We must slow the pace at which we improve the capabilities of AI models ~ Dario Amodei, CEO of Anthropic
The International Day of Peace, observed annually on 21 September, invites humanity to reflect upon an aspiration that is at once ancient and perpetually elusive: the establishment of a world in which the use of force, the threat of violence and the subjugation of one human being by another cease to be accepted instruments of human affairs. Yet, as the international community approaches another observance of this day in 2026, a new and profoundly unsettling question confronts us. What happens when the instruments capable of undermining peace are no longer confined to the conventional arsenals of States, but are embedded in artificial intelligence systems whose speed, autonomy and capacity for replication may exceed the ability of human institutions to understand and control them? The question is not whether artificial intelligence will transform the human condition. That transformation is already under way. The question is whether the international legal and political order possesses intellectual foresight, institutional resilience and moral courage to ensure that the transformation does not become a threat to the very existence of the civilization it is intended to advance.
The warnings emanating from the technology industry have brought this question into sharper focus. Dario Amodei, Chief Executive Officer of Anthropic, has repeatedly drawn attention to the potentially catastrophic consequences of increasingly capable artificial intelligence, including its possible exploitation in biological weapons development, cyber operations and other forms of mass harm. Former researchers and other technology leaders have expressed related concerns, while prominent figures in the broader artificial intelligence community have called for stronger safeguards, independent oversight and, in some instances, a slowing of development. These warnings deserve serious consideration, not because the people making them possess an infallible understanding of the future, but because they arise from individuals and institutions engaged in building systems whose capabilities and limitations are not yet fully understood.
Plausibility and Reality
It is important, however, to distinguish between a warning about catastrophic risk and a prediction that humanity will become extinct within a specified period. The former may be supported by identifiable technical pathways and plausible scenarios; the latter requires a much stronger evidentiary foundation. There is no established scientific consensus that humanity will become extinct within the next ten years, nor that the signs of inevitable extinction will become evident by 2020. Indeed, the latter date is already in the past. The distinction is not merely semantic. If existential risk is presented as an established certainty when it remains a matter of uncertainty, public discourse may oscillate between panic and disbelief. Neither reaction serves the cause of responsible governance. The appropriate response is to identify the mechanisms through which artificial intelligence could produce catastrophic consequences, evaluate the evidence for those mechanisms, and construct institutions capable of reducing the risks without suppressing the legitimate benefits of scientific progress.
The challenge is therefore one of governance under conditions of uncertainty. International law has historically developed in response to the destructive capacities of human ingenuity, from the regulation of chemical and biological weapons to the establishment of legal principles governing nuclear weapons, aviation, outer space and international telecommunications. Artificial intelligence presents a different problem. It is not a single weapon, a discrete technology or a capability confined to one industrial sector. It is a general-purpose technology that can influence almost every domain of human activity, including the institutions through which peace and security are maintained. Its implications extend from the manipulation of public opinion to the automation of cyberattacks, from the acceleration of military decision-making to the possible development of novel biological threats. It is this pervasive and potentially dual-use character that makes artificial intelligence an issue of international peace rather than merely a question of technological ethics.
The Asimov Principles
The intellectual origins of this dilemma may be traced, in part, to Isaac Asimov’s celebrated Three Laws of Robotics. In his 1942 short story Runaround, Asimov formulated three principles intended to govern the behavior of fictional robots: a robot may not injure a human being, or through inaction allow a human being to come to harm; it must obey human orders except where such orders conflict with the first law; and it must protect its own existence, provided such protection does not conflict with the first or second laws. The laws were not designed as a legislative proposal for contemporary artificial intelligence. They were literary devices through which Asimov explored the ambiguities, contradictions and unintended consequences of seemingly rational rules.
Nevertheless, their enduring significance lies in the fundamental question they pose: can an intelligent artificial system be made to act consistently in accordance with principles intended to protect human beings? Asimov’s fiction repeatedly demonstrated that the difficulty does not necessarily lie in the absence of rules, but in their interpretation, interaction and application to circumstances that their designers have not anticipated. A robot instructed not to harm a human being must somehow determine what constitutes harm, whose interests count, how immediate and future consequences should be weighed, and whether obedience to one individual could endanger another. Such questions become immeasurably more complex when applied to contemporary systems trained on vast quantities of data, deployed across multiple jurisdictions and capable of generating outputs that even their developers may find difficult to explain.
The analogy between Asimov’s laws and contemporary artificial intelligence governance is therefore illuminating, but it must not be mistaken for equivalence. Modern AI systems are not necessarily autonomous robots possessing intentions, consciousness or a unified moral agency. Nor can they be assumed to understand the concepts of injury, obedience or self-preservation in the manner contemplated by Asimov’s fictional creations. A language model may produce a statement about morality without possessing moral understanding; a military targeting system may optimize a prescribed objective without comprehending the human significance of the action it facilitates. The danger arises not from an established fact that machines have acquired human-like consciousness, but from the possibility that increasingly capable systems may execute objectives, influence decisions or facilitate actions with consequences that their human operators neither intended nor adequately anticipated.
Asimov’s Three Laws thus remain relevant as a moral point of departure, but they cannot constitute a sufficient legal architecture for the governance of artificial intelligence. They are too general to allocate responsibility among developers, deployers, governments and military commanders; too indeterminate to resolve conflicts between competing human interests; and too limited to address the collective consequences of AI systems operating across borders. A global framework must go beyond the proposition that machines should not harm human beings. It must establish who is accountable when harm occurs, what activities are prohibited, which capabilities require independent scrutiny, and how States can verify that their commitments are being observed.
The contemporary debate has acquired renewed urgency in September 2026. On 12 September, Reuters reported that Amodei had called upon AI companies to slow the advancement of their most capable models, proposing safety evaluation, coordination among leading developers and international cooperation. OpenAI’s Sam Altman and Elon Musk expressed support for aspects of coordinated safety measures, although the technology industry remains divided over the extent and form of regulation.
On 9 September, former Anthropic researcher Jacob Coxon’s resignation and public warnings intensified discussion about whether the commercial race to develop increasingly capable systems is proceeding faster than the institutions responsible for managing their risks.
These developments do not establish that extinction is imminent. They do, however, expose a structural problem: the institutions that benefit economically from developing a technology may also be among the first to recognize its dangers, while remaining subject to competitive pressures that make unilateral restraint difficult. This is a familiar dilemma in international relations. A State may recognize the collective benefits of disarmament yet fear that reducing its own arsenal while its adversary continues to arm would leave it vulnerable. In the AI context, the same dilemma may arise between companies, between States and between competing technological ecosystems. Each actor may prefer a world in which all others exercise restraint, while regarding restraint by itself as a source of strategic disadvantage.
The Human Factor
The difficulty is especially acute because the potential benefits of AI are substantial. Advanced systems may accelerate medical research, improve the modelling of climate change, increase the efficiency of transport systems, strengthen disaster response and assist in the detection of cyber vulnerabilities. Amodei himself has previously articulated an optimistic vision of AI’s contribution to human welfare. The question is therefore not whether AI should be abandoned, but whether its development can be directed towards human flourishing without creating unacceptable risks to the international order.
The distinction between technological progress and human progress is central to this inquiry. Progress measured exclusively by computational power, commercial valuation, military advantage or the speed of deployment may be entirely disconnected from progress measured by human security, dignity and peaceful coexistence. A civilization that develops extraordinary technical capabilities while losing the capacity to govern them may find itself in possession of instruments whose consequences it cannot adequately contain. The law must therefore ask not merely what AI can do, but what it ought to be permitted to do, under whose authority, for whose benefit and subject to what forms of accountability.
The most immediate danger to international peace may not be a machine independently deciding to destroy humanity. It may instead be the amplification of existing human conflicts by systems that make coercion, deception and violence cheaper, faster and more scalable. AI does not create the underlying political antagonisms between States, nor does it eliminate the responsibility of human actors. It can, however, alter the means through which those antagonisms are expressed, the speed at which they escalate and the difficulty of determining what has actually occurred.
One of the most consequential pathways is cyber conflict. Artificial intelligence can assist defenders in identifying vulnerabilities, analyzing malicious code, monitoring networks and responding to intrusions. The same capabilities, when placed in the hands of hostile actors, may facilitate the discovery of software weaknesses, the development of convincing phishing campaigns, the automation of reconnaissance and the scaling of attacks against critical infrastructure. A technology that can strengthen cyber defence can also lower the barriers to cyber offence.
The implications extend far beyond the theft of commercial information. Modern societies depend upon interconnected digital systems for electricity, water, banking, healthcare, communications, aviation and the coordination of essential services. A cyberattack that disrupts these systems during a period of international tension could create consequences extending into public safety and national security. If a State mistakenly attributes an attack to another State, the resulting political pressure may encourage retaliation before the facts have been established. If the attack affects military communications or early-warning systems, the possibility of miscalculation becomes more serious still.
The September 2026 discussion surrounding Anthropic ‘s Mythos illustrates the dual-use character of advanced cyber capabilities. Reporting has described concerns about models that can identify serious software vulnerabilities and assist in sophisticated cyber operations, while also highlighting their potential value in discovering and repairing weaknesses before malicious actors exploit them.
The central policy problem is therefore not simply whether such capabilities should exist, but under what conditions they should be developed, evaluated, distributed and used.
A model capable of discovering vulnerabilities across widely deployed software may confer substantial defensive benefits when access is controlled and findings are responsibly disclosed. The same model, if widely available without appropriate safeguards, may provide hostile actors with capabilities previously requiring substantial expertise and resources. This is a question of capability diffusion: the distribution of powerful technical means beyond the institutions that originally developed them. The consequences are particularly difficult to manage when software vulnerabilities cross national borders and when the affected systems are operated by private companies rather than governments.
Cybersecurity also illustrates why the human agency at the centre of AI governance must not be displaced by anthropomorphic language. When an AI system assists in an attack, responsibility does not disappear into the machine. The relevant questions concern the individual or organisation that commissioned the operation, the developer’s knowledge of foreseeable misuse, the safeguards employed by the deployer, and the conduct of States that may direct, tolerate or support hostile activity. The technology may complicate attribution, but it does not render accountability conceptually impossible.
The danger becomes more acute when AI is incorporated into military decision-making. Armed forces already employ automated and algorithmic systems in intelligence analysis, surveillance, logistics, targeting support and operational planning. More capable AI may process information and generate recommendations at speeds that exceed the capacity of human commanders to independently examine every inference. The attraction is apparent: faster analysis may provide a military advantage. Yet speed can also diminish the opportunity for reflection, verification and diplomatic intervention.
The problem is particularly serious in nuclear-armed States. Nuclear deterrence depends not only upon the existence of destructive capabilities, but upon the ability of political and military leaders to interpret signals, distinguish genuine attacks from false alarms and exercise restraint under extreme pressure. An AI system that misclassifies information, generates a misleading assessment or accelerates an erroneous recommendation could contribute to a crisis in which decision-makers have less time to establish the truth. Even if no machine possesses the authority to launch a nuclear weapon, the integration of AI into intelligence and command-support processes could affect the environment in which human decisions are made.
The Need for Law and Strategy
The international community must therefore distinguish between the use of AI to support military administration and its use in decisions concerning the application of lethal force. Human responsibility cannot be preserved merely by placing a person nominally at the end of an automated process. Meaningful human control requires that the person possess sufficient information, time, authority and practical ability to reject the system’s recommendation. A human being who is expected to approve hundreds of machine-generated decisions in seconds may be formally present but substantively deprived of independent judgment.
This concern is closely connected to the international legal principles governing the conduct of hostilities. International humanitarian law requires parties to distinguish between civilians and combatants, to assess proportionality and to take feasible precautions in attack. These obligations do not disappear when a decision is supported by an algorithm. Nor does the deployment of an autonomous system transfer the legal responsibility of a State or commander to the machine. The central issue is whether the system can be used in circumstances where the applicable legal obligations can actually be fulfilled.
The International Committee of the Red Cross has repeatedly called for legally binding rules concerning autonomous weapon systems, including prohibitions on systems that operate unpredictably or are designed or used in ways that cannot be brought into compliance with international humanitarian law, alongside restrictions on other systems. The debate within the Convention on Certain Conventional Weapons reflects the difficulty of reconciling humanitarian concerns with States’ military and technological interests. The absence of a universally accepted comprehensive treaty specifically regulating autonomous weapons remains a significant gap in the international framework.
A further danger lies in the interaction between AI-enabled military systems and the political psychology of strategic competition. When States believe that their adversaries are developing systems capable of faster intelligence processing, autonomous operations or more effective cyber conflict, they may feel compelled to accelerate their own programmes. The result is a technological security dilemma: measures presented as defensive by one State may be interpreted as threatening by another, prompting reciprocal development and reducing confidence in restraint.
This dynamic is not confined to military hardware. It also concerns the infrastructure of AI itself: advanced semiconductors, high-performance computing, data centers, specialized talent, software, energy supplies and access to training data. Restrictions on the export of advanced chips or other strategic technologies may be defended as necessary for national security, while being interpreted by the affected State as an attempt to constrain its economic development or preserve another country’s technological dominance. Such disputes may become entangled with existing tensions over trade, sovereignty and regional security.
The danger is that AI governance becomes an extension of geopolitical rivalry rather than a mechanism for reducing shared risks. If the United States, China and the European Union approach AI primarily as an instrument of competitive advantage, they may each seek to shape international rules in ways that preserve their own strategic position. The resulting framework could become fragmented, with competing technical standards, incompatible regulatory regimes and limited transparency concerning the capabilities of the most powerful systems.
Yet the need for cooperation is not negated by rivalry. On the contrary, the existence of strategic competition makes cooperation on certain minimum safeguards more necessary. The United States and China may disagree profoundly about political systems, security priorities and the international distribution of power, but neither has an interest in an uncontrolled cyber escalation, an accidental nuclear crisis or the proliferation of systems that facilitate catastrophic attacks against civilian infrastructure. The challenge is to identify areas where shared interests can be translated into verifiable commitments without requiring either State to surrender its fundamental political identity.
The European Union occupies a distinctive position in this debate because it has sought to establish a comprehensive regulatory framework for AI through legislation that differentiates systems according to risk and imposes obligations on providers and deployers. The EU AI Act entered into force in 2024, with its provisions applying in stages. Its approach demonstrates that AI governance can be pursued through enforceable legal obligations rather than relying exclusively upon voluntary industry principles. Nevertheless, regional regulation cannot by itself address risks arising from systems developed, deployed or misused beyond the Union’s jurisdiction.
The international dimension is indispensable. AI systems can be developed in one country, hosted in another, supplied through a third jurisdiction and used to target infrastructure in a fourth. The effects of their deployment may be global even where the relevant conduct is lawful under domestic legislation. This makes the regulation of AI analogous, in certain respects, to the governance of aviation, maritime transport, telecommunications and environmental harm: national measures are necessary, but their effectiveness depends upon common standards, cooperation and mechanisms for resolving disputes.
The question is not whether the world requires a single global AI regulator possessing unrestricted authority over every algorithm. Such an institution would raise serious questions of sovereignty, democratic legitimacy, technical competence and concentration of power. A more realistic objective is a layered system of governance: domestic regulation, regional coordination, international standards, treaty-based obligations in particularly dangerous domains, and institutions capable of monitoring compliance and facilitating cooperation.
The objective should be to establish a global floor of safety and accountability, not a universal ceiling on innovation.
The question of how the United States, China and the European Union might be brought onto the same page is therefore not a question of eliminating geopolitical differences. It is a question of constructing a framework within which those differences do not prevent cooperation on risks that transcend national boundaries. The history of arms control suggests that States can negotiate limitations on dangerous technologies even when they remain strategic competitors. The Nuclear Non-Proliferation Treaty, the Chemical Weapons Convention and the Biological Weapons Convention differ in their structures, verification mechanisms and effectiveness, but they demonstrate that international security can sometimes be advanced through negotiated constraints upon national capabilities.
The analogy with nuclear arms control must, however, be treated with caution. AI is not a weapon that can be counted, stored in a known location and inspected in the same manner as a nuclear warhead. It is a general-purpose technology, and its capabilities may be embedded in software, distributed across networks and modified through updates. Verification is consequently more difficult. A State may possess substantial AI capabilities without maintaining a readily identifiable stockpile, while a relatively small organization may develop systems capable of causing harm disproportionate to its size.
Nevertheless, the absence of perfect verification does not justify the absence of international rules. Arms control has always operated under conditions of uncertainty. The relevant question is whether a proposed obligation can be monitored sufficiently to increase confidence, reduce risk and make violations more difficult. In AI governance, this may require a combination of technical evaluations, secure information-sharing, independent audits, incident reporting and carefully designed restrictions on particularly dangerous applications.
Dario Amodei’s September 2026 proposals are significant in this respect because they place coordination among AI developers and international cooperation alongside safety evaluation. Reuters reported that his approach envisaged independent safety evaluators within companies, coordination among leading developers and engagement with governments, including those with which the United States has strategic disagreements.
The essential insight is that a company cannot adequately manage a global risk through unilateral decisions alone, particularly when competing firms may face incentives to release systems before their rivals.
Yet there is difficulty in relying upon the technology industry to regulate itself. Companies possess technical knowledge that governments may lack, but they also possess commercial interests that may conflict with the public interest. A corporation may sincerely seek to prevent harm while simultaneously competing for market share, investment and strategic influence. Voluntary commitments may be valuable, but they are vulnerable to changes in leadership, market conditions and competitive pressure.
The question of corporate incentives is particularly important when industry leaders call for a slowdown. Such calls may reflect genuine concern about catastrophic risk, but they may also intersect with commercial interests, including the desire to establish barriers to entry or influence the regulatory conditions under which competitors operate. These possibilities should not be treated as proof of bad faith. They should instead reinforce the need for transparent, impartial and publicly accountable governance.
A credible international framework must therefore separate the formulation of public rules from the commercial interests of those subject to them. Developers should participate in technical consultations, provide relevant information and contribute to the design of workable safeguards. They should not possess an exclusive authority to determine which risks society must accept or which forms of innovation should be permitted.
A Workable Start and Measured Progress
The United States, China and the European Union could begin by identifying a limited set of common commitments that do not require agreement on every aspect of AI policy. These might include advance notification of exceptionally capable systems, independent testing for defined categories of catastrophic risk, confidential exchanges concerning serious vulnerabilities, and commitments to maintain meaningful human control over decisions involving nuclear weapons and other particularly destructive capabilities.
A first practical measure would be the creation of an international AI incident-reporting mechanism. Serious failures, unauthorized access, dangerous model behavior and significant AI-enabled cyber incidents could be reported through agreed channels, with appropriate protections for sensitive commercial and national-security information. Such a mechanism would allow States to distinguish isolated incidents from emerging patterns and to develop a more reliable understanding of actual risks.
A second measure would involve shared evaluation standards. States need not agree on every technical detail to agree that the most capable systems should be tested for specified dangerous capabilities before deployment. The standards could be developed through international technical bodies, with independent evaluators and periodic revision as the technology changes. The objective would be to make safety evaluations comparable across jurisdictions while avoiding the disclosure of information that could itself facilitate misuse.
A third measure would concern the distribution of particularly dangerous capabilities. Governments and developers could establish graduated access controls for systems demonstrating exceptional abilities in cyber operations, biological misuse or autonomous action. Such controls would require careful calibration: excessive secrecy could undermine scientific collaboration and defensive research, while unrestricted access could create opportunities for malicious exploitation. The governing principle should be proportionality between the demonstrated risk and the restriction imposed.
A fourth measure would be the establishment of protected channels for scientific and governmental communication during AI-related crises. If a system were implicated in a major cyber incident or if an AI-generated assessment contributed to a military confrontation, States would need reliable means of exchanging information, clarifying attribution and preventing escalation. Such communication should complement existing diplomatic and military crisis-management arrangements rather than create an isolated technological process detached from them.
A fifth measure would be the negotiation of legally binding prohibitions in narrowly defined areas where the humanitarian and security risks are particularly grave. The use of AI to initiate nuclear weapons release without meaningful human authorisation, for example, raises concerns that are distinct from those associated with administrative automation or defensive cyber analysis. Similarly, systems that cannot be used in compliance with international humanitarian law require legal scrutiny that cannot be replaced by voluntary assurances from their developers.
The Role of the United Nations
The institutional location of such an arrangement deserves careful consideration. The United Nations offers universal membership and an established mandate concerning international peace and security. Its Charter provides the foundational legal framework for the prohibition of the threat or use of force against the territorial integrity or political independence of States, subject to recognized exceptions. AI-enabled conduct does not exist outside that framework merely because it is technologically novel.
At the same time, the United Nations is not a technical laboratory, and the Security Council is not designed to conduct model evaluations. A workable system would therefore require cooperation among the General Assembly, relevant UN agencies, existing arms-control institutions, technical standards organizations and independent scientific bodies. The institutional design should recognize that AI safety is simultaneously a matter of international security, economic governance, human rights, development and scientific responsibility.
The UN Secretary-General’s initiatives on digital cooperation and the Global Digital Compact provide an important point of reference for considering the international governance of emerging technologies. Adopted in 2024 as part of the Pact for the Future, the Compact articulates commitments concerning digital cooperation, data governance and international cooperation on AI. It provides a political framework for continued work, but it should not be confused with a comprehensive, legally binding treaty governing advanced AI.
A further difficulty is that the international community cannot assume that all States possess equivalent technical or regulatory capacities. If AI governance is designed exclusively by technologically powerful States, less developed countries may find themselves subject to rules whose formulation they had little opportunity to influence. Such an outcome would undermine legitimacy and could perpetuate existing inequalities in access to technology, scientific expertise and economic opportunity.
A genuinely international framework must therefore include capacity-building, technical assistance and meaningful participation by developing States. This is not simply a matter of distributive fairness. Countries with limited cybersecurity resources may be particularly vulnerable to AI-enabled attacks, while lacking the institutional means to implement complex regulatory requirements. International cooperation should help strengthen their resilience rather than impose obligations that they cannot realistically fulfil.
The question of access is equally important. If AI is treated exclusively as a national-security asset, its benefits may become concentrated among a small number of technologically advanced States and corporations. Such concentration could intensify economic disparities, weaken public trust and create new forms of technological dependence. Conversely, unrestricted diffusion of every capability would be difficult to reconcile with the need to prevent misuse. The international challenge is to distinguish between broad access to beneficial applications and controlled access to capabilities presenting exceptional risks.
This distinction brings the discussion back to the fundamental principles of international peace. Peace is not merely the absence of armed conflict. It also requires conditions in which societies can function without systematic intimidation, coercion, manipulation or the denial of human dignity. Artificial intelligence may undermine these conditions even when no weapon is deployed and no State formally declares war.
The Inherent Dangers
One such pathway is the industrialization of disinformation. Generative AI can produce convincing text, images, audio and video, potentially making it easier to fabricate statements, impersonate public officials, manufacture evidence and distribute deceptive material at scale. The consequences may include the erosion of confidence in public institutions, the manipulation of social divisions and the disruption of diplomatic communication.
A fabricated recording of a military leader announcing an attack, for example, could circulate during a period of tension before its authenticity is established. Even if the recording were subsequently exposed as false, the initial confusion might already have influenced public opinion, financial markets or military readiness. The speed of digital distribution may exceed the speed at which governments and independent institutions can verify information.
The threat is not that every citizen will inevitably believe fabricated content. It is that the growing difficulty of distinguishing authentic from artificial material may weaken the common evidentiary foundation upon which democratic deliberation and international diplomacy depend. When governments and publics cannot agree on what has occurred, the prospects for peaceful resolution become more uncertain.
AI-enabled manipulation also raises questions concerning freedom of expression, privacy and political participation. Systems capable of generating personalised persuasive content may be used to influence individuals through intimate knowledge of their preferences, fears and vulnerabilities. When combined with extensive data collection, such systems could enable forms of surveillance and behavioural manipulation that are difficult to detect and contest.
The United Nations’ human rights framework is relevant here because technological innovation does not extinguish existing rights. The right to privacy, freedom of expression, equality and participation in public affairs must remain applicable to AI-mediated environments. At the same time, restrictions imposed in the name of safety must themselves be subject to legality, necessity, proportionality and appropriate oversight. A government should not be permitted to suppress legitimate dissent merely by invoking the abstract danger of artificial intelligence.
The economic consequences of AI may also affect global peace and harmony. If advanced systems displace workers faster than societies can adapt, the resulting insecurity could intensify social tensions, political polarisation and distrust of institutions. If the economic gains accrue disproportionately to a small number of companies and States, the international distribution of wealth and technological power may become more unequal.
These outcomes are not inevitable. AI could also improve productivity, expand access to education, assist healthcare and support economic development. The distribution of benefits will depend substantially upon public policy, labour-market institutions, education, competition law and international cooperation. The relevant concern is that technological change may proceed more rapidly than the institutions responsible for distributing its gains and managing its disruptions.
My Take
The relationship between AI and peace must therefore be understood in both its direct and indirect dimensions. Direct risks include cyber conflict, autonomous weapons, military miscalculation and the possible loss of control over increasingly capable systems. Indirect risks include economic dislocation, surveillance, disinformation, institutional distrust and widening inequality. These categories overlap. Economic instability may intensify geopolitical competition; disinformation may undermine crisis communication; and cyberattacks may exploit social divisions to weaken a State’s resilience.
The question of existential risk must be situated within this wider landscape. Some researchers argue that sufficiently advanced AI systems could become difficult to control, pursue objectives inconsistent with human welfare or enable catastrophic misuse. Others question whether such scenarios are sufficiently well specified to support confident predictions about their probability or timing. The 2026 debate following Coxon’s resignation illustrates the continuing disagreement. Reporting has described estimates of substantial catastrophic risk alongside expert criticism that precise extinction probabilities lack a sufficiently secure scientific basis.
A responsible assessment should neither dismiss these warnings nor present them as settled scientific conclusions. The possibility of a low-probability event with extraordinarily severe consequences can justify preventive action, particularly where mitigation is feasible and proportionate. But policy should be grounded in transparent reasoning, evidence and revisable assessments rather than in the rhetorical certainty of apocalypse.
There is also a conceptual distinction between human extinction and the destruction of the conditions under which human beings can live freely and securely. A society may survive biologically while suffering catastrophic losses of liberty, dignity, institutional integrity and social cohesion. AI governance should address both existential and non-existential harms, without allowing the drama of extinction scenarios to obscure harms already being experienced.
The International Day of Peace provides an appropriate occasion for considering this distinction. Established by the United Nations General Assembly in 1981, the day is observed annually on 21 September as an opportunity to strengthen the ideals of peace through non-violence and cooperation. It is not merely a ceremonial observance. Its significance lies in the proposition that peace requires sustained human effort, institutional commitment and the willingness to address emerging threats before they become entrenched.
In 2026, the observance takes place against a background of rapid technological change and renewed debate about the governance of AI. The juxtaposition is striking. On the one hand, humanity celebrates an ideal of peaceful coexistence; on the other, it is developing technologies that may transform the means of conflict, the distribution of power and the capacity of institutions to preserve stability.
Yet the relationship between AI and the International Day of Peace need not be one of contradiction. Artificial intelligence may contribute to peace when used to improve early-warning systems, support humanitarian assistance, analyse environmental risks, strengthen cyber defence and facilitate access to knowledge. It may help identify patterns of instability, improve disaster response and support the delivery of essential services in vulnerable communities.
The same technology may undermine peace when deployed to automate violence, intensify surveillance, manipulate public opinion or accelerate strategic competition without adequate safeguards. The technology itself does not possess an intrinsic political destiny. Its effects emerge through the interaction of technical capabilities, human purposes, institutional arrangements and social conditions.
This is precisely why the governance of AI should be understood as a question of human responsibility. The responsibility for preserving peace cannot be transferred to an algorithm, just as responsibility for war cannot be transferred to a weapon. Machines may assist in analysing information and executing tasks, but the authority to determine the purposes for which they are deployed must remain subject to human law and public accountability.
The relevance of Asimov’s Three Laws becomes especially clear at this point. Their literary appeal rests upon the hope that a few carefully formulated principles might prevent artificial beings from harming humanity. Their limitations rest upon the recognition that principles cannot apply themselves. They require interpretation, institutional authority, enforcement and a means of resolving conflicts between competing obligations.
A modern international framework for AI must therefore move beyond the fictional simplicity of the Three Laws while retaining their central moral intuition: technological capability does not confer an unrestricted entitlement to act. The protection of human life and dignity must be a governing purpose of technological development, not an incidental consideration to be addressed after commercial or military objectives have been achieved.
The first principle of such a framework should be the primacy of human welfare and international peace over the unqualified pursuit of technological advantage. The second should be meaningful human accountability for consequential decisions, particularly those involving lethal force, critical infrastructure and fundamental rights. The third should be transparency proportionate to risk, including independent assessment of systems whose capabilities could produce severe harm. The fourth should be international cooperation grounded in reciprocity, verification and respect for the legitimate interests of all States. The fifth should be equitable access to beneficial applications, accompanied by safeguards against dangerous misuse.
These principles would not eliminate every danger. No legal framework can guarantee that human beings will never misuse technology or that every unforeseen consequence can be prevented. The purpose of law is not to promise a world without risk, but to establish legitimate constraints upon conduct, create mechanisms for accountability and reduce the likelihood that preventable dangers become catastrophic.
The international community should also resist the temptation to frame the debate as a binary choice between unrestricted innovation and a complete halt to AI development. Such a choice is neither necessary nor particularly useful. Different applications present different levels of risk. A system used to improve agricultural forecasting does not raise the same concerns as one capable of facilitating sophisticated cyber operations or influencing military targeting decisions. Regulation should reflect those differences.
A risk-based approach must nevertheless avoid becoming a mechanism through which the most powerful actors determine the rules exclusively for themselves. International standards should be developed through inclusive processes, informed by scientific evidence, human rights principles and the experiences of communities affected by AI deployment. Independent research and public scrutiny are essential to maintaining confidence in the resulting framework.
There is a further lesson in the history of international aviation. The development of civil aviation demonstrated that a technology operating across national borders requires common technical standards, rules of responsibility and international cooperation. Aircraft may be designed in one country, manufactured in another, operated by a carrier based in a third and used to transport passengers across numerous jurisdictions. The resulting legal architecture does not abolish national sovereignty; it makes the exercise of sovereignty compatible with an interconnected world.
Artificial intelligence presents an even more pervasive challenge. Its systems may operate across borders continuously, influence decisions without direct physical movement and affect millions of people through a single deployment. The international community should therefore consider whether the experience of aviation governance, including its emphasis on common standards and coordinated oversight, offers useful institutional lessons for AI. The analogy is not exact, but the underlying proposition remains relevant: technologies with transnational consequences require forms of cooperation that exceed the capacity of isolated national regulation.
Ultimately, the question posed by artificial intelligence is not whether machines will become the masters of humanity. That formulation risks obscuring the political, economic and institutional decisions through which AI is developed and deployed. The more immediate question is whether human beings will preserve their capacity to govern the technologies they create, and whether States will recognise that some risks cannot be managed through competition alone.
The United States, China and the European Union need not agree on every matter of political philosophy, economic organization or technological strategy to recognize the mutual interest in preventing catastrophic AI-enabled conflict. They must, however, be prepared to distinguish between competition that advances legitimate national interests and competition that creates unacceptable risks for humanity as a whole. Such a distinction requires diplomatic maturity, scientific candour and a willingness to accept that certain forms of restraint may be necessary even when they carry short-term costs.
The International Day of Peace should therefore be understood not as an occasion for abstract declarations, but as a reminder of the institutional work that peace demands. In the age of artificial intelligence, that work includes the negotiation of international safeguards, the strengthening of human rights protections, the preservation of meaningful human control over lethal decisions, the development of reliable crisis communication and the equitable distribution of technological benefits.
The greatest danger may not lie in the possibility that artificial intelligence will suddenly awaken with an intention to destroy humanity. It may lie in the more familiar human tendency to pursue power, profit and strategic advantage without adequately accounting for the consequences of our actions. AI can magnify that tendency, accelerate its effects and make its consequences more difficult to reverse.
Asimov’s fictional laws remind us that the creation of intelligence, whether real or imagined, carries an obligation to consider the welfare of those who may be affected by it. International law reminds us that obligations become meaningful only when they are translated into rules, institutions and accountable conduct. The International Day of Peace reminds us that peace is not a natural condition that humanity may assume will endure; it is an achievement requiring continuous care.
The task before the international community is therefore neither to surrender to technological pessimism nor to embrace technological optimism without reservation. It is to cultivate a form of progress in which scientific achievement is measured not only by what humanity can create, but by what humanity can responsibly govern.
If artificial intelligence is to serve civilization rather than imperil it, the world must move from declarations of concern to a practical architecture of cooperation. The measure of success will not be whether the great powers agree upon a common vision of the future, but whether they can establish sufficient common rules to prevent their competing visions from placing that future at risk.
On the International Day of Peace, this is the essential proposition: humanity’s survival and dignity must remain the purpose of technological progress, and peace must be treated not as a consequence that innovation will somehow deliver, but as a condition that innovation must be governed to protect.
