AI companies have developed bespoke legal structures that purport to hardcode public interest into their DNA: non-profit foundations, capped-profit joint ventures, public benefit corporations and long-term benefit trusts. On their own, these private ordering mechanisms do not function as a meaningful layer of AI governance. Drawing on the corporate structures of three frontier AI labs (Google DeepMind, OpenAI, Anthropic), this project maps how each company has designed its legal architecture to reconcile its stated mission to “benefit humanity” with financial imperatives. It then tests whether structure shapes behaviour across three case studies: positions on AI regulation (EU AI Act, California AI safety regulation), voluntary scaling policies, and military contracting.
The findings suggest that corporate structure creates friction but does not determine governance outcomes. In fact, companies tended to modify their structures to be more responsive to investors as AI transitioned from proof of concept to lucrative product. Where structure constrained action, it was dismantled. Crucially, the ability to exit a constraining corporate form is itself a governance failure. OpenAI’s conversion from non-profit to public benefit corporation illustrates this point: the original structure was effective enough to attract legal scrutiny, but ultimately insufficient to maintain mission above profits.
The paper identifies leverage points available to attorneys general, shareholders, and civil society under existing corporate law, while documenting their limits. The broader implication for multi-level governance is that private, reversible choices of corporate vehicles alone are too brittle to ensure accountability for a technology with consequences of this magnitude. As frontier AI companies move toward public listings, the stakes of those structural choices intensify. An IPO does not merely dilute mission-driven ownership; it imports shareholder primacy and market logics into architectures originally designed to resist them.