Recently, the White House and some of the leading artificial intelligence (AI) companies released a “Joint Commitment on Frontier Responsibilities.” This voluntary document charts a path toward a self-regulation standard for governing frontier AI systems. The document is an innovative approach that provides both the government and industry with the flexibility to adapt to the rapidly evolving capabilities of AI systems and helps prevent some of the most common pitfalls of government intervention. This commitment should provide a valuable building block for the regulation of the AI industry, but some questions are left unanswered.
While the responsibilities and commitments of the industry are explicit and clear, the document does not address key questions regarding the role of government and leaves the industry with a lack of clarity on when, how, and under what justification the government can intervene in the market and what tools it is allowed to use.
The four commitments made by the leading frontier AI labs try to tackle the biggest risks associated with AI development. First, they must implement “robust internal controls” that help them monitor the capabilities and behavior of their models. Second, they must “empower” an internal team to ensure the implementation of these controls goes smoothly. Third, they must partner with an external auditor to evaluate the adoption of these two commitments. Finally, the companies must designate an independent committee of the board of directors to oversee the internal control team and evaluate the reports created by these independent evaluators.
Using a publicly announced voluntary commitment is an innovative solution that might give the industry and government enough flexibility to adapt to the technology’s changing needs while still giving the government enough teeth to punish bad actors, thereby quelling some of the concerns fueling calls for further AI regulation. As some have pointed out, making these commitments public could enable federal agencies, such as the Federal Trade Commission (FTC), to step in and punish the labs if they fail to live up to their promises. While this is, in theory, a policy of self-regulation, the government has carved out a method to secure enforcement, ensuring that the parties live up to their bargain.
The commitment’s approach of requiring labs to incorporate internal teams and measures to mitigate risk and having these protocols audited by independent verifiers is a similar approach to some of the proposals currently floated in Congress. In essence, this approach allows for proactive monitoring of safety risks and has the advantage of potentially discovering process flaws before any harm materializes. Usually, ex ante regulation runs the risk of being overly prescriptive and harming innovation, but the flexible approach where labs are still free to choose their preferred auditor, without the need for it to be licensed by the government—therefore avoiding the common pitfalls of government-run licensing—curtails some of the risks of ex ante regulation.
Essentially, labs must demonstrate that their risk mitigating processes are sound. That claim is then scrutinized by a specialized, knowledgeable, and highly technical entity: the auditor. This gives the public assurance that the labs are adhering to the highest standards of care in developing frontier AI. And, in case something does indeed go wrong, the public and the courts will be able to discern if damages are a consequence of reckless or negligent action versus a true accident or a blind spot in the pre-existing best practices.
Where the Commitment Falls Short
While this commitment is clear in its mandate to industry and charts a clear picture of their responsibilities going forward, it still leaves a significant number of questions left unanswered. Most salient, and a reasonable tradeoff of its voluntary, opt-in nature, is how sustainable these rules are for governing AI long-term. The major players of today that were able to sit at the table and sign the agreement might not be the major players of tomorrow. If a new frontier lab emerges tomorrow and decides, because it has not signed onto this agreement, not to follow this new standard, what would be the consequence? One could argue that the agreement could become an industry standard or best practice and become, in the eyes of the court, de facto law as a legal custom. But this process would need to play out in court, which could create a lag in adoption.
Additionally, relying on this approach means that smaller players will not have a say in shaping the rules that will eventually govern them, as they lack a seat at the table. While the agreement creates significant legal certainty by charting a path for the major players on what they should do, it fails to extend that certainty to smaller or future actors in the market.
Additionally, the document fails to set out a set of duties and responsibilities or delineate the government’s powers under this scheme. While we know the duties industry must fulfill, we do not know what sanctions the government could impose on companies that fail to do so. What agency or document would determine if the internal controls are “robust”? Is receiving a positive audit an indication that the government will not pursue legal action against a lab? Will sanctions be merely monetary, or will they face de facto market bans through tools such as export controls (as with Anthropic and its Mythos/Fable models)?
If the government’s main enforcement tool is through lawsuits filed by the FTC, will it wait for a mandate or a settlement before taking further action, or could it unilaterally impose sanctions while it waits for a resolution? There are no clear boundaries of what the government can and cannot do as the enforcer of this agreement. This gives the executive wide latitude to impose sanctions on AI labs, with little need for justification. As the Anthropic-Pentagon saga demonstrated, this is ripe for abuse and contravenes the rule of law.
Conclusion
The Joint Commitment on Frontier AI is a sophisticated, robust self-governance mechanism that can capture most of the benefits of a light-touch, flexible approach while still responding to concerns about potential negligence or recklessness in frontier AI development. For now, it promises to provide a good roadmap to a common safety standard to mitigate risks in AI, but the reality of its strength or enforceability remains to be seen in any challenge.
The framework is in many ways a positive step; however, other questions in frontier AI governance remain unanswered, especially regarding the constraints and limits of government action.
