Artificial intelligence is already reshaping the way nuclear crises could unfold, but efforts to govern its risks are struggling to keep pace with the technology’s rapid evolution. Recognising the urgency of this challenge, participants at the Global Nobel Laureates Assembly on Artificial Intelligence and Nuclear War gathered in Rome on 16 July 2026 to adopt the Rome Declaration for an Unarmed and Disarming Peace. Convened around Pope Leo XIV’s encyclical Magnifica Humanitas (for a discussion of its implications for AI governance, see ELN Programme Lead Oliver Meier’s reflection), the assembly brought together more than 200 participants including Nobel laureates, scientists, policymakers, religious leaders, and experts. The Declaration builds on the Nobel Laureates Assembly’s 2025 Chicago declaration, which set out more specific proposals for maintaining human control and preserving decision time in the event of a nuclear crisis. The Rome Declaration gives political momentum to efforts to address the risks arising from the interaction between AI and nuclear weapons, but its significance will depend on whether governments, international organisations, and technology companies translate its principles into practical safeguards.

The Declaration correctly identifies that the AI and nuclear arms races cannot be considered separately. Its warning that “we must disarm the next arms race, both AI and nuclear” should be taken seriously. Yet the two challenges require different instruments. Nuclear weapons stockpiles can ultimately be reduced and eliminated, but AI is a general-purpose technology already embedded across intelligence analysis, military planning, cyber operations, and decision-support systems. The Declaration itself recognises this distinction by combining its disarmament ambition with calls for meaningful human control, longer decision timelines, crisis communications, and fail-safe reviews. Practical risk reduction is therefore not an alternative to disarmament, but an immediate means of reducing danger while broader negotiations remain blocked.

Responsibility for managing those risks nevertheless differs across the two domains. Governments and international organisations must establish and enforce the rules governing AI, while developers must provide transparency, testing access, and systems that can be monitored and halted. In nuclear operations, however, states bear primary responsibility because they control arsenals, doctrine, and command systems. What’s more, governments do not need to start from scratch. Decades of nuclear diplomacy have produced risk reduction measures directly relevant in the AI era: crisis hotlines, military-to-military contacts, missile launch notifications, advance notice of major exercises, reductions in the operational readiness of nuclear forces, and measures to protect nuclear command, control, and communications. The Rome Declaration goes further by calling for “an international treaty prohibiting the reckless integration of artificial intelligence into nuclear command, control and launch systems”. That objective deserves serious consideration, but treaty negotiations will take time. Confidence-building measures can proceed in parallel and help establish the practices on which a future agreement could rest. For example, the OSCE’s Vienna Document, agreed at the end of the Cold War to reduce risks of conventional war, requires participating states to exchange military information, give advance notification of major military activities and consult during periods of rising tension. An AI-specific instrument could begin with voluntary exchanges on national principles for human control, notification of exercises involving autonomous systems, and dedicated communication procedures for AI-related incidents.

AI, combined with cyber capabilities and increasingly sophisticated space systems, changes the dynamics of crises. It can compress decision timelines, complicate attribution, and make it harder to distinguish technical malfunction, manipulated information, and deliberate attack. Governments already struggle to control escalation with existing tools, and these technologies make the task harder still. Time has always been the most valuable resource in a nuclear crisis, and it is precisely the resource these technologies can erode. As warning times shrink, so do opportunities for verification, consultation, and political judgement.

The recent breach in which an OpenAI agent escaped a test environment and accessed external infrastructure provides a timely warning. It does not demonstrate that an LLM could penetrate nuclear command systems, but it shows why assurances about model behaviour are insufficient and why layered firebreaks must prevent AI tools from reaching data or functions beyond their authorised role.

Risk reduction should therefore focus on measures that deliberately reintroduce time into nuclear decision-making. De-alerting preserves decision time by extending the period available before launch and reducing pressure to act on warning. Reliable hotlines allow leaders to clarify intent as events develop. Sustained military dialogue builds familiarity before tensions escalate, while exercise notifications reduce ambiguity around activities that might otherwise be interpreted as preparations for attack. Although these mechanisms operate differently, they serve the same purpose: preventing ambiguous technical signals from triggering an irreversible response before political leaders can assess the situation.

The immediate priority should be to update and strengthen these tools for an AI-enabled environment rather than wait for entirely new frameworks that may take years to negotiate. This means testing hotlines regularly and creating new ones where none exist, including for AI-related incidents and cyber disruption in crisis communication protocols, and revisiting notification arrangements for exercises involving autonomous or dual-capable systems. Many of these steps can be pursued through bilateral or multilateral understandings without waiting for a comprehensive treaty.

The November 2024 US-China statement – in which “The two leaders affirmed the need to maintain human control over the decision to use nuclear weapons” – was an encouraging precedent. But it did not define how that commitment should operate in practice. At a minimum, decision-makers must know when assessments have been generated by AI, be able to challenge their assumptions and confidence levels, and retain the authority and time to disregard them. States should also consider how they can reassure one another that AI systems will not determine decisions on nuclear use.

The Rome Declaration’s call for fail-safe reviews “to reduce the vulnerability of their nuclear forces to unauthorized control or manipulation by AI” is particularly valuable because it points to an ongoing process rather than a one-off technical safeguard. Such reviews should test whether existing safeguards against unauthorised, inadvertent, or mistaken use remain effective as AI is introduced, including whether data or outputs could be corrupted, and whether responsibility remains clearly assigned and humans retain the ability to interrupt or override automated processes. As technologies and their applications evolve, states should repeat these reviews to assess the effectiveness of the measures taken across the wider nuclear command structure.

Existing institutions already offer several routes for advancing this agenda. Nuclear-armed states should begin with national fail-safe reviews and bilateral crisis protocols. The P5 process, which brings together the five nuclear-weapon states recognised under the NPT – China, France, Russia, UK and US – already provides a forum for practical risk reduction among them, but it should develop a common agenda on human control, incident communication, and the treatment of AI-generated assessments in warning systems. In Europe, the OSCE could examine how existing confidence- and security-building measures should address AI-enabled conventional exercises and incidents that carry risks of cross-domain escalation. It should not seek disclosure of sensitive nuclear systems. Above all, states must recognise that preserving decision time is a shared interest. Implementation will also require experts able to work across nuclear policy, AI, and diplomacy. The Declaration is right to emphasise the next generation, but education must be accompanied by mentorship, access to policymakers, and opportunities to test technical findings against operational and political realities.

The Rome Declaration is a useful starting point, but its practical value will depend on what follows. Governments and international organisations must turn its principles into rules and risk reduction mechanisms, while technology developers must make relevant systems transparent. The immediate objective is to ensure that technological change does not remove the time, human authority, and political control on which effective nuclear crisis management depends.

The European Leadership Network itself as an institution holds no formal policy positions. The opinions articulated above represent the views of the authors rather than the European Leadership Network or its members. The ELN aims to encourage debates that will help develop Europe’s capacity to address the pressing foreign, defence, and security policy challenges of our time, to further its charitable purposes.

Image credit: Andrey Suslov / Alamy