On 21 July 2023 the White House convened seven leading AI companies and announced a set of voluntary commitments. The companies were Amazon, Anthropic, Google, Inflection, Meta, Microsoft and OpenAI. The commitments covered safety testing, content watermarking, transparency reporting, risk management and safety investment, under the theme of safety, security and trust.
The commitments have no legal force and no regulator. But they offer an observation window: where government has not yet legislated, which practices leading firms are willing to state publicly, and how those commitments travel down the supply chain.
What the commitments say
Public information indicates that the seven companies committed to safety testing before model release, watermarking or other provenance mechanisms for generated content, public safety and transparency reports, prioritising high-risk capabilities, strengthening safety investment and internal governance, and sharing safety information with government and academia.
Their common thread is making process visible. Safety testing is no longer only an internal step but a record that can be explained externally. Watermarking is not merely a technology but a mechanism for users to identify generated content. Transparency reports are not public-relations documents but a basis on which outsiders can ask questions.
The limits of self-regulation
The weaknesses are equally clear. First, coverage is limited to a few leading firms; many smaller developers are outside. Second, the commitments are abstract, lacking verifiable metrics and third-party audit. Third, there is no consequence for breach; companies can adjust wording without legal cost.
Treating voluntary commitments as regulation would therefore misjudge them. They are closer to an industry baseline, telling the market what minimally responsible practice looks like. Real constraint comes from other mechanisms: safety clauses in customer contracts, diligence lists of investors, platform access policies, insurance underwriting conditions and state-level legislative attempts.
Transmission to Chinese companies
Chinese AI companies may be affected even without direct US regulation. If a product enters the US market, uses US cloud services or works with US customers, contracts and diligence may ask whether the company conducts pre-release safety testing, provides content provenance and publishes safety reports.
The more realistic impact is in the supply chain. Upstream providers of foundation models, compute and toolchains may turn commitments into procurement conditions; downstream customers may write them into supplier codes of conduct. Companies need to locate themselves in that chain rather than look only at their place of incorporation.
What companies can borrow
First, establish a pre-release safety testing process with clear scope, standards and records. Second, design labelling and provenance for generated content across text, images, audio and video. Third, prepare a disclosable safety report covering capability boundaries, known risks and handling mechanisms. Fourth, translate commitments into contract terms, allocating safety responsibility with customers and suppliers.
These steps need not be justified by “compliance with US commitments”; they can serve as internal product governance standards. While rules remain fragmented, being able to explain one’s safety process is itself a form of market trust.
For AI product safety and launch documents, see our AI product launch documents service and cybersecurity compliance; for multi-jurisdiction differences, see multi-jurisdiction; for safety clauses in contracts, see commercial contracts.
This article is based on public information about the voluntary AI commitments announced by the White House on 21 July 2023. It is for reference only and does not constitute legal advice.