The Digital Geneva Convention is a proposal advanced by Brad Smith, President of Microsoft, in 2017 for a treaty-like framework that would commit governments not to attack civilian technology infrastructure. It was never formalized as a binding treaty, but its design — voluntary multi-stakeholder commitments binding by reputation rather than law — has been characterized as a template later adopted across AI governance, including the Bletchley, Seoul, and Paris declarations, the AI Safety Institute network, and voluntary frontier-lab frameworks.
Origin
Smith proposed a Digital Geneva Convention in 2017 in response to nation-state cyberattacks targeting civilian infrastructure, including WannaCry and NotPetya. The proposal called for an international agreement analogous to the 1949 Geneva Conventions, binding governments not to attack civilian technology infrastructure during peacetime.
Smith and Carol Ann Browne elaborated the idea in *Tools and Weapons* (2019), setting out six principles for tech-government partnership: (1) transparency, under which policies and practices are made public; (2) accountability for impacts; (3) privacy by design, built in rather than added afterward; (4) non-discrimination through binding fairness commitments; (5) security to safeguard users and infrastructure; and (6) ethics, framed as explicit commitments adopted even without a legal requirement.
Evolution into AI governance
Although the convention was never formalized as a treaty, its architectural features — voluntary multi-stakeholder commitments, enforcement by reputation rather than law, and participation by industry, government, and civil society — have been described as the model on which subsequent AI governance was built.
Several international declarations follow this pattern. The Bletchley Declaration (November 2023) was the first multilateral AI safety declaration; the Seoul Frontier AI Safety Commitments (May 2024) set out voluntary lab commitments; the G7 Hiroshima Code of Conduct (October 2023) established a developer code of conduct; and the Paris AI Action Summit Declaration (February 2025) was issued at the Paris summit, though the United States and United Kingdom declined to sign it, an outcome cited as illustrating the limits of the voluntary model.
The framework is carried institutionally by a set of bodies. The AI Safety Institute network — the UK AISI, the US AISI, and analogous institutes in Japan, Korea, and Singapore — forms a distributed technical-audit capability. The Frontier Model Forum serves as an industry coordination body, and the Partnership on AI provides broader stakeholder coordination.
Voluntary frontier frameworks operate inside individual labs as pre-commitments. The Anthropic RSP v3.1 and OpenAI Preparedness V2 have been mapped onto Smith's transparency, accountability, security, and ethics principles. The GPAI Code of Practice represents the EU's voluntary-pre-commitment track, and the Frontier Compliance Framework provides for explicit cross-lab coordination.
Cybersecurity Tech Accord and AI accords
Smith led the launch of the Cybersecurity Tech Accord in April 2018, initially comprising 34 technology companies that pledged not to assist government cyberattacks and to protect their customers. By 2024 the accord had grown to more than 100 signatories. It served as a precedent for the Munich Tech Accord on AI Election Integrity (February 2024), under which 20 technology companies agreed to combat AI-generated deceptive election content, and further AI-specific accords have been discussed.
Debates and assessments
Proponents identify several strengths in the voluntary model. Treaty negotiations can take decades, whereas voluntary frameworks can deploy within months. The frameworks can evolve as technology changes without a formal amendment process. Companies that co-design commitments are described as more likely to comply, and the model is credited with making rogue-state cyberattacks reputationally costly.
Assessments over 2019–2026 have also identified weaknesses. Because the commitments are non-binding, there is no enforcement and signatories can defect at low cost, as illustrated by the US and UK refusal of the Paris declaration. Authoritarian actors are largely excluded: Russia and China never signed meaningfully, so the framework has been described as protecting only the liberal-democratic perimeter. Critics point to a "washing" risk, under which companies can signal compliance without substantive change, and to state-actor evasion, noting that NotPetya-type attacks continued despite the framework.
The model has been discussed in relation to several adjacent frameworks. Compared with Suleyman's containment framework, Smith's voluntary approach has been characterized as "thin" containment and Suleyman's 10-point framework as "thick," with the two described as complementary rather than conflicting. In relation to the Brussels Effect, a voluntary convention combined with EU hard law has been described as potentially approximating a binding global floor. Framings around AI Sovereignty, by contrast, are described as weakening voluntary-multilateral mechanisms.
Status
The architecture is described as surviving but fragmenting. In the United States, the Trump administration's EO 14179 rescinded the Biden-era executive order, and America's AI Action Plan explicitly deprioritizes international safety commitments; the United States also declined to sign the Paris declaration. Chinese engagement remains selective: China has signed some declarations, including Bletchley and Seoul, but operates outside Western voluntary frameworks. The EU binding-law path, comprising the EU AI Act (Regulation 2024/1689) and the EU General-Purpose AI Code of Practice (Final Version, 2025), offers an alternative described as voluntary-with-binding-backstop rather than purely voluntary.
Relationships
- supports: Tools and Weapons — Smith and Browne (2019) — book source.
- supports: The Bletchley Declaration (AI Safety Summit, 1–2 November 2023) / Frontier AI Safety Commitments (Seoul, 2024) / G7 Hiroshima Code of Conduct for Advanced AI (2023) / Paris AI Action Summit Declaration (2025) — operationalizations.
- supports: International AI Safety Institute Network (INSAI / AISIN) / Frontier Model Forum / Partnership on AI (PAI) — institutional carriers.
- supports: Anthropic's Responsible Scaling Policy (Version 3.1) / OpenAI Preparedness Framework V.2 / EU General-Purpose AI Code of Practice (Final Version, 2025) — voluntary frontier frameworks.
- related: Brad Smith — entity page for originator.
- related: AI Safety Cases and Frameworks — broader concept this nests under.
- related: AI and Cybersecurity — original problem domain.
- contrasts-with: EU AI Act (Regulation 2024/1689) — mandatory-regulation alternative path.