On September 29, 2026, U.S. President Donald Trump and representatives of six major AI companies reached a voluntary agreement for company-led oversight of the safety of frontier AI models.
The published "White House Accord on Super Intelligence" commits the companies to a four-tier system: internal controls, in-house oversight, independent external audits, and board oversight. However, it does not specify how often external audits must be conducted, nor whether the audit results must be made public.
On the same day, Trump also signed an executive order directing the executive branch to use the term "Super Intelligence (SI)" in place of "Artificial Intelligence (AI)." But corporate self-regulation on safety and an executive order changing terminology used within government agencies differ in both scope and legal standing.
To judge how effective the accord will be, it helps to separate two questions: how far safety problems and audit findings are reported inside a company, and how much of those findings users and researchers outside the company can actually see.
The four-tier oversight structure set out in the accord
The signatories reportedly include representatives of Google, Anthropic, Meta, OpenAI, xAI, and NVIDIA, along with Trump.
Executives from other companies also attended the White House meeting, and attending the meeting is not the same as signing the accord. Not every participating company signed.
The full reprint of the published accord describes the four tiers of controls and audits. We have not been able to directly verify the original document that Trump posted on social media.
The accord applies to companies that train frontier models and make them available for actual use.
Companies must first monitor model capabilities that could lead to cyberattacks or biological and chemical threats. They must also check whether models behave in line with human intent and whether they might unintentionally break into or gain access to technical systems.
They must then oversee whether these safeguards actually work, through the following four tiers.
| Tier | Body or mechanism | Role required by the accord |
|---|---|---|
| 1 | Internal controls | Monitor model capabilities and behavior during training and after release, and prevent unintended intrusion or access |
| 2 | In-house oversight team | Confirm that controls, monitoring, and detection mechanisms work as intended, and correct any problems |
| 3 | Independent external auditors and evaluators | Assess from a third-party standpoint whether controls, monitoring, and detection mechanisms work |
| 4 | Independent board committee | Receive reports from in-house teams and from internal and external audits, and oversee whether identified problems are corrected |
The table organizes the four items in the published accord by responsible party and role.
Because it builds in independent external audits and requires the results to be reported up to the board, the system differs from one in which a development team tests its own model and judges it problem-free. Its distinguishing feature is a path that connects technical safety issues to management-level oversight.
However, these four tiers are what the signatories have promised to implement going forward. They do not show that all six signatories have already completed external audit contracts or put board committees into operation.
The independent board committee is also an internal body that oversees each company's management; the accord does not create a new government review body.
Whether the audit findings and problems received there are also made public will greatly change how far customers and researchers can verify the safeguards.
External audits will happen, but results are not necessarily disclosed
As far as the published accord shows, it does not specify how many months apart external audits must occur, who selects the auditors, or what conditions a model must meet to be judged as having cleared a safety standard.
It also contains no deadlines for correcting problems that are found and no procedure for publicly releasing audit results. We could not find any provision imposing legal penalties on companies that violate it, and it is largely a voluntary joint commitment by the companies.
That said, the absence of such provisions in the joint accord does not mean each company lacks its own safety standards or audit systems.
Rules that individual companies have set separately need to be distinguished from what multiple companies have jointly committed to in this accord.
For example, if an external audit finds a problem, the scope of external verification differs greatly depending on whether the information stops at a report to the board or is also disclosed to customers and the general public.
What the accord clearly establishes is a path for finding problems inside the company, reporting them to management level, and correcting them.
This is not the first voluntary safety commitment by AI companies.
The commitments OpenAI published on July 21, 2023 included conducting tests in which internal and external experts probe a model's weaknesses, as well as a policy of reporting on capabilities, limitations, and safety evaluations when releasing major new models covered by the commitments.
The new accord specifies an oversight structure that reaches the board, but it does not require an equivalent public-facing report.
It is therefore not possible to judge simply by counting safety measures whether regulation has become stronger or weaker than before. Nor can comparing the documents alone show how far the 2023 commitments have actually been carried out.
A mechanism for finding and correcting problems inside a company and disclosure that lets users and society check the results need to be evaluated as separate mechanisms.
Voluntary corporate audits and government model evaluations are separate systems
Model evaluation by the U.S. government and AI companies was under way before this accord.
According to an official announcement from the National Institute of Standards and Technology (NIST) on May 5, the Center for AI Standards and Innovation (CAISI) is working with Google DeepMind, Microsoft, and xAI on pre-release evaluations and research on models.
This is a model evaluation conducted jointly by companies and the government, and it is a separate mechanism from the external auditors and board committees described in the accord.
In addition, Section 3 of the June 2 executive order directs the relevant agencies to design a voluntary framework under which frontier models meeting certain criteria would be evaluated for cyber capabilities and the government would gain access to them before release.
The period of government access is set at up to 30 days before the date on which the company plans to make the model available to "other trusted partners."
This is not a system requiring every company to undergo government review 30 days before public release.
The order also states explicitly that the framework must not be interpreted as a basis for creating a mandatory government licensing system or a pre-release approval system for models.
The September 29 executive order, meanwhile, mainly deals with the terminology used for AI within the federal government.
The two can be compared as follows.
| Item | June 2, 2026 executive order, Section 3 | September 29, 2026 executive order, Sections 2 and 3 |
|---|---|---|
| Main subject | Frontier models meeting certain criteria and their developers | Terminology used in executive branch documents and official communications |
| Main instruction to agencies | Design model capability evaluations and a voluntary pre-release access system | Rename AI as SI and draft legislation for a new federal definition |
| Main conditions | Access of up to 30 days is voluntary; no mandatory licensing or pre-approval system | For now, use the definition of AI in existing law; past documents need not be changed |
The table sets out the subjects and the actions required of government agencies under Section 3 of the June 2 order and Sections 2 and 3 of the September 29 order.
Government evaluation of models, voluntary audits by companies, and changing the name used in government agencies are each separate mechanisms.
Signing this voluntary accord also does not relieve companies of their responsibilities under existing law.
Section 4 of the June 2 order directs the Attorney General to prioritize enforcement of applicable federal criminal laws against illegal computer intrusion, system damage, and similar acts carried out using AI.
A policy of not creating a new government pre-approval system for AI models and a policy of cracking down on violations of existing law are compatible.
Renaming AI as "SI" is not a finding of superhuman AI
The September 29 executive order calls for using "Super Intelligence" and its abbreviation "SI" in executive branch official communications, websites, reports, and similar materials, to the extent permitted by law.
However, previously issued rules and executive orders need not be rewritten. The order also does not require changes to existing contracts, grants, or historical documents.
The key point here is the definition in Section 3.
Unless another law sets a different definition, the scope of "SI" for now includes the technologies and systems defined as "artificial intelligence" in 15 U.S.C. 9401(3).
In other words, it is not the case that only models that have passed certain capability tests and reached human-surpassing intelligence are being certified as "SI."
Even if the name "SI" comes into use in federal documents, that alone does not represent a technical assessment that a model has surpassed human intelligence.
The order further directs the Assistant to the President for Science and Technology to submit to the President, within 60 days, proposed legislation to establish new federal definitions of "Super Intelligence" and "SI."
That process would consider whether to change, broaden, or replace the existing legal definition of AI, and would also present draft amendments to related laws as needed.
The 60 days is the deadline for submitting the legislative proposal. It does not mean that a new law will be enacted and take effect within 60 days.
At this point, therefore, changing the name from "AI" to "SI" and the technology itself reaching so-called superintelligence need to be considered separately.
What is made public after signing will determine the accord's effectiveness
Under the accord, the participating companies say they will continue to meet regularly to develop standards and practical methods for improving safety.
It also touches on the possibility of eventually formalizing such measures as laws or regulations. In other words, the document does not position this self-regulation as the final system.
At the same time, it does not give a timeline for legislation or the specific content of any law.
Going forward, the factors for judging whether the accord actually works will be who is chosen as external auditors, which models are evaluated and at what stage, how often audits are conducted, and what action is taken after problems are found.
If it is shown what was examined in which model in an external audit, what problems were found, and how they were reflected in the release or provision of the model, customers and researchers will be able to evaluate the claim that "an audit was conducted" more concretely.
If, on the other hand, audit results are not published and the process ends with a report to the board, all that can be confirmed from outside is that an oversight structure exists.
The internal oversight system working and that function being verifiable from outside are not the same thing.
Whether the six companies' signatures become a basis for judging the reliability of AI services will become clearer depending on how much is disclosed about the scope and results of audits and the corrective actions taken when problems are found.