Rome Declaration Calls for AI Nuclear Safeguards

For practitioners, the Rome Declaration connects familiar AI governance concerns, including human oversight, adversarial cyber risk, and reliable information environments, to nuclear command-and-control scenarios where errors can carry exceptional consequences. Industry context: work on comparable high-consequence systems commonly treats model performance as only one part of safety, alongside authority boundaries, monitoring, communications integrity, and escalation procedures.

The Global Nobel Laureates Assembly adopted the Rome Declaration for an Unarmed and Disarming Peace on July 16 at Rome's Capitoline Hill, Vatican News reports. The assembly ran from July 14 through July 16, with sessions at Borgo Laudato Si' in Castel Gandolfo and a concluding session in Rome. Vatican News reports that more than 200 figures and representatives of research institutions participated.

The official declaration states that AI embedded in nuclear systems can leave less time for, or replace, human judgment during a crisis. It also identifies AI-enabled cyber attacks against critical civilian and strategic infrastructure, and information warfare that can undermine the shared factual basis needed for dialogue, verification, and restraint.

Pugwash, which served as an organizing partner, reports that conference panels and working groups examined AI applications in military systems, nuclear command and control, autonomous weapons, and escalation at "machine speed." Pugwash also reports that a recurring theme was how to define meaningful human control when AI affects information integrity and reliable communication.

The declaration further states that advanced AI capabilities, computing resources, and data infrastructure are increasingly concentrated among a small number of countries and corporations. It frames that concentration as a power asymmetry with limited incentives for cooperation.

For practitioners, the document is not presented in the sources as a technical standard or enforceable regulation. Its relevance is in the risk framing: teams building, evaluating, or supplying AI systems used near defense, critical infrastructure, cybersecurity, or communications workflows may encounter stronger demands for auditable human decision authority and failure-mode analysis.

Observed patterns in similar safety debates show that "human in the loop" is insufficient unless people have meaningful authority, usable time to intervene, and reliable situational information. The Pugwash account's focus on machine-speed escalation makes latency, alert design, automation thresholds, and degraded-mode behavior relevant engineering questions, not solely policy concerns.

According to Vatican News, University of Chicago professor Daniel Holz, founding director of the Existential Risk Lab, told participants that it was a time of "unprecedented danger" but that there were actions available to reduce risks from nuclear weapons and AI. Cardinal Baldo Reina stated that no machine or algorithm can decide existential issues, Vatican News reports.

The sources describe the Rome Declaration as a joint public appeal adopted by assembly participants, rather than as a treaty or regulatory instrument. Its immediate effect is therefore agenda-setting rather than creating mandatory requirements for model developers or infrastructure operators.