Trump's voluntary AI safety pact lacks effective oversight until auditors gain authority

Trump and AI companies presented a voluntary safety pact with four layers of oversight on frontier AI on 29 September 2026. Treat that promise as a supplement, not a replacement for public oversight: without common standards, secured independence and authority to stop training, control remains dependent on the companies themselves.

According to Reuters, Trump said on 29 September that technology executives had agreed voluntary AI standards after a meeting at the White House. Among the reported signatories was Dario Amodei of Anthropic. The companies promised internal controls, independent auditors and measures to prevent systems from gaining access to technical environments in unintended ways. Trump described the arrangement as a form of protection, while the commitments remained voluntary.

The substance is set out in the White House Accord on Superintelligence. In our assessment, the core of this news is not that a safety pact exists, but the question that underlies it: will the announced external evaluators be given enough independence and authority to enforce anything? As long as that answer is missing, the four-layer structure is mainly an internal discipline instrument.

Which four layers of oversight are in the 29 September pact?

The pact describes four layers that together are meant to secure the safety of frontier models during training and rollout:

• Internal technical controls during training and deployment.
• An internal team that verifies whether those controls and the monitoring actually function.
• Independent external auditors or evaluators who test the whole.
• An independent committee of the board that receives reports and oversees the remediation of identified problems.

The document explicitly names these four points as voluntary commitments by the sector, not as binding legal obligations. That distinction determines how much weight the structure carries. For anyone who recognises a layered set-up in which every safety claim has to be tied to evidence, this aligns with the idea of a safety case that ties safety claims to evidence before a run proceeds.

Why does the voluntary nature make the promises hard to enforce?

The accord does not clearly establish a common evaluation standard, auditor-accreditation system, disclosure duties, remediation deadlines or sanctions, although it calls for companies to work toward standards and best practices. POLITICO described on 15 September that AI companies increasingly advocate independent third-party review, while lawmakers and critics doubt whether evaluators chosen by the companies themselves are sufficiently independent. In that piece, conflicts of interest, regulatory capture, the qualifications of evaluators and the absence of authority to enforce changes emerge as core problems.

The Brookings Institution argues that AI safety oversight requires explicit choices about applicable standards, covered models, disclosure, qualifications of evaluators, funding and enforcement authority. Rules that a regulated company effectively draws up itself and executes itself may, according to Brookings, shift from self-regulation to self-interest. In our assessment, that is the weakest spot of the pact: it describes who looks, but not clearly against what standard the testing is done or what follows a negative finding. The same tension between voluntary promises and demonstrable execution is central to this article's assessment of AI governance and oversight, a theme explored further in our topic hub on AI governance and oversight.

What does the comparison with banking supervision teach about missing authority?

CNBC described on 16 September that proposed embedded AI evaluators may be given extensive access, but have no authority to stop training or rollout. Experts compare that unfavourably with banking supervision, where supervisors can require corrective measures, restrict growth or close an institution. CNBC names the independence of evaluators, control over scope, publication rights and conflicts of interest as unresolved issues.

The difference can be made sharp in a short list:

• Banking supervision: according to the experts cited by CNBC, supervisors can require corrective measures, restrict growth or close an institution.
• The pact: cooperation with external evaluators, but no stated authority for them to pause training or deployment and no explicit legal safeguards for their independence.

Our editorial conclusion: access without enforcement mainly delivers reassurance, not accountability. That is the distinction between a promise and evidence: an assurance becomes more meaningful when it is backed by a traceable record of testing, findings and remediation.

What evidence must an AI provider be able to show before the promise is worth anything?

For professionals who work with confidential or high-value information, a self-regulation promise is only meaningful once the provider can produce verifiable evidence. In our assessment, the following checklist is useful when assessing a provider:

1. Demonstrable model tests: which evaluations were carried out, by whom and against which standard.
2. Independent access: did the external evaluator actually have access to relevant evidence, or only to what the company chose to show.
3. Open findings: which problems were identified and not yet resolved.
4. Remediation with deadline: who is responsible for remediation and within what timeframe.
5. Accountable decision-makers: which board committee receives the reports and oversees remediation.

Without these elements, the question that POLITICO and Brookings put at the centre remains open: who chooses the auditor, which standard do they apply and what follows a negative verdict. A comparable principle applies to any reporting framework, including OpenAI's reporting framework for model misalignment and the US assessment framework and duty of care for frontier models: reporting a category has practical value only when there is a traceable record, an accountable owner and a defined response.

The conclusion is sober. The 29 September pact shows that self-regulation can yield useful internal discipline and audit evidence. But without public standards and enforceable consequences, it cannot replace government oversight. For the reader that means: assess a provider on what is demonstrable, not on the promise itself.

Sources: The article relies on Reuters and the published text of the White House Accord on Superintelligence, supplemented by analyses from POLITICO, CNBC and the Brookings Institution.