freely implement their discoveries. Insofar as safety and reliability are a market advantage in their own right, companies are incentivized not to share their advances even if doing so would benefit the public. To address this challenge, we can turn to a different form of catalytic regulation: market guarantees.
Government spending influences private-sector decision-making. A big part of Operation Warp Speed — which used federal dollars to incentivize private-sector COVID-19 vaccine development and safety assessment — consisted of using the federal government's purchasing power as a regulatory lever. Similarly, we could develop an “AI Safety Charter” program to give firms demonstrating safety leadership a bundle of concrete benefits: priority consideration in federal procurement, expedited security-clearance processes, streamlined regulatory compliance, and access to government datasets for safety research. As illustrated by the recent face-off between Anthropic and the Department of War, federal purchasing can be incredibly consequential even for market-leading AI labs.
Patents can regularize safety proliferation. The patent system can address the appropriability problem more directly. A lab that published a breakthrough interpretability method cannot prevent competitors from adopting it, yet the social interest in rapid diffusion conflicts with the private interest in maintaining a competitive advantage. Mandatory “fair, reasonable, and non-discriminatory” (FRAND) licensing at modest royalty rates could give innovators a period of competitive advantage without bottlenecking the diffusion of safety knowledge.
The mechanisms above operate through material incentives. But material incentives address only part of what drives organizational behavior. Status competition — the pursuit of professional standing relative to others — is at least as powerful a motivator, and, unlike with tax credits, its effects tend to persist after the initial intervention.
Federal recognition can bring market benefits. Governments already harness companies’ desire for prestige to shape their behavior, sometimes to striking effect. The Malcolm Baldrige National Quality Award, created in 1987, to promote “improved quality of goods and services” in the US, carries no cash prize. Yet an analysis commissioned by the National Institute of Standards and Technology estimated a benefit-to-cost ratio of roughly 820 to 1, with gains flowing primarily from process improvements that firms undertook in pursuit of recognition. The Occupational Safety and Health Administration’s (OSHA) Voluntary Protection Programs (VPP) tell a similar story: participants receive a worksite flag indicating that they meet VPP standards, and a listing on the OSHA website. Moreover, they directly see the benefits of adhering to the standards, seeing injury rates approximately 50% below industry averages. These programs work because visible honors shift what good engineers optimize for.
The AI industry appears well-positioned for such interventions. Researchers already compete intensely for conference acceptances, citation counts, and benchmark rankings. Labs recruit by advertising leaderboard positions. A “Presidential Frontier AI Safety Medal,” modeled on the Baldrige Award, would recognize breakthrough safety research through White House ceremonies, with deliberate scarcity (honoring perhaps three organizations annually) and a requirement that recipients publish certain safety practices or tools. A public safety leaderboard maintained by an attestation body would track which organizations solve recognized safety challenges first, harnessing the competitive instincts that currently drive benchmark races.
Prestige cannot be easily gamed. As a part of catalytic regulation, prestige incentives could help fill some of the gaps left over by financial incentives alone. One of those is “gameability.” Tax credits can, in principle, be gamed; prestige cannot be reliably fabricated. A lab cannot consistently fake the technical achievement required to top a public leaderboard or win peer recognition. If prestige incentives are well-recognized and hard to game, they might meaningfully shift research incentives. Companies may also see prestige from safety work as a way of attracting top talent, a scarce resource for which they are already competing ferociously.
Ultimately, catalytic regulation aims not to outspend the labs but to shift the competitive equilibrium. Financial incentives provide the activation energy, but the bigger, deeper, and more sustainable impact will come from cultural and competitive dynamics. When leading labs compete visibly on safety, they set norms that smaller labs adopt too — not because regulations require it but because the organizations they aspire to emulate have made safety a core part of what excellence looks like. The same competitive dynamics that accelerate risk-taking in the current capabilities race can, with modest institutional design, be redirected toward safety leadership.
Regulation is hard, especially for an unknowable future. Yet, during a period of intense skepticism toward traditional regulatory measures, catalytic regulation offers a path forward. At its best, it redefines the terms of the AI race, positioning safety as a non-negotiable reputational priority for American labs, and showing that safety guarantees make their systems more attractive to governments, enterprises, and end users worldwide. Even at its most modest, catalytic regulation still spurs safety advancements that seed the ground for any regulatory architecture that comes next. In 1959, Volvo gave away the seat belt and led the world. The question is whether US policymakers will recognize the same opportunity for American innovation, and whether they will do so before another nation does.
Current export rules focus on keeping chips a generation behind. They should focus on keeping America's total compute ahead.
Embodied AI is arriving faster than the regulations meant to govern it. If we start now, that’s a problem we can still fix.
