Transparency From AI Labs and a U.S.–China AI Agreement

A new piece by Frazier in The Regulatory Review, "Transparency From AI Labs Is Key to a U.S.–China AI Agreement" (September 28, 2026), argues that AI labs should not wait for governments to write the rules and should proactively adopt the transparency practices that a future U.S.–China agreement would require. That can sound like a small procedural suggestion, but it is a claim about how trust gets built between rivals who can't see inside each other's most important systems. Below, I mark where I am summarizing Frazier and where I am adding my own analysis.

The rest of this article adds context the headline doesn't carry and works out what the argument means for businesses that depend on AI.

What is the argument, and why does it matter now?

The core of it is a sequencing idea. Governments negotiating over frontier AI face a basic problem: you can't agree to limits you can't verify, and you can't verify what you can't see. Nuclear arms control had inspectors, satellites, and physical facilities that were hard to hide. AI development has data centers, to be sure, but the thing being governed is mostly software, weights, training choices, and evaluation results. Much of that lives inside private companies.

So if labs begin disclosing more on their own, in standardized and comparable ways, they create the raw material that negotiators will need. They also create a track record. A diplomat can't credibly promise that American labs will publish safety evaluations if no American lab has ever done so in a consistent format. And a Chinese counterpart has little reason to believe a promise like that.

That sequencing is the strongest part of the case. Transparency here isn't a favor to the public or a public relations exercise. It is a precondition for the kind of agreement people keep saying they want.

Why is verification the hard part of any U.S.–China AI deal?

There is a long history of the two countries agreeing on principles and struggling on specifics. On November 16, 2024, President Biden and President Xi Jinping met in Lima, Peru, and the White House readout of the meeting (published on whitehouse.gov) said they affirmed the need to maintain human control over the decision to use nuclear weapons. It was a narrow statement, but it showed that the two governments could name at least one AI-related risk they both wanted to keep humans in charge of.

What that statement did not include was any way to check. No inspection regime, no shared reporting format, no agreed definition of what a covered system even is. That is the gap Frazier's piece addresses.

There are at least three things any verification scheme would need:

• A shared vocabulary for what counts as a frontier system, a dangerous capability, and a meaningful evaluation.
• A way to confirm that disclosed information is accurate, which usually means some form of third-party or reciprocal access.
• Enough prior practice that disclosure feels routine rather than like a concession.

Labs can't supply the second item alone, but they can make real progress on the first and third. Those two are where a proactive move would count.

What transparency practices already exist?

The proposal is more realistic because parts of it already exist, unevenly, in several places. Here is a rough comparison of the major regimes a lab operating internationally has to think about.

Look at the table and a pattern shows up. Both countries already ask something of AI providers. But the asks point in different directions. The Western regimes mostly ask developers to describe their own risk practices and, in California's case, to report serious incidents. The Chinese measures focus heavily on content, registration, and state oversight of what services say and do.

That difference matters for the idea of a shared standard. A U.S.–China agreement would need to sit in the overlap, and the overlap is thinner than either side's full rulebook. It probably includes things like incident reporting, disclosure of dangerous-capability evaluations, and some common description of how models are tested before release. These are also the very things labs could start doing in a standard format today.

Why should labs move first rather than wait?

There are a few reasons, and they pull in different directions, which is part of what makes this interesting.

The first is timing. Negotiated agreements are slow, and capability progress is not. If the only transparency that exists is whatever a treaty eventually mandates, the world gets visibility years after it was most needed. Voluntary disclosure compresses that timeline.

The second is shaping. Whoever has already built a workable disclosure practice has a strong hand in deciding what the eventual rules look like. A lab that has published a coherent safety framework and honest evaluation results can say to negotiators: this is feasible, here is what it costs, here is what it reveals and what it doesn't. A lab that has done nothing can only object.

The third is trust, and disclosure only builds trust when it is real. If a lab publishes a polished framework that never changes a decision, it is transparency theater and it will make a future agreement harder, not easier, because it teaches counterparties to discount everything they read.

A practical test follows: does the disclosure name conditions under which the lab would act against its own commercial interest, such as pausing a release or restricting access? A disclosure that commits to nothing costly gives a counterparty nothing to verify.

What are the obstacles?

There are real tensions, and some of them lack clean answers. For each, existing practice offers partial precedent.

Competitive secrecy

Labs treat model weights, training recipes, and some evaluation methods as their core assets. Asking them to disclose more invites a fair worry that disclosure leaks advantage to competitors, including foreign ones. The answer probably lies in disclosing outcomes and processes rather than the underlying assets: what was tested, what was found, what thresholds triggered what response, without publishing the weights or the recipe. A useful decision criterion is whether a counterparty could replicate the capability from the disclosure. If not, it is probably safe to publish.

Security risks of disclosure itself

Detailed descriptions of dangerous capabilities can serve as a map for misuse. Any serious disclosure standard has to separate information the public needs from information that should go only to vetted parties. That is a hard line to draw, and it will be drawn differently in Washington and Beijing. Tiered disclosure is the usual way to handle it: summary results go public, while sensitive detail goes to vetted evaluators under confidentiality. Government testing bodies already work this way. The U.S. AI Safety Institute (now the Center for AI Standards and Innovation, CAISI) and the UK AI Security Institute have run pre-deployment evaluations of frontier models under agreements with labs. The open question is whether a comparable arrangement could be made reciprocal across borders.

Asymmetric government roles

In the United States, labs are private companies that choose, or are required by statute, to disclose. In China, the relationship between companies and the state is different, and disclosure to the government is part of the regulatory design. A reciprocal arrangement has to work across those two models, and I am not aware of a worked-out design for doing so.

Domestic politics

Any U.S. administration that appears to be sharing AI information with China will face criticism, and any Chinese leadership that appears to be exposing its systems to foreign eyes will face its own. The advantage of lab-led transparency is that it lets a lot of groundwork happen before the political stakes of a formal deal arrive. The disadvantage is that it can also be accused of freelancing on questions of national security.

None of these obstacles defeats the idea, but they make the proposal a starting point rather than a solution.

What does this mean for businesses that use AI?

Most readers of this site are not running frontier labs, so it is fair to ask why any of this matters to them. There are three reasons.

First, the disclosures labs adopt flow downstream. The EU AI Act's Article 53(1)(b) requires general-purpose model providers to make information and documentation available to the providers that integrate their models, subject to the open-source exemption and the dates noted in the table above. If a baseline of transparency becomes normal at the lab level, businesses that deploy those models inherit better information about what they are using, and they can reasonably start asking for it. Procurement conversations change when documentation is expected rather than exceptional.

Second, geopolitics is a supply-chain issue. To map your exposure, build a simple inventory with one row per AI system you use and these columns: the model and its provider, where the provider is headquartered, the cloud region where inference runs, the chip vendor behind that region if the provider discloses it, and the fallback model if access is cut off. Rows with no fallback, or with a single region or vendor, are your concentration risks. A credible agreement, even a limited one, would reduce one category of uncertainty. Without one, that inventory tells you how much of your stack is exposed.

Third, what labs disclose becomes a reference point for what you can be asked about. If a model provider publishes its testing practices and incident history, customers, regulators, and insurers will eventually expect the businesses built on top of it to understand them. The standard set at the top tends to become the question asked at every level below.

Businesses can also ask their vendors for this transparency directly. Concrete steps:

• Add questionnaire items asking whether the vendor publishes a safety framework, which capability thresholds trigger added safeguards, whether it runs pre-deployment dangerous-capability evaluations, and how it reports and remediates serious incidents.
• Add a contract clause requiring the vendor to provide its Article 53 documentation (or an equivalent model card and technical summary) on request and to update it when the model changes materially.
• Add a notification clause requiring prompt notice of serious incidents, model deprecations, and changes to hosting region or subprocessors.
• Ask where the model is trained, hosted, and served, so the exposure inventory above can be filled in from vendor answers rather than guesses.

Why do small, repeated acts matter?

The underlying point follows from Frazier's argument. Agreements between rivals have rarely begun with a signed document. They usually begin with small, repeated, verifiable acts that make a signed document imaginable. Transparency from labs is one of those acts. It won't resolve the harder disagreements over export controls, military use, or who gets to set the global standards. But it could change what the two sides are able to say to each other.

What would a good lab-led disclosure look like?

A serious version of lab-led disclosure would include a short list of things that are checkable and comparable:

• A published safety framework that names the capability thresholds at which the lab will pause, restrict, or add safeguards.
• Results from pre-deployment evaluations of dangerous capabilities, described in enough detail that another party could tell what was actually tested.
• A practice of reporting serious safety incidents, with some indication of what changed afterward.
• Documentation for downstream users, which the EU already requires of general-purpose model providers.
• An honest account of what the lab does not know or could not test.

That last item matters most. A disclosure that contains no admission of uncertainty is usually marketing. The ones that say what hasn't been figured out are the ones that earn belief.

None of this requires waiting for a treaty. All of it would make a treaty easier to write and easier to trust.

What I am still unsure about

This article draws on Frazier's piece and the public record on existing regimes. No signed U.S.–China AI agreement exists as of this writing, so the idea may be ahead of the politics. Lab transparency may not change the strategic calculus for either government.

The alternative, though, is that the first serious look inside these systems happens after something goes wrong, which is a poor time to start building trust. Labs that can begin now have little reason to wait.

Last updated: 2026-10-14

Frequently Asked Questions

Why is transparency from AI labs important for a U.S.–China AI agreement?

Any agreement between the two countries needs a way to check that each side is keeping its commitments. Because frontier AI is mostly software and private research, outsiders can't easily see what is being built or tested. Lab disclosures about safety frameworks, evaluations, and incidents supply the information negotiators would need, and they build a track record that makes promises credible.

Do AI labs already have to be transparent under existing laws?

In part. Article 53 of the EU AI Act (Regulation (EU) 2024/1689) requires general-purpose AI model providers to keep technical documentation and share information with downstream providers. California's SB 53, signed in September 2025, requires large frontier developers to publish risk frameworks and report certain critical safety incidents. China also imposes requirements on generative AI services, including the Interim Measures effective August 15, 2023. The regimes differ in what they emphasize.

Is there currently a U.S.–China AI agreement?

I'm not aware of a signed comprehensive agreement. The most visible joint statement I can point to is the November 2024 meeting in Lima, where President Biden and President Xi affirmed that humans, not AI, should control decisions on the use of nuclear weapons. That was a narrow statement and did not establish a verification mechanism.

How does AI lab transparency affect businesses that use AI?

Disclosures made at the lab level flow downstream to the companies building on those models, giving them better documentation for procurement, risk assessment, and customer questions. A credible international framework would also reduce some geopolitical uncertainty in AI supply chains, and businesses can encourage the trend by asking their vendors for this information.

What are the main obstacles to lab-led AI transparency?

The main obstacles are competitive secrecy around model weights and training methods, the security risk of publishing details about dangerous capabilities, differences between how U.S. and Chinese companies relate to their governments, and domestic political sensitivity about sharing AI information across the rivalry. Disclosing processes and outcomes rather than underlying assets is one way to ease the first two.