Tag: Autocrats Never Lose

Something remarkable has happened in artificial intelligence over the past several days. Some of the people building the world’s most powerful AI systems are asking to slow down. Anthropic CEO Dario Amodei has called for “pacing the frontier”—deliberately giving safety research, monitoring, and institutions more time to catch up with rapidly improving AI capabilities. OpenAI CEO Sam Altman and Elon Musk have publicly supported the basic idea. Major laboratories are discussing independent safety evaluation, increased transparency, and mechanisms for preventing competitive pressures from pushing everyone into a race that none of them necessarily wants to run at maximum speed. These are not environmental activists demanding that an oil company stop drilling, nor are they outsiders chaining themselves to the gates of a laboratory. They are the people running the laboratories.

Almost immediately, however, the most powerful Republican leaders in Washington have responded with some version of: not so fast on slowing down. President Donald Trump was asked on September 13 whether AI companies should slow development or face greater regulation. His answer centered not on the safety concerns being raised by AI researchers, but on China. “We’re leading China in AI,” Trump told reporters. “We’re the most sophisticated country in the world, and frankly I want to keep it that way because whoever wins AI wins.” Trump allowed that “we could put guardrails” around the technology, but then dismissed some of the more serious warnings as coming from “very negative forces” raising scenarios that, in his words, “won’t happen.”

House Speaker Mike Johnson has taken a somewhat more measured position, but his practical conclusion is similar. Johnson acknowledges that some safety measures are necessary. He has spoken about preventing AI from “running away” and says the issue is a priority. Yet when confronted with calls for Congress to intervene quickly, Johnson said lawmakers should “resist Congress jumping in and imposing some sort of emergency moratorium.” He suggested that AI companies themselves can stop or slow their work and said he would prefer industry leaders to assume primary responsibility rather than having Congress impose the solution. Johnson also emphasized the same geopolitical concern as Trump: slowing American AI too much, he warned, could allow China to gain an advantage and thereby create a national-security threat.

This position is consistent with the broader policy of the Trump administration. A June executive order created mechanisms through which developers can voluntarily provide the federal government early access to certain frontier models for evaluation and cybersecurity work. But the same order goes out of its way to say that nothing in it should be interpreted as creating mandatory governmental licensing, preclearance, or permitting requirements for developing or releasing new AI models. The administration wants evaluation. It wants cooperation. It wants cybersecurity. What it does not appear to want is a government speed limit on the frontier.

Perhaps that will prove to be the correct judgment. Perhaps the AI laboratories can police themselves. Perhaps the most extreme warnings are indeed exaggerated. Perhaps competition among American companies, combined with voluntary safety agreements and targeted government oversight, will turn out to be enough. But anyone who remembers the political aftermath of Sandy Hook should recognize another possibility. We may be watching a policy window begin to close almost as soon as it opened.

On December 14, 2012, a gunman entered Sandy Hook Elementary School in Newtown, Connecticut, and murdered twenty children and six educators. The political response was immediate. There are moments in American politics when an event seems sufficiently horrifying that the normal rules briefly disappear. Positions that appeared immovable suddenly look negotiable. Politicians previously reluctant to discuss an issue begin demanding action. Public attention becomes overwhelming. Something that had been politically impossible on Wednesday can appear nearly inevitable by Friday. Sandy Hook created such a moment.

President Barack Obama made gun legislation one of the central priorities of the beginning of his second term. Families of Sandy Hook victims traveled to Washington. Senators began negotiations. The resulting Manchin-Toomey proposal would have expanded background checks for commercial gun sales. It was hardly revolutionary gun-control legislation, and public support for the underlying policy was extraordinary. In May 2013, Pew Research Center found that 81 percent of Americans favored background checks for private gun sales and gun-show purchases. Remarkably, the topline number was essentially identical among Republicans, Democrats, and independents. Even after the Senate legislation itself became politically contested, 73 percent of Americans still wanted Congress to pass a background-check bill.

And yet the legislation died. On April 17, 2013—barely four months after twenty children were murdered at Sandy Hook—the Manchin-Toomey amendment received 54 votes in the Senate and 46 against. It had a majority, but it needed 60 under the procedure being used. The proposal failed. The extraordinary national moment produced by Sandy Hook did not disappear instantly. States including Connecticut enacted substantial reforms, executive actions followed, and gun-control organizations continued their work. But the federal window had effectively closed. By the end of 2013, Pew observed that the overall gun debate had already begun drifting back toward its previous political equilibrium. The horror remained. Public support for background checks remained broad. But the extraordinary sense that Congress simply had to do something had dissipated. Major federal gun-safety legislation would not finally pass until the Bipartisan Safer Communities Act in June 2022, following another horrifying elementary-school shooting, this time in Uvalde, Texas. Nearly a decade had passed.

That sequence deserves considerably more attention in the AI debate than it is receiving, because political windows do not remain open simply because the underlying problem remains dangerous. They close. Public alarm fades, legislators find other priorities, industries organize their opposition, and proposals that seemed urgent become subjects for another committee hearing or another study. A danger can remain exactly as serious while the political willingness to address it steadily disappears.

The most striking similarity between the gun debate after Sandy Hook and the AI debate today is therefore not that guns and artificial intelligence present similar dangers. They plainly do not. The similarity is the transition from abstract agreement to concrete policy. Almost everyone can agree with an abstract noun: safety, responsibility, guardrails, common sense. After Sandy Hook, one could find politicians across the ideological spectrum saying that children should be protected and dangerous people should not have access to firearms. That broad agreement became much less useful once legislators had to answer the next question: what law, exactly, are you willing to vote for? That was where political consensus evaporated.

Something similar may now be beginning with AI. Trump says guardrails may be appropriate. Johnson says guardrails are necessary. The AI companies say safety is important. Democrats say safety is important. Republicans say safety is important. China presumably does not want uncontrolled AI systems destroying Chinese infrastructure any more than Americans want them destroying American infrastructure. In the abstract, then, nearly everyone supports “AI safety.” The meaningful question is what happens when safety costs something.

What happens when a laboratory has a model ready to train and an independent evaluator says the experiment should wait six weeks? What happens when a company has spent several billion dollars preparing a training run and a federal regulator says it cannot proceed? What happens when OpenAI wants to continue while Anthropic wants to pause, or when Anthropic slows down while Google does not? What happens if the leading American companies agree to exercise restraint but intelligence officials announce that a Chinese laboratory may be six months away from a comparable breakthrough? At that point, AI safety stops being an aspiration and becomes a policy. Policies impose constraints, and constraints are precisely where political agreement tends to disappear.

Trump has already identified the argument that could become the most powerful weapon against mandatory AI safety regulation: China. It is an extremely potent argument because, unlike some political talking points, it describes a genuine problem. The United States and China are engaged in strategic competition over artificial intelligence. Frontier AI may eventually affect intelligence gathering, cyberwarfare, weapons development, biotechnology, economic productivity, and military command systems. No American president can casually ignore the possibility that the United States might voluntarily slow its development while China continues racing forward. Dario Amodei himself has acknowledged how difficult this problem is, because the strategic rewards for ignoring an international slowdown could be enormous.

Trump is therefore identifying a legitimate dilemma. The danger is that a legitimate dilemma can become a convenient excuse. If “China might get ahead” becomes sufficient reason to reject mandatory safety measures, then almost no safety measure will survive. The more powerful AI becomes, the more strategically important it becomes, and the more strategically important it becomes, the more dangerous it appears to slow down. Greater AI capability could perversely produce less willingness to regulate it.

Consider where that logic leads. Suppose a frontier system becomes dramatically better than humans at offensive cybersecurity. Should development be slowed while researchers determine whether it can be controlled? The answer could be that China might get ahead. Suppose a system becomes exceptionally capable at biological design. Again, China might get ahead. Suppose autonomous AI researchers become capable of meaningfully accelerating AI development itself. Once again, China might get ahead. If the response to every dangerous capability is identical, then “we cannot let China win” ceases to be an argument about balancing risks and becomes an argument that no amount of risk can ever justify slowing down. That is no longer ordinary risk management. It is the logic of an arms race.

This is where Sandy Hook becomes useful—not as an analogy between the technologies, but as an analogy between political processes. After Sandy Hook, one of the central arguments against additional gun restrictions was that regulation would impose costs on law-abiding Americans without preventing determined criminals from obtaining weapons. A criminal could still acquire a gun illegally, so regulation could not guarantee safety, and therefore additional rules might merely burden people who followed the law. It is possible to argue endlessly about the merits of that reasoning in firearm policy, but its structural resemblance to the emerging AI argument is striking.

China might ignore American rules. Open-source developers might ignore them. A rogue laboratory might ignore them. Other countries might ignore them. Therefore, the argument goes, responsible American frontier laboratories should not be constrained. In both debates, imperfect enforcement threatens to become an argument against enforcement itself. Unless a regulation can eliminate the danger everywhere, its inability to eliminate the danger everywhere is offered as a reason not to reduce it anywhere. That is an extraordinarily high standard to demand from public policy, and almost no regulation could survive it.

The comparison with gun politics becomes even more intriguing when one notices where it breaks down. After Sandy Hook, the firearms industry was not asking Congress to slow gun sales. The National Rifle Association was not demanding tougher federal regulation of itself. Gun manufacturers were not going on television saying that they were moving too quickly and needed Washington to establish a common framework preventing competitors from behaving irresponsibly. Yet something close to that is happening with artificial intelligence. The leaders of major frontier laboratories are publicly acknowledging that the competitive dynamics of their own industry may be unsafe. Several have endorsed slowing development enough for safety systems to catch up.

This creates an extraordinary political paradox. Imagine that after Sandy Hook, several of the largest firearms manufacturers had gone to Congress and said that the competitive structure of their industry made voluntary restraint difficult, that they feared where the market was heading, and that they wanted enforceable standards applying equally to all major companies. Then imagine that congressional leadership responded that these companies were free to restrain themselves whenever they wished. That hypothetical contains something close to the contradiction now emerging around frontier AI.

Johnson’s argument—that the laboratories can simply slow themselves down—is perfectly logical if each laboratory operates independently. But that is precisely what the laboratory leaders say is not happening. Their concern is a coordination problem. If Anthropic slows down while OpenAI accelerates, Anthropic loses. If OpenAI slows while Google accelerates, OpenAI loses. If all three exercise restraint while another competitor continues, the competitor gains. The problem is not necessarily that the executives personally want to race recklessly. The problem is that they inhabit a system that rewards whichever participant races fastest. Government exists, in part, to solve exactly these kinds of collective-action problems. A speed limit would be fairly useless if every driver were simply invited to choose whatever speed seemed responsible.

There is nevertheless an obvious attraction to the Trump-Johnson approach. It avoids bureaucracy and reduces the danger of placing enormously consequential technical decisions in the hands of legislators who may understand the technology poorly. It decreases the possibility that regulation will freeze today’s market leaders permanently into place and allows safety systems to evolve quickly rather than waiting for Congress. It also avoids creating a powerful government licensing regime over one of the most important technologies of the century. Those are serious concerns and deserve serious consideration.

One should also be wary when enormous corporations ask to be regulated. Large incumbents sometimes favor regulations they are uniquely capable of satisfying. A federal frontier-model licensing regime requiring hundreds of millions of dollars in compliance infrastructure could improve public safety while conveniently making it much harder for smaller competitors to challenge OpenAI, Anthropic, Google, or other existing giants. There are legitimate libertarian, economic, and technological arguments for moving cautiously. But “move cautiously when regulating” is different from “leave the decision to slow down entirely to the companies engaged in the race.” The former is prudence. The latter is itself a major policy choice. It means accepting voluntary self-regulation as the principal mechanism protecting the public from technologies that the developers themselves increasingly describe as potentially catastrophic.

Sandy Hook also revealed another unpleasant feature of democratic politics: public attention has a half-life. Immediately after catastrophe, almost everything seems possible. Four months later, considerably less seems possible. A year later, politicians have moved on. The underlying danger may not have changed at all. What changed was attention.

AI may now be experiencing an unusual inversion of that pattern. There has not yet been a Sandy Hook-scale AI catastrophe. Instead, capability advances, unsettling incidents, resignations by researchers, and increasingly serious warnings from laboratory leaders have created something resembling a pre-crisis policy window. That is rare. Normally governments receive warnings, postpone difficult decisions, and eventually regulate after something terrible happens. Here, at least in theory, we have an opportunity to attempt the reverse.

We can establish institutions before the disaster. We can determine dangerous capability thresholds before one is crossed accidentally. We can create reporting requirements before the incident everyone later wishes had been reported. We can establish independent evaluators before corporate and national competition becomes too intense to permit them. We can begin negotiating international agreements before the United States and China both possess systems whose development neither side can adequately monitor or verify.

Or we can wait. Waiting has enormous political advantages. Nobody gets blamed for slowing innovation. Nobody gets blamed for losing to China. Nobody has to decide precisely what constitutes a dangerous frontier model. Nobody has to create a new regulator. Nobody has to explain why a multibillion-dollar training run was postponed. If nothing goes wrong, waiting looks wise. The trouble is that this asymmetry works only until something does.

The most sobering aspect of the Sandy Hook aftermath was not that Americans disagreed about guns. Americans had disagreed about guns for decades. It was that even a moment of extraordinary agreement could evaporate without producing the federal legislation many people believed was inevitable. Twenty children had been murdered. The president was personally lobbying senators. Victims’ families were walking the halls of Congress. More than eight in ten Americans supported expanded background checks. A bipartisan compromise received a majority vote in the United States Senate. And still the legislative effort collapsed.

That is what political drift looks like. It rarely begins with a dramatic announcement that nothing will ever be done. It sounds much more reasonable than that. We need more study. We need consensus. Industry should lead. We cannot move hastily. The proposal is imperfect. The other side will not cooperate. Now is not the right time. Eventually an urgent problem becomes an ordinary problem, and ordinary problems can remain unsolved for decades.

This is the danger in the comments from Trump and Johnson. Neither man is saying that AI safety is irrelevant. In some respects that would actually be politically easier to confront. Instead, both are accepting the principle while resisting the mechanism. Trump says there can be guardrails, but emphasizes winning the AI race and dismisses some warnings as scenarios that will not happen. Johnson says safety measures are necessary, but rejects an emergency congressional moratorium and suggests the industry itself should assume the leading role. That may sound like a temporary disagreement over methods. History suggests it can become something considerably more consequential.

After Sandy Hook, the question rapidly shifted from whether the country should do something to whether Congress should enact this particular restriction, this particular background-check requirement, or this particular compromise. Once the debate reached that second stage, the coalition supporting action fractured. Artificial intelligence may now be entering exactly the same political phase. Almost everyone supports AI safety. The argument is becoming whether safety should ever be allowed to compel delay; whether independent evaluations should merely advise companies or possess actual authority; whether laboratories should disclose dangerous capabilities voluntarily or be legally required to do so; whether a frontier laboratory should ever need government permission before deploying an exceptionally capable model.

Ultimately, the question is whether Congress should ever be able to tell an AI company to stop—not forever, not because artificial intelligence is evil, and not because technological progress should end, but because human beings may occasionally need several months to understand what they have built before building something substantially more powerful. If the answer to that question is always no, then talk of “guardrails” means considerably less than it appears to mean. A guardrail that cannot prevent the vehicle from continuing forward under any circumstances is largely decorative.

No analogy between Sandy Hook and artificial intelligence should obscure the most important difference between them. Sandy Hook was already a tragedy. Twenty children and six adults were already dead before the political window opened. Nothing Congress did afterward could save them. AI safety presents us with the possibility—still, for the moment—of acting before an equivalent catalytic event.

Perhaps the catastrophic forecasts will prove wrong. Perhaps artificial general intelligence will arrive gradually. Perhaps alignment techniques will work. Perhaps the current generation of researchers is overestimating the risks of autonomy, cyber capabilities, biological design, or recursive improvement. That would be an excellent outcome. Good safety regulations sometimes look unnecessary in retrospect precisely because the catastrophe they were designed to prevent never occurs.

But American politics has a dangerous tendency to demand catastrophe as proof. After industrial accidents, we improve industrial safety. After financial crises, we strengthen financial regulation. After terrorist attacks, we close security vulnerabilities. After school shootings, we briefly debate gun laws, and then the urgency begins to dissipate. The extraordinary thing about the current AI moment is that some of the people closest to the technology are attempting to reverse that sequence. They are effectively asking whether society can argue about the regulations before discovering, through disaster, why those regulations were necessary.

Trump and Johnson are not wrong to worry about China. They are not wrong to worry about governmental overreach, incompetent regulation, or rules that inadvertently entrench today’s AI giants. But those objections need to become components of an AI-safety policy rather than substitutes for one. A serious response to the China problem might involve treaties, verification regimes, compute monitoring, narrowly targeted capability thresholds, and agreements between rival laboratories. A serious response to regulatory capture might involve independent oversight and protections for smaller competitors. These are difficult problems, but difficulty is not an argument for doing nothing.

There is another lesson buried in the aftermath of Sandy Hook. A country can recognize a danger. Its citizens can demand action. Experts can propose solutions. Politicians can express sympathy. Majorities can support reform. Everyone can insist that something must be done. And then, slowly and almost imperceptibly, the moment can pass.

That may be what is beginning to happen with artificial intelligence. The laboratories are asking for time. Researchers are asking for safeguards. Some political leaders are asking for binding action. Meanwhile, the president and the speaker of the House are signaling that America cannot afford to take its foot too far off the accelerator. Perhaps they will prove correct. Perhaps history will look back on the current safety fears as exaggerated.

But before accepting that gamble, we should remember how often American politics has followed the same sequence: a warning, a moment of clarity, a burst of political possibility, a thousand plausible reasons to wait, and eventually a tragedy that causes everyone to ask why nobody acted when there was still time.

After Sandy Hook, the tragedy came first and the opportunity for prevention came too late for twenty children and six educators. With artificial intelligence, for now at least, we have been given the order in reverse.

We should be very careful about throwing that advantage away.

There is an uncomfortable comparison beginning to suggest itself in the debate over artificial intelligence.

In December 2012, a gunman entered Sandy Hook Elementary School in Newtown, Connecticut, and murdered twenty children and six educators. The horror of the crime produced something that had become increasingly rare in American politics: a moment in which a large portion of the country seemed to agree that an intolerable risk demanded a political response. The Obama administration proposed new gun-safety measures. Families of the victims went to Washington. Senators negotiated a bipartisan compromise. Polls showed extraordinary public support for expanded background checks.

And then, at the federal level, very little happened.

Today, the United States may be approaching another such policy window, this time involving artificial intelligence. The analogy should not be pushed too far. AI systems are not firearms. The dangers posed by highly advanced AI are different in kind, probability and immediacy from gun violence, whose human costs are tragically well established. Nor should Sandy Hook be reduced to a convenient metaphor for some other political cause.

But the political comparison is worth making because Sandy Hook demonstrated something disturbing about American institutions: even overwhelming public alarm, elite attention and apparent agreement that a danger is real do not necessarily produce preventive legislation. Sometimes a society recognizes a risk, debates it intensely, develops plausible safeguards—and still fails to act.

That is precisely the possibility now confronting AI policy.

The change in the AI debate during the summer and early fall of 2026 has been extraordinary. For several years, arguments about existential AI risk could be dismissed as speculative discussions among researchers, science-fiction-minded technologists and a relatively small community of AI-safety advocates. That characterization has become increasingly difficult to maintain.

Anthropic CEO Dario Amodei has now explicitly called for the industry to “pace the frontier.” His proposal does not call for abandoning artificial intelligence. Instead, he argues that the development of increasingly powerful systems should be slowed enough for safety mechanisms to keep pace. Among his proposals is an unusually intrusive form of independent oversight: frontier companies would give embedded third-party evaluators ongoing, employee-like access so that outsiders could examine safety practices, training procedures and dangerous capabilities rather than merely testing the polished model shortly before release. He also calls for broader industry standards and eventually international coordination.

OpenAI has moved in much the same direction. On September 9, the company publicly called for mandatory national AI-safety requirements, specifically saying that voluntary commitments are no longer sufficient. Its proposal includes capability-based federal regulation, independent safety assessments, cybersecurity requirements and incident reporting. The company is simultaneously supporting several California bills involving independent safety evaluations, auditor standards, biological risks and protections for young users.

This shift did not occur in a vacuum. OpenAI recently designated Astra as reaching its “Critical” cybersecurity capability threshold, meaning that, with appropriate tools and access, the system can discover previously unknown vulnerabilities and develop exploits against well-protected systems without a human directing every individual step. OpenAI says it delayed parts of Astra’s development and strengthened safeguards before release. The company had already temporarily slowed some frontier work following the OpenAI-Hugging Face incident, imposing stronger workload and network isolation on higher-risk research.

Political demands are escalating as well. On September 3, Senator Bernie Sanders and Representative Greg Casar announced legislation that would go far beyond ordinary technology regulation. Their proposed Ban Artificial Superintelligence Act would prohibit the development and deployment of systems classified as superintelligent and temporarily pause advanced AI development until a federal regulator establishes safety standards. It would also direct the United States to pursue international agreements intended to prevent an uncontrolled international race toward superintelligence.

Meanwhile, Republican Senator Josh Hawley has opened an investigation into OpenAI’s recent agent behavior and the Hugging Face incident. The significance of this should not be overlooked. AI risk is increasingly producing concern from politicians who agree on very little else. The specific remedies differ dramatically, but unease about allowing a handful of private companies to develop systems of potentially enormous power with limited external oversight is no longer confined to one ideological faction.

There is even the beginning of something remarkably unusual in Silicon Valley: competing AI companies discussing whether they should deliberately slow themselves down. That creates its own problems. OpenAI has reportedly sought clarification about whether coordination among rival companies to slow frontier development might violate antitrust law. What sounds like responsible cooperation from an AI-safety perspective can look remarkably like competitors agreeing to restrict production from the perspective of traditional antitrust law.

This is where the shadow of Sandy Hook becomes relevant.

Following the December 2012 massacre, the political ingredients for action appeared to be present. President Obama announced twenty-three executive actions aimed at reducing gun violence and urged Congress to go further. The centerpiece of the congressional effort became the bipartisan Manchin-Toomey amendment, which would have expanded background checks to additional commercial gun sales.

It was hardly a radical proposal. In early 2013, polling repeatedly found extraordinarily high support for expanded background checks. Pew later found that 81 percent of Americans favored subjecting private sales and gun-show sales to background checks; even among gun owners, support was substantial.

The amendment nevertheless failed.

On April 17, 2013, Manchin-Toomey received 54 votes in the United States Senate and 46 votes against. Because 60 votes were required under the Senate procedure being used, a proposal supported by a majority of senators—and by an overwhelming majority of the public—did not advance.

It is important to be precise about what happened afterward. Sandy Hook did not produce literally no policy response. Connecticut enacted substantial new gun restrictions, including expanded background checks, restrictions on large-capacity magazines and an expanded assault-weapons ban. Other states also tightened their laws, while still others moved in the opposite direction. The Obama administration implemented executive actions where it believed it had authority to do so.

What failed was the larger attempt to convert a national moment of horror and unusually broad public agreement into major federal legislation.

Congress would not enact another major federal gun-safety package until the Bipartisan Safer Communities Act in June 2022—nearly a decade later, after the country had experienced many more mass shootings, including the killings in Buffalo and Uvalde. That law enhanced background checks for buyers under twenty-one, addressed some domestic-violence restrictions and funded mental-health and school-safety programs, among other provisions. It was significant precisely because meaningful federal action had been so difficult for so long.

The lesson is not simply that “Congress is dysfunctional.” It is more specific and more troubling.

Public policy has windows of opportunity. A shocking event, technological breakthrough or sudden shift in public consciousness can temporarily change what politicians consider possible. For a short time, previously abstract risks become tangible. Journalists pay attention. Citizens demand answers. Politicians who normally avoid the subject feel pressure to take positions. Opposing interest groups have not yet fully reorganized around the new political landscape.

Then the window begins to close.

The immediate fear fades. Other stories dominate the news. Proposed regulations acquire details, and details create opponents. Economic interests calculate what they might lose. Politicians discover that vague support for “doing something” fragments when the conversation turns to a particular bill. Arguments that sounded inappropriate immediately after a crisis become politically effective again.

Sandy Hook demonstrated the difference between salience and power. An issue can command overwhelming attention without its supporters possessing sufficient political power to overcome concentrated opposition.

That distinction ought to worry people concerned about AI.

At the moment, AI safety has extraordinary salience. Frontier researchers are resigning and issuing warnings. CEOs are publicly discussing slowing development. Companies are calling for government regulation of their own industry. Politicians from different ideological camps are demanding investigations or legislation. Recent incidents have provided concrete examples around which previously abstract safety concerns can coalesce.

But almost all of the incentives that produced the AI race remain intact.

Billions of dollars are at stake. Companies fear losing market share. Researchers fear that another laboratory will reach the next capability threshold first. Investors have enormous sums committed to infrastructure and model development. Governments increasingly view artificial intelligence not simply as an industry but as a strategic national asset.

And the most powerful argument against slowing down is already obvious: China.

President Donald Trump has resisted broad calls to slow American AI development, emphasizing the danger of surrendering technological leadership to China. That concern is not frivolous. Unlike domestic gun regulation, frontier AI policy really does involve an international strategic competition. A unilateral American slowdown could conceivably reduce one category of risk while increasing another.

But this argument can also become the AI equivalent of an all-purpose veto.

Every proposed safeguard can be answered with the claim that China will not adopt it. Every delay can be described as surrendering the technological race. Every safety requirement can be portrayed as a burden on American innovation. If that logic becomes absolute, then there is effectively no capability threshold dangerous enough to justify restraint, because greater danger would merely make winning the race seem more important.

That is structurally similar to what happened in the gun debate. The arguments are not the same, but the political mechanism can be. A broadly popular principle—“dangerous technology should have reasonable safeguards”—collides with a much more intensely motivated constituency for whom the regulation carries concentrated ideological or economic costs.

There is another similarity. Both debates contain a powerful form of fatalism.

After mass shootings, opponents of gun restrictions have often argued that criminals will obtain guns regardless of regulation, that another intervention would not have prevented the particular shooting being discussed, or that determined attackers will simply find another method. These arguments can transform uncertainty about whether a regulation will prevent every tragedy into an argument against preventing any tragedies.

AI policy risks developing its own version: if America slows down, somebody else will build it; if one company refuses, another company will proceed; if regulated laboratories stop, open-source developers will continue; if democratic countries impose restrictions, authoritarian governments will ignore them.

There is truth in every one of those objections. Taken together, however, they can produce paralysis. The inability to guarantee universal compliance becomes a reason not to reduce risk at all.

Yet the analogy also has important limits, and those limits may actually make AI regulation more achievable.

Firearms are deeply distributed throughout American society. Hundreds of millions are already in private hands. Gun ownership is tied to constitutional law, regional culture, personal identity and an enormous commercial ecosystem. Any significant regulation therefore collides with millions of individual stakeholders as well as organized political groups.

Frontier AI is currently far more concentrated.

Training the most capable systems requires enormous amounts of capital, specialized chips, data-center infrastructure, electricity and technical expertise. The number of organizations capable of operating at the frontier remains relatively small. That gives policymakers potential regulatory chokepoints that do not exist in the same way with firearms.

It is much easier to inspect five or ten frontier laboratories than hundreds of millions of gun owners.

There is another profound difference. The major firearms industry and gun-rights organizations did not emerge from Sandy Hook asking Congress to regulate them aggressively. In the current AI debate, some of the companies standing to be regulated are themselves asking for regulation.

That should be welcomed, but not accepted uncritically.

Large incumbents often prefer regulations they can afford to comply with, particularly if those regulations create barriers that smaller competitors cannot. A licensing or evaluation system designed around the resources of OpenAI, Anthropic or Google DeepMind might genuinely improve safety while also conveniently solidifying their market positions. Critics are therefore justified in asking whether industry-supported AI regulation protects humanity, protects incumbent companies, or does some combination of both.

That is an argument for designing regulation carefully, not for abandoning regulation.

The biggest difference, however, concerns evidence.

Gun violence does not require a forecast. Its consequences are measurable. Sandy Hook happened. Uvalde happened. Buffalo happened. Tens of thousands of Americans die from firearms in a typical year when suicides, homicides and accidents are counted together. One can argue endlessly about which policies would reduce those deaths, but the underlying harm is not hypothetical.

The most extreme AI scenarios remain prospective. Nobody can demonstrate that a future superintelligence will escape human control, seize infrastructure, develop biological weapons or destroy civilization. Nobody can assign a reliable probability to those outcomes.

That uncertainty can become another excuse for inaction.

It should instead force policymakers to confront a basic principle of risk management: when consequences become sufficiently large, uncertainty about probability does not eliminate the need for precautions. We do not require engineers to prove that a bridge will collapse before inspecting it. We do not require nuclear regulators to demonstrate that a particular reactor will melt down before establishing containment requirements. Safety systems exist precisely because waiting for conclusive empirical evidence sometimes means waiting for the accident.

And that may be the most important lesson Sandy Hook offers the artificial-intelligence debate.

America has an unfortunate political habit of treating catastrophe as the price of admission for serious regulation. We wait for the bridge to collapse, the market to crash, the terrorist attack to occur, the school to be attacked or the industrial system to fail. Then, for a brief period, everyone asks why obvious vulnerabilities were tolerated.

With artificial intelligence, that sequence may be extraordinarily dangerous.

If the strongest claims made by AI-safety researchers are exaggerated, stringent safeguards might cost us money, slow technological progress or postpone beneficial applications. Those are real costs and should be acknowledged rather than waved away.

But if even some of the stronger warnings are substantially correct, waiting for an unmistakable AI catastrophe before creating serious oversight could be a disastrous strategy. An accident involving autonomous cyber capabilities, biological design, critical infrastructure or recursively improving systems might not provide the clean second chance that policymakers assume every technology will offer.

The goal therefore should not be to “ban AI” or freeze technological civilization in place. Amodei’s phrase—pace the frontier—is useful because it describes something more modest and more defensible: capability should not advance faster than our ability to understand, monitor and control it.

Independent evaluators should have meaningful access rather than ceremonial access. Frontier developers should face mandatory incident-reporting rules. Dangerous capability thresholds should trigger stronger security requirements automatically rather than depending on corporate discretion. Government should possess enough technical expertise to evaluate claims made by the companies it regulates. And international negotiations on the most dangerous capabilities should begin while those capabilities remain concentrated among a relatively small number of actors.

None of those policies guarantees safety. Neither would a temporary slowdown. Neither would a treaty with China. Complex risks rarely have single solutions.

But the alternative cannot be that because perfect safety is impossible, deliberate safety is unnecessary.

Sandy Hook should have taught the United States something about the difference between recognizing danger and governing it. Twenty children could be murdered in their elementary school. Public support for a concrete reform could exceed 80 percent. A bipartisan bill could receive a majority of votes in the Senate. A president could expend substantial political capital on the issue.

And the policy window could still close.

The AI-safety debate may now be entering such a window.

There is, however, one enormous difference. With Sandy Hook, the catastrophe that opened the window had already happened.

With artificial intelligence, we still have the extraordinary luxury of arguing about what to do before the event we might someday name the legislation after.

It would be a remarkable failure of imagination if we decided that we needed the tragedy first.