The White House Voluntary Commitments, July 21, 2023

---

The Context

The voluntary commitments were announced in the midst of the most intense period of AI policy activity in American history. In the six months before the July 21 announcement, a cascade of events had pushed AI to the top of the political agenda. ChatGPT had been released in November 2022 and had become the fastest-growing consumer application in history. GPT-4 had been released in March 2023. The Future of Life Institute’s pause letter (see B41) had been published on March 22, 2023, calling for a six-month moratorium on AI training. Sam Altman had testified before the Senate (see B89) on May 16, 2023, calling for federal regulation. The CAIS Statement on AI Risk (see B38) had been published on May 30, 2023, warning that AI posed an extinction risk.

The Biden administration, faced with this cascade, was under pressure to act. But the administration’s options were limited. Congress had not passed any significant AI legislation, and the prospects for legislation were, given the political dynamics of the US Congress, uncertain. The administration could, however, use its executive authority — through executive orders, through agency rulemaking, and through the bully pulpit — to shape the AI policy landscape. The voluntary commitments were, in some ways, an exercise of the bully pulpit — a way for the administration to bring the AI companies to the table, to extract public commitments, and to create a framework that could be built upon.

The approach was, in some ways, modelled on earlier voluntary commitments in the technology industry. The administration had, in 2022, secured voluntary commitments from social media companies on election integrity. The pattern — bringing companies to the White House, extracting public commitments, and using the publicity to create pressure for compliance — was familiar. The difference, in the case of AI, was that the technology was newer, the risks were less understood, and the regulatory framework was even less developed.

---

The Commitments

The seven companies agreed to eight voluntary commitments:

1. Red-teaming. The companies pledged to conduct internal and external security testing (“red-teaming”) of their AI systems before release, to identify and mitigate risks. Red-teaming — the practice of having people deliberately try to break or misuse a system, to find its vulnerabilities — was borrowed from cybersecurity, where it had been a standard practice for decades. The commitment to red-team AI systems before release was, by most accounts, a meaningful step — it acknowledged that AI systems could be dangerous, and it committed the companies to testing for those dangers before deploying the systems.

2. Information sharing. The companies pledged to share information about AI risks — including across the industry and with governments, academics, and civil society. The commitment was, in some ways, a recognition that the risks of AI were not specific to any one company — that a vulnerability discovered in one company’s system might be present in others, and that sharing information about risks could benefit the entire industry.

3. Investment in cybersecurity. The companies pledged to invest in cybersecurity measures to protect their AI models from being stolen or compromised. The commitment was, in some ways, a response to growing concerns about the security of AI models — concerns that were heightened by the fact that the models were valuable intellectual property, and that their theft or compromise could have serious consequences.

4. Content provenance. The companies pledged to develop technical mechanisms — such as watermarking — to help users distinguish AI-generated content from human-generated content. The commitment was, in some ways, a response to the deepfake concerns that had been growing since the release of ChatGPT — the concern that AI-generated content could be used to spread disinformation, impersonate people, and undermine trust in media.

5. Public reporting. The companies pledged to publish reports about their AI systems’ capabilities, limitations, and areas of appropriate and inappropriate use. The commitment was, by most accounts, the weakest of the eight — it did not specify what the reports should contain, how detailed they should be, or how often they should be published. The lack of specificity was, for critics, a sign that the commitment was more about the appearance of transparency than about transparency itself.

6. Prioritising safety in research. The companies pledged to prioritise research on the societal risks of AI, including bias, discrimination, and privacy.

7. Deployment of AI to address societal challenges. The companies pledged to develop AI systems to address societal challenges like cancer prevention and climate change.

8. Development of a watermarking system. The companies pledged to develop a robust system for watermarking audio and visual content to help identify AI-generated media.

On September 12, 2023, a second round of commitments was announced, with eight additional companies signing on: Adobe, Cohere, IBM, Nvidia, Palantir, Salesforce, Scale AI, and Stability AI. This brought the total to fifteen companies.

---

The Companies

The seven founding companies were, by most accounts, the right group. They included all the major American AI labs — OpenAI (the maker of ChatGPT and GPT-4), Anthropic (the maker of Claude), Google (the maker of Gemini, and the company that had developed the Transformer architecture), and Meta (the maker of LLaMA). They included the cloud providers that hosted the models — Amazon (AWS), Microsoft (Azure, and the company that had invested $13 billion in OpenAI). And they included Inflection AI, a smaller start-up that was, at the time, building its own large language model (Inflection was later absorbed by Microsoft in 2024, in a deal that reflected the rapidly changing landscape of the AI industry).

The inclusion of all the major labs was significant. It meant that the commitments covered, in effect, all of the frontier AI systems — the most capable models, the ones that posed the greatest risks. It also meant that the commitments were, in some ways, a coordinated industry position — a shared set of principles that all the major players had agreed to. The coordination was, by most accounts, a result of the administration’s behind-the-scenes work — the White House had spent weeks talking to the companies, and the commitments were the product of those conversations.

The absence of non-American companies was also significant. The commitments did not include any European companies (like Mistral AI, see B58), any Chinese companies (like DeepSeek or Baidu), or any Japanese companies. The commitments were, in this sense, an American initiative — a reflection of the fact that the frontier AI labs were, in 2023, overwhelmingly American, and that the regulatory conversation was, in 2023, overwhelmingly American as well.

---

The Significance

The voluntary commitments were significant as the first industry-wide AI safety agreement. They demonstrated that the major AI companies were willing to publicly acknowledge the risks of their technology and to commit — at least in principle — to addressing those risks. They also provided a framework that could be built upon, either through additional voluntary commitments or through binding regulation.

The commitments were particularly notable for the breadth of the companies involved. The seven founding companies included all the major American AI labs — OpenAI, Anthropic, Google, Meta — as well as the cloud providers that host their models — Amazon, Microsoft. The inclusion of Inflection AI, which was later absorbed by Microsoft in 2024, reflected the rapidly changing landscape of the AI industry.

The commitments were also significant as a political achievement. The Biden administration had, in the space of a few months, gone from a relatively hands-off approach to AI policy to a position of active engagement — bringing the companies to the White House, extracting public commitments, and setting the stage for further action. The voluntary commitments were, in this sense, a down payment — a signal that the administration was serious about AI policy, and a foundation for the more substantial actions that would follow.

---

The Criticism

The commitments were widely criticised for being voluntary and non-binding. Because they were not enforceable, there was no mechanism to ensure that the companies actually followed through on their pledges. There were no penalties for non-compliance, no independent verification, and no transparency requirements.

MIT Technology Review, in a July 2024 retrospective, noted that the commitments had produced “better red-teaming practices and watermarks, but no meaningful transparency or accountability.” The review found that the companies had made some progress on the technical commitments — like red-teaming and watermarking — but had made little progress on the transparency commitments, like publishing detailed reports about their systems’ capabilities and limitations.

The voluntary nature of the commitments also reflected the political reality. The Biden administration did not have the authority to impose binding rules on AI companies without congressional legislation, and Congress had not passed any significant AI legislation. The voluntary commitments were a way to make progress without legislation — but they were also a way to create the appearance of progress without the substance.

The commitments were also criticised for being shaped by the AI companies themselves. The companies had significant input into the content of the commitments, and the commitments reflected the companies’ preferences — particularly the preference for voluntary, flexible commitments over binding, specific rules. Critics argued that the commitments were designed to forestall more stringent regulation, not to ensure safety. The concern was, by most accounts, legitimate — companies that support voluntary regulation often do so because voluntary regulation is weaker than binding regulation, and because it gives them more control over the terms.

---

What Came After

The voluntary commitments were followed, three months later, by President Biden’s Executive Order 14110 on AI, signed on October 30, 2023. The executive order went further than the voluntary commitments — it required AI companies to share safety test results with the government, and it directed federal agencies to develop standards for AI safety. But the executive order, like the voluntary commitments, was limited by the absence of congressional legislation.

The executive order was rescinded by the Trump administration on January 20, 2025, along with much of the Biden AI policy framework (voluntary commitments, executive order, AI Safety Institute). The Trump administration signalled a preference for less regulation, and the voluntary commitments — which were already voluntary — became even less significant under the new administration. The contrast with the EU AI Act (see B86), which was binding and enforceable, was stark — and it illustrated, in vivid form, the difference between the American and European approaches to AI regulation.

---

The Legacy

The voluntary commitments were, in retrospect, a first step — an attempt to establish norms of responsible AI development in the absence of binding rules. They demonstrated that the AI industry was willing to acknowledge the risks of its technology, and they provided a framework that could be built upon. But they also demonstrated the limits of voluntary approaches — without enforcement, without penalties, and without transparency, the commitments were only as strong as the companies’ willingness to follow them.

The legacy is also, in some ways, a lesson about the difference between voluntary and binding regulation. The EU AI Act, which was being negotiated at the same time, took the opposite approach — it imposed binding rules, with significant penalties for non-compliance. The contrast between the two approaches — voluntary in the US, binding in the EU — would, over the following years, become one of the defining features of the global AI regulatory landscape. The voluntary commitments were, in this sense, not just a policy document but a statement of philosophy — a statement that the American approach to AI regulation would be, at least initially, based on cooperation rather than coercion. Whether that philosophy would prove adequate to the challenge was, as of 2026, still an open question.

---

---

---

What would change if more people understood the story behind the White House voluntary commitments? Who benefits from the current state of affairs, and who is left out? The conversation is worth having — with colleagues, with students, with anyone who uses technology without thinking about where it comes from.