Anthropic CEO Proposes AI Policy Framework Amid Rapid Technological Growth

Recently, Anthropic, a leading AI giant, CEO Dario Amodei published a 10,000-word article that went viral across the internet: “Policy in the Age of Artificial Intelligence.” The full text meticulously outlines his observations and offers systematic recommendations spanning regulation to geopolitics, organized around five policy areas.

(Original link: https://darioamodei.com/post/policy-on-the-ai-exponential, full Chinese version provided below.)

It must be acknowledged that, in the Silicon Valley circle where everyone talks endlessly about “changing the world” and only knows how to make empty promises, Dario is a rare breath of fresh air.

He was extremely candid, even to the point of “exposing his own kind,” tearing away the veneer of false harmony in the tech industry.

Insiders often attribute public concerns about AI to poor public relations, but Dario bluntly counters: people are concerned because the risks are real—it’s not a public relations issue at all.

He also directly acknowledged the truth that tech giants are极力回避: AI is very likely to trigger a severe and prolonged wave of unemployment, trapping the economy in an extreme state of "super growth and super inequality."

But if we follow the remedy proposed by this industry frontline "whistleblower," we uncover a thought-provoking structural paradox:

This responsible executive, while calling for regulations to be imposed on the industry, is objectively facilitating an unprecedented transfer of power. Silicon Valley elites are rewriting the rules of human society.

Faced with AI evolving at lightning speed, Dario is alarmed and urges the government to regulate AI as strictly as it does airplanes and pharmaceuticals.

For example, establish a mechanism similar to the Federal Aviation Administration (FAA) that requires frontier models meeting certain computational power thresholds to pass mandatory third-party security tests before being released.

The intention is undoubtedly for human safety.

From the objective laws of commercial evolution, the airline industry is highly concentrated and oligopolistic precisely because compliance costs are astronomically high.

Once the AI industry truly becomes "FAA-ized," high review fees, regular red teaming, and penetration testing will directly solidify into an impenetrable "Wall of Sighs."

Regulation can sometimes be a moat for market leaders—something startups and open-source communities simply cannot afford.

The end result was that the giants, under the guise of "love" for the safety of all humanity, objectively used public power to legally clear the field and enshrined their oligarchic status in law.

This subtle logic also exists in the fields of pharmaceuticals and geopolitics.

Dario calls on traditional institutions like the FDA to relax approval requirements for AI-driven drug development and accept AI simulations as replacements for lengthy clinical trials. While the intent is commendable, it undoubtedly clears the way for AI giants with the strongest computational power to deliver a paradigm-shifting blow to the trillion-dollar pharmaceutical industry.

At the same time, he envisioned an "Alliance of Democratic Nations AI" that would share chips internally and strictly guard against external access.

If this alliance is formed, who provides the technology? Who is excluded? Who is opposed? Who benefits throughout this process?

To maintain their technological edge, tech giants are not only technology providers but also naturally become the architects of the new rules of the game, and may even transform into a new type of "digital military-industrial complex" that承接 massive defense budgets.

If we trace the subtle clues of history, all of this was actually foretold.

Historically, when a company becomes large enough, it often "captures the state or adopts quasi-state characteristics." Dario himself mentions a very precise analogy in the text: the British East India Company.

The East India Company began as a group of merchants engaged in the spice trade, but to protect their shipping routes and manage their vast transnational interests, they gradually began recruiting armies, establishing courts, and issuing currency, eventually becoming a "quasi-state" that ruled the subcontinent.

Today’s Silicon Valley giants are following the exact same path.

Back then, the medium was gunboats and cannons; today, it is computing power.

Dario envisioned a stunning future: if AI continues its exponential evolution, it will become a "nation of geniuses inside data centers"—AI would be a gathering of geniuses from every field, and possessing powerful AI would be like having a nation’s worth of geniuses serving you. Thus, when an army equipped with powerful AI confronts one without AI, the gap is like that between World War II Marines and medieval swordsmen.

When tech giants hold super technologies capable of triggering financial collapses, creating biological weapons, or even determining global geopolitical landscapes, traditional state institutions are struggling to effectively constrain them.

The giants have already attempted to establish their own security standards, test their own models, and even devise international alliance proposals.

This is not because they are a group of evil schemers, but because, after technological development reached this scale, the power vacuum and the complexity of human nature naturally led them down a similar path.

In an era where hash power determines everything, individual strength seems infinitely compressed.

Dario ruthlessly revealed the truth: AI is not only replacing physical labor, but also humanity's most prized cognitive abilities.

When the logical reasoning and strategic planning we’ve painstakingly honed in our careers seem like child’s play against AI, the “nation that brings together all geniuses,” how should we respond?

The remedy offered by the giants is macroeconomic support such as universal basic income (UBI)—the state provides financial support to everyone, even if they don’t work.

But the job is safe—so who fills the big hole in your heart?

Perhaps, as Dario’s metaphor in the text suggests: even though machines have long surpassed humans in chess and Go, people still devote their lives to the board and are still held in high regard.

Because machines calculate cold, optimal solutions, while humans seek the vibrant experience itself.

Those qualities that cannot be standardized will become the most scarce resources in the future.

In a subplot of The Lord of the Rings, two hobbits attempt to awaken Treebeard—a wise but slow-moving sentient tree-ent—to defend his forest against an army that is cutting it down. The problem is that Treebeard’s pace is vastly different from that of the hobbits. It takes him an entire day just to greet another tree, making it nearly impossible for him and his kind to act quickly enough.

The intersection between artificial intelligence (AI) and our political systems feels a bit like Hobbits and Treebeard. AI is advancing at a lightning pace—within just four years, AI models have progressed from being unable to write a single line of coherent code to generating most of the code at major AI companies. Similar breakthroughs have occurred in biology, physics, mathematics, finance, law, translation, and many other fields. The scaling laws of AI, which predict that general cognitive abilities will grow exponentially with increased computational power, are now supported by over a decade of empirical evidence. If these scaling laws continue for just another one or two years, we are likely to achieve what I call “Powerful AI”—or a “nation of geniuses inside a data center.”

In contrast, policy—especially legislation—progresses very slowly. This is often for good reason: governments wield extremely significant power, and it is usually best not to exercise it hastily. But this misalignment in time scales remains deeply frustrating: while it may take years for Congress to act, AI could evolve from an interesting toy into a true nation of geniuses.

Over the past several years, since AI became a major business technology, those of us hoping to handle it responsibly have faced a dilemma. We can clearly see the trajectory of this exponential growth: we strongly suspect that within a few years, AI will become one of the rare technologies that fundamentally reshape the entire policy landscape—much like nuclear weapons reshaped geopolitics, or the Industrial Revolution fundamentally transformed every economic and social issue. Yet, to those focused only on what AI can do right now, it appears more like a fairly ordinary technology—perhaps similar to the latest consumer app or cryptocurrency. It is difficult to convince most policymakers and corporations that anything beyond a laissez-faire approach is meaningful. Fairly stated, the radical impacts of AI have not yet materialized, and we do not yet know precisely what form they might take, making it hard to design the right policies even when there is a willingness to act.

Given the constraints imposed by this situation, many safety advocates—including Anthropic—have so far focused on advocating for policy actions that preserve options, tee up rapid future responses, or provide the world with better insight into what’s coming down the pike—such as transparency legislation, export controls on chips, and collecting data on AI’s impact on the workforce. These aren’t enough, but they feel like everything that can realistically be done right now.

However, over the past few months, evidence of AI’s incredible power and its associated risks has become undeniable. Perhaps the most emblematic example is the Claude Mythos Preview, along with the discovery that frontier models pose a very real threat to cybersecurity, creating the potential to disrupt financial systems, critical infrastructure, and national security. The Mythos Preview has scrambled the global cybersecurity landscape. But its broader significance lies in the undeniable proof that AI models are now tools of global and strategic national importance. The cyber risks posed by Mythos-level models will not be the last we must confront. I believe biological risks may soon follow, and serious risks of AI autonomy may come right after (Note 1).

We now collectively need to activate slow and unwieldy policy institutions to address risks and opportunities that will compound at an astonishing rate from this point forward. Many policymakers are showing increasingly open attitudes toward action, and it is encouraging to see our peers come around to the same positions we have advocated for over the past several years. This is positive, but I am concerned that these early efforts are already at least a year behind the rapid progress of AI. This article seeks to close that gap: clarifying where exponential growth stands today and what collective action is needed to meet the moment.

I will focus on five enduring policy areas that need to be reimagined in the AI world: regulation and public safety, macroeconomic and tax policy, scientific innovation, the balance of power between the state and society, and geopolitics. I will primarily address these from a U.S. policy perspective, as Anthropic is a U.S.-based company, but most of the recommendations I offer are also relevant to other parts of the world.

Along with this article, Anthropic will release a legislative proposal on testing frontier models and a policy framework on job displacement, for which we intend to provide substantial financial support. We plan to do more in the future, but we view these as the first steps demonstrating our seriousness.

1. Regulation and Public Safety

Every new technology or product has both beneficial and harmful applications, creating a dilemma between innovation and safety. Regulating products can reduce the likelihood of harm and play a vital role in improving lives around the world, but it may also directly diminish their benefits and indirectly stifle innovation. There is also a Hayekian perspective that regulators typically lack the information needed to make sound decisions about complex economic trade-offs, making regulation often both ineffective and burdensome. A related concept is the Collingridge dilemma, which states that the impacts of a technology are usually difficult to predict until it is too late to manage them easily.

During 2023–2024, these dynamics loomed large in the field of AI. For Anthropic, it was clear that AI in the future could potentially have the capability to manufacture biological weapons threatening millions of people, or that its autonomous behavior, under extreme circumstances, might even threaten humanity itself. What was less clear was exactly how these risks would manifest, how best to test and mitigate them, and how they would play out in practice. As a result, preemptive legislation carries a high risk of ultimately becoming ineffective—creating meaningless or low-value compliance requirements while missing the most critical sources of actual risk (Note 2).

Ultimately, we concluded that transparency was the right approach at the time. Developers of AI models should be required to disclose their security protocols and the tests they conduct on their models, and to report any critical security incidents, so that the public and the scientific community can better understand these risks as they emerge. When and if risks become more clearly defined and their contours more apparent, the evidence gained through transparency can be used to craft smart legislation that precisely targets the most concerning risks. Therefore, in 2025, Anthropic supported transparency legislation, helping to pass California’s SB 53, New York’s RAISE, and Illinois’ SB 315 (which passed in early 2026), and advocated for federal-level transparency standards.

However, the risks are now clearly upon us. It is time to move beyond transparency and implement more serious and binding regulation of AI. I believe that, at least during this current phase of exponential growth, the best analogy is with automobiles, airplanes, or pharmaceuticals—powerful technologies essential to the modern economy, but capable of causing mass fatalities if improperly designed or operated. Therefore, I believe we should model AI regulation after agencies such as the Federal Aviation Administration (FAA). Frontier AI models, like airplanes, should be required to undergo technical testing and auditing, and if they fail to meet high safety standards, their release should be blocked or revoked, treated as a threat to public safety. I am encouraged to see the Trump administration’s executive order moving in the direction of increasing government involvement in AI, although Anthropic’s proposal calls for even more robust action. Our proposal includes the following elements:

• Models exceeding a certain computational threshold must undergo mandatory third-party testing to assess their risk levels in four specific areas: cybersecurity, biological weapons, loss of control over AI systems, and automated research that may accelerate these other risks.
• If a third-party assessment determines that the model poses unacceptable risks, the government should have the authority to prevent or deter its deployment. This authority must be limited to the four specific risks outlined above and must include safeguards against political bias or arbitrary decisions.
• Third-party evaluations can be conducted by government agencies (similar to the FAA) or by a group of private organizations authorized and inspected by the government to assess models according to certain standards (a “regulatory market” approach).
• AI companies developing advanced AI models must have strong security standards to protect their model weights, conduct regular red teaming and penetration testing, and collaborate with governments to defend against major threat actors.
• Security incidents in these four key areas must be reported promptly.

There may come a day—perhaps soon—when the most powerful AI systems no longer resemble airplanes or cars, but instead resemble weaponizable nuclear material; at that point, the threat they pose to humanity will go beyond merely public safety, and we will need to go beyond this step. If this happens, we may need to adopt regulatory measures more radical than those I’ve outlined above (note 3). However, just as it was difficult in 2024 to implement the measures I’ve suggested, I believe we should not get ahead of ourselves (idiom meaning to act too soon or prematurely). We should design policies to address the dangers emerging today, while laying the groundwork to respond more swiftly when new threats arise.

2. Macroeconomics and Tax Policy

For a long time, governments around the world have faced the challenge of encouraging economic growth while also providing essential public services and ensuring support for the most vulnerable populations. An important—and generally accurate—assumption underlying these debates is that economic growth is fragile and difficult to achieve—that while reducing inequality may bring significant benefits, it must be weighed against the economic headwinds of higher taxes or increased deficits.

I suspect that powerful AI could disrupt this assumption. If AI acquires far superior capabilities to humans in performing most cognitive tasks, it logically follows that it could drive extremely rapid and strong economic growth by accelerating science, technology, and operational efficiency. AI’s iterative ability to build better AI could further amplify this growth. But for exactly the same reasons, AI might also serve as a more widespread economic substitute for human cognitive abilities than previous technologies, and do so more rapidly. Therefore, there is good reason to believe that AI could cause much larger—and potentially more persistent—disruptions to labor markets than previous technologies. The risk we face is ending up in a world where the dials of economic trade-offs are stuck on supergrowth and superinequality, and may be difficult to dislodge. In such a world, the key challenge will no longer be incentivizing growth, but finding ways to ensure everyone shares in the benefits.

Among the topics discussed in this article, macroeconomic factors and persistent labor substitution have arguably attracted the most public attention and the most misconceptions, so I want to be very clear about two points.

First, persistent job displacement is undesirable and dangerous, and we should do everything possible to minimize or prevent it, rather than enable it. I have warned about job displacement in interviews and articles because I want policymakers and the private sector to have the best possible chance to adapt and respond—not because I am trying to be a “doomsayer.” As a company, Anthropic always strives to work with customers to find creative new use cases and new revenue streams that enable them to do more with their existing workforce, rather than focusing solely on cost savings (which often means layoffs). As these systems advance, we continually seek to imagine new interaction paradigms that allow humans participating in collaboration to play as active a role as possible within AI systems. More broadly, it is valuable for the entire world to experiment with as many new ways of using AI as possible, because this is how society discovers new potential job configurations. I do believe AI will generate vast new economic opportunities. I have predicted that AI will enable a single individual to build a billion-dollar company, and we have already seen teams of just a few people build businesses with hundreds of millions in revenue. Yet at the same time, we must acknowledge that despite all our efforts, AI is still very likely to cause significant persistent unemployment—possibly an inherent feature of the technology and its ability to broadly replicate human cognition (note 4).

Second, any response to job displacement caused by AI must address both the need to provide economic security for everyone and the deeper human need to find meaning, purpose, and agency. The latter is ultimately more important, as it hinges on fundamental questions about how society should be organized, what people should strive for, and what constitutes a good life. In fact, I am deeply optimistic that even in a world where AI surpasses humans in every domain, people can still lead lives of profound purpose and strive to create awe-inspiring and beautiful things (note 5). But this is a challenge for society as a whole to solve—not one that policy alone can fix. Policy can play its most vital role by slowing the pace of job displacement and providing economic security to those who may be affected, thereby buying us the time needed to undertake this deeper work.

In this spirit, some key policy interventions that may be helpful include:

• Measure and track. It’s easy to dismiss mere data collection and analysis as insufficient to address the scale of the problem, but without accurately measuring what is actually happening, we are unlikely to craft effective policies. Anthropic has been operating an economic index on how people use Claude for nearly a year and a half, but governments can access types of data we cannot and can significantly expand their economic statistics to more closely track job displacement caused by AI.
• Incentives to promote employment. Broad policies to encourage employment can help slow or reduce job losses, including: wage insurance programs that compensate workers when they must take lower-paying jobs (note 6), tax incentives to encourage employers to retain staff, workforce training grants, or infrastructure that facilitates better matching between employers and employees to accelerate labor market adaptation. While the specifics of which interventions are most effective will depend on the type of labor displacement brought about by AI, we should readily accept the associated costs and market inefficiencies, especially since they may be offset by productivity gains driven by AI.
• Long-term macroeconomic support. If AI-driven labor substitution eventually becomes massive and permanently reduces demand for labor, it may be necessary to move beyond incentive programs and provide long-term income support to a large portion of the workforce. Mechanisms such as universal basic income could be funded through taxes on relevant companies or higher capital gains taxes. Universal capital accounts offer another tool. Broadly speaking, rapid economic growth should create a tax base for shared prosperity.

One common focus I did not mention in economic concerns about AI is data centers, particularly their potential to drive up energy prices. My view is that AI companies should bear the cost of rate increases themselves—and Anthropic has already committed to doing so—but I believe public hostility toward data centers is largely a symbol or outlet for broader economic anxieties about AI. It is crucial that we engage in direct societal dialogue about these wider economic issues and develop compelling, concrete solutions to address them; otherwise, they will likely manifest indirectly, as they have with data centers.

3. Accelerate the positive impact of AI

Just as we must strive to balance innovation and safety in AI itself, we must also strive to achieve a similar balance in technologies that AI may accelerate—such as biomedicine, energy, or materials science. Yet while AI itself may bring entirely new challenges that emerge rapidly and for which we have no prior experience, other fields accelerated by AI may face a distinctly different problem: regulatory systems designed for slower innovation cycles are unprepared for the flood of new products and advancements that AI will enable. AI may also make these downstream technologies safer and more predictable, thereby challenging the skeptical assumptions held by regulatory bodies such as the FDA.

Therefore, for AI’s downstream applications—unlike AI itself—I am more concerned that regulators will slow progress (because they cannot keep pace with accelerating change) than that they fail to address significant risks. We least want to see AI’s benefits delayed while its risks grow ever closer, so it is crucial to act on this issue as soon as possible.

This issue and its solutions will manifest in different ways across science, business, and technology, so I will focus on a representative area: biomedical innovation. This is both because it is likely to be the source of AI’s greatest humanitarian benefits and because it is a region of particularly complex regulation. We do not yet know exactly how AI will accelerate biomedical innovation, but it appears likely to:

• Significantly increase the rate at which candidate new drugs enter the regulatory pipeline;
• Increase the efficacy of new drugs and improve their safety profile through better optimization, and potentially gain a deeper understanding of their underlying biological properties;
• Developed candidate drugs for diseases that had never been successfully treated before;
• Rapidly create entirely new forms of therapies, just as antibodies, peptides, and cell therapies have become new categories of treatments over the past few decades.

Some of these advances will naturally accelerate the regulatory timeline without requiring structural changes. Drugs with larger effect sizes can lead to smaller, lower-cost clinical trials and activate accelerated approval pathways. However, the current regulatory system is designed to impose high levels of scrutiny and multiple stages of testing, based on the assumption that candidate drugs are typically ineffective and, even when effective, usually pose serious safety concerns. Whether at the FDA or the European Medicines Agency (EMA), the typical time for a candidate drug to pass through the regulatory pipeline is 7–8 years, partly due to these pessimistic assumptions. Without reform, AI will simply cause this system to become congested or overwhelmed.

Clearly, we don’t want to change things in a way that leads to a flood of snake-oil treatments slang for fraudulent remedies or widespread safety incidents. But some relatively simple reforms could make the FDA, EMA, and similar agencies better adapted to the rapid scientific acceleration driven by AI—if that happens.

Many clinical process steps that previously required expensive and slow experiments can now be completed quickly through AI simulation or analysis. Regulatory agencies should now consider establishing standards for the conditions under which such methods can be accepted. This means that once these methods prove effective, they can be rapidly adopted, rather than enduring unnecessary prolonged periods of additional testing requirements. Areas where this may apply include:

• AI-based pharmacodynamic and pharmacokinetic (PD/PK) modeling;
• Toxicology predictions to avoid the need for multi-species animal toxicology studies;
• More precise dosage selection to reduce the need for a wide range of doses in trials;
• Validate biomarkers by analyzing large datasets;
• Synthetic control arms in clinical trials to reduce the need to recruit additional participants;
• Develop alternative endpoints (this is particularly important in aging and neurodegenerative diseases).

Beyond these specific examples, institutions should also consider more aggressive and flexible accelerated approval mechanisms. If my prediction about AI is correct, many highly effective interventions will soon emerge out of the blue, and regulatory systems should be prepared to take them seriously rather than adopting an overly skeptical stance.

The acceleration of biomedical research should significantly increase the benefits of AI, but it is worth noting that it may also help reduce AI risks. Reforming biomedical approval processes could aid in biodefense, and AI-driven biomedical advancements may also improve mental health, potentially having a stabilizing effect on society.

4. National and Civil Liberties

Every system of government must confront the issue of state power and its limits. The state has legitimate, often vital, interests in protecting its population from internal and external threats. But granting it too much power is a path to tyranny. Modern democracies have largely succeeded in managing this balance, yet even at their best, this equilibrium remains precarious. Maintaining this balance requires a vast array of legal and constitutional mechanisms developed over centuries—for example, in the United States, the First, Fourth, and Fifth Amendments, the Posse Comitatus Act, FISA, and more.

AI threatens to disrupt this balance while significantly raising the stakes. But if we respond swiftly and rise to the challenge, we can leverage AI to create a world with stronger, more enduring safeguards for freedom and better defense against threats.

A powerful AI in the wrong hands could become the ultimate tool of authoritarian dictatorship, and our existing legal and constitutional protections are not fully equipped to counter this threat. Fundamentally, the enormous rewards that intelligence brings to global power, combined with the rapid pace of AI advancement, has created a perfect storm (idiom meaning an extremely bad situation caused by the simultaneous occurrence of multiple negative factors) for various dangerous actors to accidentally seize control.

This danger may take various specific technical or operational forms, but their common thread is that AI could suddenly grant immense power while routing around existing democratic oversight mechanisms. 俚语/引申义，意为绕过障碍或监管机制 A fully autonomous drone army that sounds like science fiction today could, in the future, obey illegal orders and enable governments to unilaterally consolidate power—something trained human operators are far more likely to resist. An AI focused on surveillance could analyze vast amounts of publicly available information to infer the most intimate details of every citizen’s life—a technological capability not accounted for in current civil liberties laws. All of this could occur rapidly or in secret, making it crucial to proactively strengthen democratic nations’ commitments to freedom and civil liberties.

Here are some policy ideas we should consider:

• Establish reliable accountability rules for fully autonomous weapons. Fully autonomous weapons, particularly any autonomous systems that coordinate or command them, should be required to respond to constitutional and command accountability mechanisms—such as court orders, legislation, and accountability to senior human supervisors—rather than blindly obeying orders. This may mean that a well-designed legal review panel or judicial body holds control over a “kill switch,” or that the system is inherently trained to identify and respond to legitimate oversight authorities—or both.
• Fully automatic weapons are prohibited for domestic use. Although there may be legitimate justifications for fully automatic weapons in defending against foreign adversaries—such as Russia’s invasion of Ukraine—there is no justification for their use against Americans. While the military’s capabilities for domestic operations are already somewhat restricted, ideally, these weapons should also be banned from law enforcement use.
• Close the loophole for bulk data collection/data brokers. Under current law, data shared by Americans with private companies (such as internet providers) can be purchased and used for bulk analysis in domestic surveillance and law enforcement. This privacy loophole has existed long before AI, but AI will greatly increase the stakes by making massive analysis of such data far more revealing and practical than ever before. This loophole must be closed.
• The public has a right to access AI advice when facing adverse government actions. As a general principle, it seems essential that any individual or organization subjected to adverse government actions—such as regulatory or legal proceedings—must have the opportunity to obtain AI at least as powerful as the AI permitted to the government in such actions. This would prevent the government from gaining an unfair advantage that could effectively undermine citizens’ legitimate rights. This could be added as an extension or interpretation of the Administrative Procedure Act, due process protections, or the Sixth Amendment right to legal counsel.

Finally, it is worth noting that governments are not the only entities we should be wary of in the realm of AI-driven power grabs. At various points in history—such as America’s Gilded Age or the British East India Company—corporations grew powerful enough to capture the state or adopt quasi-state characteristics. AI will soon become so powerful that I fear it cannot be safely entrusted entirely to either governments or corporations, and must be subject to checks and balances on both sides.

Regulation is one answer to how to rein in companies (see Section 1 for my thoughts on this), but equally important is that AI companies must have greater power distribution and accountability than typical private entities. Anthropic’s Long-Term Benefit Trust—a standalone governance body designed to hold the company accountable to its mission—is one such structure, and the industry should continue exploring even more robust mechanisms. Achieving the right balance—ensuring that both corporate and governmental powers are meaningfully constrained—is critical.

5. Ensure the leadership of democratic nations

Viewing new technologies as tools of trade policy, with the goal of “spreading our tech stack globally,” has become a common instinct, perhaps developed from recent experiences in the internet and telecommunications industries. But I am firmly convinced that AI is something far more profound—it resets the entire board, and all future geopolitical strategies must be built around it—like nuclear weapons, but potentially with even greater impact.

If AI truly becomes a "genius nation inside data centers" or anything remotely similar, it is highly likely to become the primary source of military and economic power for any nation. In a virtual nation of 100 million geniuses, 10 million could be applied to military strategy, 10 million to drone manufacturing, 10 million to weapons development, 10 million to intelligence gathering and analysis, 10 million to general scientific advancement, and so on. A country with advanced AI facing a country without AI—or even one that is three years behind in AI—would have a disparity comparable to a World War II Marine corps facing a medieval army of swordsmen.

In addition, if powerful AI enables deeper and potentially permanent forms of authoritarian repression (see Section 4), it becomes all the more critical to ensure that the world’s most powerful nations are democratic—or at least have robust safeguards against AI-driven repression. This also heightens the urgency of developing a targeted geopolitical strategy.

Democratic nations should seek to establish a global alliance centered on building AI according to shared values, iteratively striving to attract the rest of the world by making membership in the alliance increasingly attractive and exclusion increasingly unappealing. This alliance should be a coordinated internationalization of the AI policy ideas discussed in Sections 1 through 4, supplemented by an effort to lock in critical supply chains necessary for building AI by sharing them within the alliance and withholding them from outsiders. Some principles and operational goals may include:

Manage the AI supply chain. Trusted alliance members should be free to share chips and semiconductor manufacturing equipment (SME) with each other while collaboratively working to deny access to adversaries. U.S. restrictions on the export of cutting-edge chips and SME to China are one of the primary reasons the U.S. maintains its overall lead in AI, and these policies need to be expanded, tightened, and coordinated with other like-minded nations. Pending legislation such as MATCH and OVERWATCH provides a strong starting point, and allied democracies should consider adopting similar measures.

• Coordinate to address AI risks. If coordinated internationally, the policies described in Section 1 aimed at addressing biological, cybersecurity, and autonomy risks will be more effective (while imposing less burden on the industry). This would mean companies can comply with mutually compatible standards, and regulators can learn from each other on how best to measure and mitigate these risks. Law enforcement and intelligence agencies should also collaborate more closely to track and disrupt abuse threats, such as terrorists attempting to use AI to develop biological weapons.
• Share the benefits of AI. Trade and regulatory policies can be used to accelerate the diffusion of AI economic benefits within the alliance and share lessons learned on how to spur innovation. Coordinating approaches to beneficial deployment can help bring the advantages of AI to developing countries. For example, harmonizing medical approval systems can enable AI-powered drugs to be tested and approved more quickly and effectively (as described in Section 3 above).
• Collective defense. Member nations should cooperate to use AI for self-defense and to counter adversaries' AI. The alliance should collectively ensure adequate production of AI-powered cyber defense, AI-driven drones, AI-driven manufacturing, secure AI computing power, AI-driven R&D, and AI-driven intelligence collection sharing.
• Reject AI-enabled repression. Alliance members must reject the high-tech, extreme authoritarian, AI-enabled tyranny I warned against in The Adolescence of Technology, and must have similar safeguards as described in Section 4 above.
• Macroeconomic cooperation. Employment or job stability crises, like any other economic crisis, are transnational in their contagion. Therefore, countries share a common interest in coordinating macroeconomic support and stabilization policies (as described in Section 2) to offset any employment impacts.

The goal should be to make membership in the alliance as attractive as possible—and make the cost of remaining outside it unmistakably clear. The alliance will be built on coordination among sovereign nations, each retaining full control over its own affairs. It can evolve iteratively, starting with ideologically aligned democratic nations (which naturally fit in) and gradually welcoming those less naturally aligned but ready to meet the alliance’s standards in exchange for its substantial benefits. Ideally, the entire world will eventually join. But even if that is not possible, establishing the alliance will position democratic nations most strongly to contain and outpace regimes fixated on oppression.

Window of opportunity

The exponential progress of AI has brought an urgency and pace of change that policymaking processes are typically ill-equipped to handle. But it has also created a unique window of opportunity. Clear and tangible evidence of AI risks, early experiences with its potential to both create and disrupt economic value, and strong public backlash against unregulated approaches to AI have converged to create an unusually open environment for policymakers to take proactive action. The trees and their forest are waking up.

It has become popular in the AI industry to frame this as a public relations issue—that AI needs “better marketing.” I completely reject this framework. People are concerned about AI because they rightly perceive its risks as real, not because AI CEOs aren’t sufficiently “Panglossian” (literary allusion/adjective, meaning blindly optimistic). I believe it is my responsibility as an AI leader to remain transparent about these risks, and the public’s concern over this transparency reflects democratic accountability functioning as it should. The key challenge is to channel this concern into constructive solutions, preventing it from devolving into aimless anger and violence.

I am optimistic about finding solutions, as many issues—from addressing unemployment and testing models before release to chip export controls and other AI-related policies on energy consumption—have common-sense appeal across the political spectrum. There is a desirable yet realistic future in which a broad, nonpartisan coalition, driven by a direct understanding of the challenges posed by AI, will pass sensible and forward-looking policies much faster than usual. The sooner we do this, the sooner we can share the incredible benefits of AI.

I would like to thank Allan Dafoe, Mariano-Florentino Cuéllar, Richard Fontaine, Buddy Shah, Vas Narasimhan, Matt Yglesias, Nick Beckstead, Jason Matheny, Brad Carson, and many employees at Anthropic for their comments and feedback on this draft.

Footnote

In my article “The Adolescence of Technology,” I discussed bio-risks and autonomy risks, among other topics. Anthropic’s research team has also released preliminary internal data in “When AI Builds Itself” on the potential for recursive self-improvement, where models can autonomously build better models.

This phenomenon is not theoretical: we have observed it repeatedly within our own voluntary governance frameworks, such as our Responsible Scaling Policy. If we were to impose a fixed or rigid checklist of safety requirements on future AI models, a very likely outcome would be that requirements proven to be inconsequential consume 95% of our compliance efforts, while at the same time we fail to anticipate some of the largest sources of risk entirely within our checklist. Voluntary frameworks can evolve and adapt, but this is far more difficult with legislation. My attempts to navigate this dilemma are evident in the two public letters I issued regarding SB 1047, a 2024 California bill attempting to address catastrophic risks, for which I hold deeply complex feelings due to the reasons outlined above.

For example, serious biological risks may be harder to manage than cyber risks, because attackers have a significant advantage over defenders, and the potential severity of disasters could be much greater.

Refer to "The Adolescence of Technology" for a more detailed analysis of why the logic that led to rapid labor market recovery without lasting labor displacement in other technologies may not apply to AI, particularly why typical adaptation mechanisms such as Jevons’ paradox or comparative advantage might be overwhelmed by the pace of the technology.

For example, people still devote their lives to playing chess or Go, or climbing mountains, and are still highly respected for these activities, even though machines can perform all of them better.

This actually gives people an additional incentive to switch jobs and begin training for a new career path, even if it’s painful in the short term, by covering the difference between their old and new salaries.

For more information on this topic, see "The Teenage Years of Technology." (This article was originally published on the Titanium Media APP, author | Silicon Valley Tech_news, editor | Lin Shen)