Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Scan for outdated or missing drivers - takes under a minute3Repair Windows errors before they cause bigger problemsiTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Voluntary AI safety pledges can give companies a shared plan for testing, security and disclosure, but a pledge is not the same as independent verification or an enforceable legal duty. The central risk is that companies often retain substantial discretion over what counts as dangerous, what evidence is enough and what the public gets to see. That does not make every commitment empty; it means the value of a promise depends on whether outsiders can check it and whether failure has consequences.
This question is especially relevant to frontier AI commitments: policy frameworks aimed at the most capable AI systems. They are not a complete account of every AI system or every kind of harm.
What have AI companies promised to do?
The 2023 White House commitments
In July 2023, seven leading companies made voluntary commitments through a White House initiative. The archived commitment document covers risk testing, information sharing, cybersecurity, ways to identify AI-generated content, public reporting, research into societal risks and beneficial applications.
The 2024 Seoul commitments
The AI Seoul Summit commitments set out a more explicit structure for frontier AI safety. Signatories committed to assess risks during development and deployment, establish thresholds for risks they consider intolerable, describe mitigations and specify what they would do if a threshold were reached. They also committed to internal accountability and public transparency, with room to limit disclosure for security or sensitive commercial reasons.
#1 Best Overall
The Seoul text says that, in an extreme case, signatories commit “not to develop or deploy a model or system at all, if mitigations cannot be applied to keep risks below the thresholds.” That is a strong stated response, but its practical force depends on how thresholds are set and how anyone can determine whether they have been crossed.
The populations covered by different public figures should not be conflated: the White House initiative involved seven companies in 2023; the Seoul initial signatory list included 16 organisations, with four added later; and the 2026 International AI Safety Report says at least 20 developers had published public transparency reports in the G7 reporting context. These are different groups and measures, not a count of companies proven to comply.
What could go wrong when companies set and assess their own rules?
Thresholds can reflect company judgment
The Seoul commitments ask organisations to set risk thresholds and explain how they chose them. That creates a framework, but it leaves consequential choices with the developers unless governments, independent experts or other outside actors can meaningfully shape or review those choices. Two companies could describe similar risks yet set different thresholds or require different evidence before acting.
Public disclosure may not be enough to verify a claim
The Seoul text supports public transparency while allowing limits where disclosure could increase risk or reveal sensitive commercial information disproportionately. Restricting detailed security information can be sensible; publishing it indiscriminately could create its own risks. But if important evidence is available only inside a company, outsiders may struggle to compare safety claims or judge whether a mitigation works. The commitments contemplate more detailed sharing with trusted actors, such as governments or appointed bodies, as one way to address that gap.
Internal oversight can face competing incentives
A company can assign responsibility for safety reviews and still face commercial and competitive pressure around release timing and capability. That is a governance risk, not proof that every company’s safety staff are compromised or that every release decision is unsafe. The relevant question is whether review functions have enough authority, information and independence to influence decisions when commercial goals and safety assessments diverge.
A written framework is not evidence that safeguards work
A framework can describe intended tests and procedures without demonstrating that mitigations are effective in practice or that incidents will be prevented. Public reports are useful evidence of what a company says it does; they do not necessarily expose all of its internal operations or establish real-world safety outcomes.
Rank #3
Commitments can lag behind a changing technology
The UK guidance on frontier AI safety processes describes those processes as emerging and says the guidance is not final. The Seoul commitments also allow approaches to evolve with the science, alongside public updates explaining changes. Adaptability matters, but it makes revision histories important: readers need to know what changed, when, and why a company departed from an earlier approach.
Recommended Free Tools
What does public evidence show—and what does it not show?
A 2025 arXiv preprint by Jennifer Wang, Kayla Huang, Kevin Klyman and Rishi Bommasani assessed publicly disclosed company behavior against a rubric based on the eight White House commitments. Under that rubric, the authors reported a 52% average overall score across companies, a 17% average score on model-weight security, and 11 of 16 companies scoring 0% on model-weight security.
Those figures describe the authors’ assessment of public disclosures under their chosen rubric. They are not an official audit, a legal finding, or a direct measure of real-world harm; a low disclosure score does not by itself establish what a company does privately, and a high score would not prove that its controls prevent harm. The authors’ findings support the narrower point that claims are more useful when disclosures are proactive and verifiable.
The 2026 International AI Safety Report gives broader context: it describes frontier safety frameworks as a prominent organisational approach, notes that the EU General-Purpose AI Code of Practice is voluntary in its current form, and records at least 20 developers publishing G7 transparency reports. The report also discusses researchers’ argument that third-party auditing, verification and standardisation could strengthen risk management. Publication of a report, however, is not itself proof of compliance or effective safeguards.
How do the main accountability approaches differ?
| Approach | Who sets requirements and thresholds? | Who evaluates the work? | What is disclosed? | Is it mandatory, and what follows from failure? |
|---|---|---|---|---|
| Internal company framework | The company sets its own framework and thresholds. | Company staff and governance processes; external scrutiny depends on the arrangement. | Determined by company policy and any applicable commitments or rules. | The framework itself is not necessarily binding law; consequences depend on other commitments and applicable law. |
| Voluntary multi-company or government-backed commitment | Signatories agree to shared stated practices, while retaining discretion over important implementation choices such as thresholds. | Primarily the signatories’ own processes, with independent evaluations and government consideration encouraged in the Seoul text. | Public reporting is expected, with possible limits for security and sensitive commercial information; more detail may go to trusted actors. | The Seoul commitments are voluntary. The commitment alone does not establish the same duties or consequences as legislation; other applicable laws still matter. |
| Voluntary standard, such as NIST AI RMF | NIST provides a risk-management framework organisations can use to incorporate trustworthiness into design, development, use and evaluation. | Organisations apply the framework; the framework does not itself make an independent evaluator mandatory. | The framework does not itself require a particular public disclosure package. | Voluntary, not binding law. NIST released AI RMF 1.0 on January 26, 2023, a Generative AI Profile on July 26, 2024, and says the framework is being revised. |
| Independent evaluation | Evaluation criteria may come from commitments, standards, regulators or the evaluator’s remit. | A third party evaluates claims or systems, potentially improving verification of developer claims. | What becomes public depends on the evaluation arrangement and any security or confidentiality limits. | Evaluation can be recommended or required by a separate regime; it does not automatically create a consequence for failure. |
| Binding regulatory requirement | Requirements are set by the relevant law or regulator for its jurisdiction and scope. | Oversight and evidence requirements depend on that regime. | Disclosure and reporting depend on the applicable law. | Mandatory within the law’s scope; available consequences depend on the jurisdiction and specific requirement. |
No single row is a universal answer for every AI risk or jurisdiction. The practical difference is whether requirements are externally set, independently checked, visible to affected publics or authorities, and backed by a credible consequence.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Scan for outdated or missing drivers - takes under a minuteDriver Scan →Repair Windows errors before they cause bigger problemsFix Now →What would make a promise more accountable?
Government guidance and the commitments themselves point toward several ways to make voluntary pledges more testable, without treating recommendations as obligations that already apply everywhere.
- Independent evaluation and red-teaming: UK guidance says third-party evaluation can help verify safety claims. Seoul signatories committed to give appropriate consideration to independent third-party and government evaluations.
- Comparable public reports: Useful reporting should identify what was tested, which risks were considered, what thresholds apply, what mitigations were used, what remains uncertain and how the approach changed. That level of detail makes claims easier to compare while still allowing justified security limits.
- Trusted access where public disclosure is unsafe: Governments or appointed bodies may be able to examine details that cannot responsibly be published to everyone.
- Consequences for failure: NTIA’s 2024 recommendations, as described in the available official summary, call for independent evaluation and consequences when commitments are not delivered or risks are not properly managed. This is a policy recommendation, not evidence that every voluntary pledge already carries such penalties.
- Practical risk-management methods: NIST’s AI Risk Management Framework offers a voluntary way to incorporate trustworthiness into organisational processes. It can help structure risk work, but it is not a substitute for binding legal duties where those apply.
How should readers judge a safety pledge?
Look past the promise’s headline and ask whether it answers these questions:
- Are risks and thresholds defined clearly enough that another party could assess them?
- Does the company explain what evidence would trigger mitigation, delay, or a decision not to deploy?
- Can an evaluator independent of the developer inspect relevant evidence, including evidence not suitable for public release?
- Does public reporting state what was tested, what was not tested, what remains uncertain, and what changed?
- Is there a credible consequence if the company misses its stated threshold or fails to follow its own process?
Voluntary commitments can establish shared practices faster than legislation and help companies organise testing, security and reporting. Their weak point is assurance: without clear criteria, evidence outsiders can examine and credible follow-through, a reader may be able to see that a company made a promise without being able to tell whether it kept it.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
