On 13 July 2026, Pugwash convened a one-day consultation in Rome on “AI and Nuclear Weapons: Risks and Opportunities.” The workshop gathered 20 specialists from computer science, physics, international law, geopolitics, and industry to assess how AI affects nuclear risk and to future discuss Pugwash activities. The agenda focused on four broad areas:
- Overall challenges posed by AI
- AI in nuclear early-warning, intelligence assessment, and command-and-control systems
- AI in arms control: verification, compliance, and disarmament
- Mapping experiences of nuclear and other WMD arms control for future governance of AI
The report reflects the range of views expressed and should not be read as a consensus statement or formal policy position.
The consultation found that AI is already affecting the strategic environment, even though its military applications and limitations remain poorly understood. Civilian AI systems can be adapted for military purposes, including by being incorporated into intelligence analysis, and accelerating operational decisions. This creates risks of overreliance, automation bias, opaque reasoning, adversarial manipulation, and the gradual erosion of meaningful human control. Simply placing a person “in the loop” is not sufficient if AI systems determine which information is presented, frame the available options, or compress the time available for reflection. Meaningful human control requires that decision-makers understand the systems they use, can challenge their outputs, have access to independent assessments, and retain the authority and time needed to reject AI-generated recommendations.
These concerns are especially serious in systems related to nuclear weapons. AI may improve the speed at which large volumes of information are processed, but any time gained could be offset by the need to verify outputs and guard against excessive trust in automated assessments. The greatest danger may arise from the interaction between human and machine weaknesses: AI can amplify existing assumptions and biases, while decision-makers operating under pressure may be unable to recognize errors in complex or opaque systems. A human who formally approves a recommendation at the end of the process may therefore exercise only nominal, rather than meaningful, control.
The report also highlights the growing influence of private technology companies. Frontier AI firms increasingly shape the technology itself, how threats are understood, and what forms of restraint are treated as realistic. Competitive pressures, commercial incentives, government dependence on private systems, and the strategic framing of AI development as an international race all encourage rapid deployment. At the same time, companies often lack expertise in nuclear issues. Governments and international organizations similarly need a better understanding of how commercial AI systems are designed and deployed. Addressing this two-way knowledge gap will require closer cooperation between different expert communities.
AI may also create opportunities for nuclear risk reduction, arms control, and disarmament. It could assist with analysing satellite imagery, identifying possible nuclear tests, strengthening chains of custody, and combining information from multiple sources. AI-supported remote sensing may help supplement on-site inspections, particularly where access is restricted. However, verification systems must themselves be transparent and open to scrutiny. Inspectors and decision-makers need to understand the origin of the data, the analytical process, the uncertainty involved, and the possibility that states may conceal activities or deliberately mislead automated systems. Crucially, it must be emphasized that more capable technical tools cannot resolve political disputes when states are unwilling to cooperate or comply.
The consultation also cautioned that greater availability of information does not automatically produce greater stability. AI-enabled surveillance may create a situation in which states gain detailed information of their adversaries’ capabilities without negotiated access or consent. Although this could support verification, it may also expose vulnerabilities, encourage concealment and deception, threaten the survivability of nuclear forces, and increase pressure for rapid decision-making. Arms control should therefore not be understood simply as a means of obtaining more information. Its deeper purpose is to create predictability, support restraint, and reduce the risk of fear and miscalculation.
Traditional arms control models offer useful lessons but cannot be transferred directly to AI. AI is a general-purpose technology rather than a single weapon system, and much of its development takes place within private companies rather than state-controlled military programs. A single treaty or institution is therefore unlikely to be a sufficient form of governance. Instead, several layers of targeted rules addressing particular actors, applications, capabilities, and risks will likely be required. These could include legally binding prohibitions, political commitments, voluntary standards, confidence-building measures, export controls, and industry codes. The difficulty of verification should not prevent early agreement on core norms, including the principle that AI must never make the final decision to use nuclear weapons.
Future governance should focus on the weaponization and application of AI rather than attempting to regulate all AI development. Use cases of particular concern include those that could delegate the nuclear use decisions to AI, compress nuclear decision-making, undermine information integrity, weaken command-and-control systems, enable counterforce strategies, or increase the likelihood of accidental or deliberate escalation. Examples of possible risk-reduction measures include stronger authentication of orders and information, technical interlocks, visible and manageable failure modes, restrictions on silent software updates, agreements not to attack nuclear command-and-control systems, and greater transparency regarding the functions for which AI is used.
The report concludes that continued international and interdisciplinary dialogue is essential. Future work should move beyond broad studies of AI and nuclear command and control and concentrate on practical proposals for verification, transparency, confidence-building, and institutional design. Options identified include an AI security summit, an international coordinating or standard-setting organization, and a global industry association for AI safety.
Pugwash can contribute by convening experts who do not normally interact, building shared understanding across technical and policy communities, and developing more concrete approaches to managing the most destabilizing applications of AI. Indeed, all participants welcomed the opportunity of the convening by Pugwash of such an interdisciplinary group – in particular the inclusion of voices from industry – and encouraged the Leadership to pursue further work in this area.