At the White House in July 2023, seven leading US AI companies including OpenAI, Anthropic, Google, Meta, and Microsoft committed to a set of voluntary safety, security, and transparency measures. At the Bletchley AI Safety Summit in November 2023, twenty-eight countries signed a declaration acknowledging shared concern about frontier AI risks. At the Seoul AI Safety Summit in June 2024, sixteen AI companies signed the Frontier AI Safety Commitments, covering safety thresholds, model evaluations, and information sharing.
These are significant political achievements. They represent the first time that the leading developers of advanced AI systems have publicly accepted that their work poses serious risks and that some form of governance is appropriate. Dismissing them would be wrong. But accepting them as adequate would be a more costly mistake.
What voluntary commitments actually require
The Frontier AI Safety Commitments signed at Seoul are the most substantive of the major pledges. Signatories committed to: identifying safety thresholds at which they would not deploy or continue training a model; conducting and publishing evaluations of their frontier models against those thresholds; sharing information about safety incidents and evaluations with governments and other AI companies; and investing in safety research. These are genuine commitments in the sense that they require visible actions that can be observed and commented on publicly.
What they do not require: independent external verification of whether thresholds are being met; binding consequences for non-compliance; any mechanism for enforcement against a signatory that falls short; and participation by companies that choose not to sign. The commitments are aspirational statements of intent by companies that retain full discretion over how to interpret and implement them.
This is the structure of a voluntary commitment, and the distinction from a binding obligation is not subtle. A voluntary commitment binds the party morally and reputationally. A binding legal obligation binds them in law, with formal consequences for non-compliance. The gap between moral and legal obligation is precisely where the most consequential safety decisions get made under competitive pressure.
The pattern from other high-risk industries
Voluntary corporate safety commitments have preceded binding regulation in most major high-risk industries, and the pattern of what those commitments did and did not accomplish is fairly consistent.
In financial services, the major banks developed industry-led risk management codes and practices throughout the 1990s and early 2000s. These codes were in place and nominally in operation when the financial crisis of 2007 to 2009 revealed that voluntary risk management had failed to prevent systemic risk accumulation. The Financial Stability Board, the Basel III framework, and expanded mandatory disclosure requirements followed. Voluntary commitments had improved baseline practices but were insufficient at the tail risks where the costs of failure were highest.
In the chemical industry, the Responsible Care program was launched in 1985 following a series of major chemical plant accidents, including Bhopal in 1984. Participating companies committed to improving safety, health, and environmental performance and to reporting transparently on their progress. Responsible Care improved industry safety culture and reduced the incidence of routine accidents over subsequent decades. It did not prevent significant incidents and was eventually supplemented by binding regulatory requirements. The program's own assessment found that its greatest limitation was the absence of independent verification.
The chemical industry's Responsible Care program, launched in 1985, is the closest historical parallel to the current AI safety commitments: a voluntary industry pledge covering safety, transparency, and information sharing. The program's own assessment after two decades found that its primary limitation was the absence of independent verification of reported performance. The same limitation is structural in the current AI safety commitments, and it produces the same gap: commitment to outcomes without any mechanism to confirm whether those outcomes are being achieved.
Nuclear power plant safety in the United States was initially governed primarily through self-regulation and industry standards. The Three Mile Island accident in March 1979 demonstrated that voluntary industry safety culture was insufficient for a technology whose failure modes could affect large populations. The Nuclear Regulatory Commission's mandatory oversight framework, with independent inspections, mandatory incident reporting, and binding operational standards, followed. The industry fought mandatory regulation at the time; it is now generally credited with producing the safety culture improvements that voluntary approaches had not achieved.
The competitive pressure problem
Voluntary commitments have a structural vulnerability that binding regulation does not: the competitive pressure problem. If a commitment creates genuine cost, such as slowing development timelines, requiring expensive safety evaluations, or excluding high-revenue applications, then any company that does not sign the commitment gains competitive advantage by avoiding that cost. Companies that do sign face a choice between relaxing their interpretation of the commitment to remain competitive, or absorbing the cost and falling behind.
This dynamic is not hypothetical in the AI sector. The companies that signed the Frontier AI Safety Commitments at Seoul are competing directly with companies and state-sponsored programs in China that signed no equivalent document. They are also competing with each other, and the commitments they signed are sufficiently broad in language that reasonable interpretive disagreements about what compliance requires are inevitable.
Binding regulation solves this by creating a floor that all competitors must meet, eliminating the competitive advantage from non-compliance. This is why binding regulation produces safety outcomes that voluntary commitments generally cannot: it removes the incentive structure that makes safety shortcutting attractive.
What voluntary commitments accomplish well
The most valuable function of voluntary commitments in the history of industrial safety governance has been norm establishment. Voluntary commitments create a record of industry acknowledgment that a problem exists and that certain practices are appropriate responses. This record becomes the baseline from which binding regulation works.
When regulators or legislators come to AI governance, the Frontier AI Safety Commitments will have established that leading companies accept the legitimacy of safety thresholds, model evaluations, and information sharing as governance concepts. Industry will be in a weaker position to argue that these requirements are technically infeasible or commercially inappropriate, having already committed to them voluntarily. In this sense, voluntary commitments are a first step toward binding governance, not a substitute for it.
Voluntary commitments also create institutional infrastructure. The Seoul commitments required companies to develop internal processes for safety threshold setting and model evaluation. These processes, and the institutional knowledge embedded in them, are the foundation on which a binding verification regime can build. A regulatory body overseeing mandatory AI safety evaluations will be able to draw on existing company practices rather than designing compliance infrastructure from scratch.
"Voluntary commitments perform a useful function in the early stage of governance development, and the Bletchley and Seoul commitments have done something real. But the history of every major high-risk industry shows the same thing: voluntary commitments improve baseline practices and fail at tail risks. Tail risks are exactly what we are governing for with superintelligence."
What binding governance requires that voluntary commitments cannot provide
Four elements distinguish binding from voluntary governance, and all four matter for frontier AI.
Independent verification. A voluntary commitment is self-reported. A binding obligation is subject to third-party verification, whether by a regulatory body, an international inspection agency, or an independent auditor with legal standing. Self-reporting produces accurate information when reporting organizations have no incentive to misrepresent. In competitive markets, that condition is rarely met.
Defined consequences for non-compliance. Voluntary commitments create reputational cost for non-compliance, which is real but limited. Binding obligations create legal consequences: fines, license revocations, market access restrictions, or treaty-based penalties. The credibility of a governance framework depends on the credibility of its consequences.
Coverage of non-signatories. The most consequential AI development might come from companies or governments that never sign a voluntary commitment. Binding national regulation and international treaty obligations create governance that applies to actors regardless of whether they consent in advance. A voluntary commitment framework leaves the actors most likely to create serious risks outside the framework if they choose to be.
Stability across leadership changes. Voluntary commitments bind the current leadership of a company. When leadership changes, board composition changes, or competitive pressure intensifies, the commitment has no legal force to survive those changes. Binding regulation is embedded in law and survives ordinary organizational change.
The voluntary commitments made at Bletchley and Seoul are genuine achievements that should be built on. Building on them means acknowledging honestly what they accomplish and what they cannot, and using the political and institutional momentum they represent to pursue binding governance that fills the gaps they leave open.
Common questions.
What are the main voluntary AI safety commitments that have been made?
The most significant are: the Bletchley Declaration (November 2023), signed by 28 countries at the UK AI Safety Summit, acknowledging shared concern about frontier AI risks; the Frontier AI Safety Commitments (June 2024), signed by 16 AI companies at the Seoul AI Safety Summit, covering risk assessment, safety thresholds, and information sharing; and the White House voluntary commitments (July 2023), signed by seven leading US AI companies covering safety, security, and transparency. These are political declarations, not legally binding instruments, and they carry no formal consequences for non-compliance.
Have voluntary corporate commitments worked in other high-risk industries?
The record is mixed. In financial services, industry-led risk management codes failed to prevent the 2008 crisis. In the chemical industry, the Responsible Care program improved baseline safety culture but was eventually supplemented by binding requirements, and the program's own assessment identified independent verification as its primary gap. Nuclear safety moved from voluntary to mandatory oversight after Three Mile Island. The consistent pattern is that voluntary commitments improve routine operations but fail at tail risks, which is precisely where governance matters most for AI.
What is the competitive pressure problem with voluntary AI commitments?
If a commitment creates genuine competitive cost, a non-signatory gains advantage by avoiding it. This puts signatories under pressure to relax their interpretation of commitments or absorb the competitive disadvantage. Binding regulation eliminates this by setting a floor all competitors must meet. The Frontier AI Safety Commitments are observed by companies competing directly with Chinese programs that signed no equivalent document, and the language of the commitments is broad enough that competitive pressures will shape how companies interpret their obligations over time.
Do voluntary commitments play any useful role in AI governance?
Yes, as a norm-setting step toward binding governance. Voluntary commitments create a record of industry acknowledgment that certain practices are appropriate responses to real risks. This makes it harder for industry to argue later that binding requirements are technically infeasible or commercially inappropriate. The Seoul commitments have established that safety thresholds, model evaluations, and information sharing are concepts the leading AI companies accept in principle. Binding governance can build on this foundation rather than starting from zero.