The Reliance Gap in AI Corporate Governance

Leading AI developers have built their safety commitments out of corporate law, such as in the form of public benefit corporations, controlling foundations, special share classes, trusts and published safety frameworks. As it turns out, each of these commitments can be lawfully loosened by those it was meant to constrain, to which few outsiders have the means to object. Yet regulators rely on these commitments when deciding whether to intervene. This post names the resulting shortfall the “reliance gap”, explains why corporate law addresses it inadequately, and proposes remedies such as consent rights and a one-way amendment ratchet which can be applied to any company whose public commitments the public relies upon.

Promises Made of Corporate Law

In October 2025, California’s Attorney General agreed not to object to OpenAI’s restructuring. Under the memorandum of understanding, the non-profit Foundation controls the new public benefit corporation insofar as it holds Class N stock, and “The decision whether to relinquish the Class N Common Stock, if ever, is in the control of the NFP Board” (para 2). A Safety and Security Committee may require mitigations “up to and including halting the release of models or AI systems” (para 11). Likewise, Anthropic’s Long-Term Benefit Trust gives trustees a special share class entitled to elect a growing part of the board.

These are guardrail devices built from material designed for another job. Organisational law lets owners commit to one another, by partitioning assets and giving governance a veto over their diversion (Hansmann and Kraakman; Ricks), and it does not bind owners to outsiders. Indeed, Delaware gives a benefit corporation’s beneficiaries no standing to sue (§ 367), and its stockholders are, as Reiser puts it, “uniquely hamstrung as enforcers”.

Lawful Ways Out

Firstly, a company can leave the public benefit form with very little friction. Since 2020, Delaware has required only a simple majority, unless the charter demands more, to convert out of public benefit status (Skadden). To date, xAI ended its Nevada benefit status in 2024, a change reported a year later in 2025.

Secondly, the guardian appointed can be amended easily. OpenAI’s Foundation may relinquish its shares by its own decision, and the Attorney General is owed 21 days’ notice, not a veto, before control changes (para 19). On the other hand, Anthropic’s trust can be ended by a stockholder supermajority, a “kill switch” that Fried and Reiter commend as a check on overbearing guardians.

Lastly, the trigger can be rewritten and the goal post shifted internally. California’s frontier-AI statute requires developers to “comply with” their published safety framework, but lets them modify it and publish the change “and a justification for that modification within 30 days” (SB 53, § 22757.12(a)–(b)). Not surprisingly, OpenAI’s 2025 framework says it “may adjust” its requirements if a rival releases a high-risk system (TechCrunch), while Anthropic’s 2026 revision dropped certain language implying a pause (GovAI). As the law stands, none of these changes is unlawful.

The Reliance Gap

A reversible safeguard is not necessarily worse than having no safeguards at all, but it is misleading when the public is led to believe that it is fixed. California’s non-objection was formed in reliance upon OpenAI’s representations about its structure. Indeed, institutions often interpret the existence of a compliance structure as evidence of compliance (Edelman). The reliance gap is the difference between the purported safeguards committed to by AI companies when regulators rely on it for decision making and the actual safeguard in place once regulatory attention has moved on.

While Fried and Reiter worry that guardians may be too strong, the reliance gap poses a different risk. That is, who actually holds the switch? Where only the boards and investors have de facto control, regulators that relied on the seemingly reasonably safe structure bears an asymmetrical risk. The structures might add friction, and the memorandum technically only ends by mutual consent (para 27). Nevertheless, in practice, when the Attorney General subpoenaed OpenAI in September 2026 over cybersecurity incidents, the announcement did not invoke the memorandum at all.

Closing the Gap

How then should the gap be closed? It is submitted that, firstly, where a regulator’s forbearance rests on a commitment, notice should become consent and changes of form, control, mission or safety authority should need an independent body’s approval. Secondly, published safety commitments should move through a one-way ratchet such that strengthening changes should take effect on publication while weakening changes would need independent approval. A UK community interest company already cannot simply leave its form without regulatory decision (2004 Act, s 55). Indeed, such powers must respect constitutional and, in Europe, free-movement limits (Commission v United Kingdom) and should be confined to the largest firms. Currently, AI safety promises written in corporate law are written in pencil, easily erased but hard for the public to detect.

This post draws on the author’s working paper, “Written in Pencil” (to be posted on SSRN).

Wen Yi Leong is a former Judicial Clerk to the Chief Justice of Malaysia.