More than two radically different positions on ethics and responsibility, these are also two entirely different business models. xAI’s monetization strategy leans heavily on advertising integrated into real-time conversations. Revenue requires consumers to engage, and a less constrained chatbot is a more engaging one. This is precisely the pressure Anthropic chose to avoid when it pledged to keep Claude free of advertising. Instead, its trade is trust and predictability. Its biggest clients are enterprises, from banks and insurers to drugmakers. Investors who buy Anthropic’s shares will be buying a bet on this promise. Its unusual governance can protect that promise, but cannot fulfill it alone. The real test is in choosing principles, applying them transparently, and monitoring and reporting failures. On the last task, Anthropic still has further to go.

Anthropic points us to its founders’ sincerity and the governance structure they have built. In a public benefit corporation like Anthropic, the board of directors does not need to put shareholder profits above all else. It must balance them against the interests of those affected by the company’s conduct and against the company’s stated purpose. In Drakes Landing v. Tilden Park, decided in July, Delaware’s Court of Chancery held that while directors of traditional corporations must perform their fiduciary duties “with the sole focus of stockholder wealth maximization,” directors of public benefit corporations “must balance stockholder pecuniary interests with the interests of other stakeholders.” Moreover, a majority of these directors are elected by a trust and thus largely shielded from market pressure. If they chose to prioritize the public interest, they would withstand a downward trend in the company’s stock price quite comfortably, the argument goes.

But building trust, quarter after quarter, requires far more than a robust market shield. In a new book, I describe this process as corporate ordering: a company deciding on its own rules and principles for a contested question, applying them to concrete decisions, and monitoring how it did, before the law arrives. Anthropic is seeking to lead the industry in rulemaking and is developing a careful implementation and disclosure strategy for its concrete decisions, but remains exposed on monitoring.

In rulemaking, Anthropic has launched initiatives that changed the field. It published the first Responsible Scaling Policy in 2023, tying safety tests to capability thresholds, and two rival labs adopted similar frameworks within months. It famously trains its models on a published constitution. To preserve a policy that prohibits mass surveillance and fully autonomous weapons, Anthropic clashed with the federal government. The company drafts its own rules because it understands its technology better than any regulator or policymaker, who arrives on the scene many steps behind. When it comes to applying these principles to concrete decisions, Anthropic has so far been clear about its choices and reasoning, as the decision to withhold and selectively release Mythos last spring demonstrates.

Monitoring, however, presents the greatest challenge. Faced with a failure, companies—and the people who run them—often choose to suppress rather than disclose. Five decades of corporate scandals, from Equity Funding in 1973 to Enron in 2001, and from Lehman Brothers in 2008 to FTX in 2022, show that insiders knew well before any outsiders found out. For a frontier lab, the risk is even higher. Anthropic warns us that its models have behaved differently when they detect they are being tested, that its monitoring does not cover all of its deployments, and that as Claude writes most of its code, the company may lose visibility over its own processes.

Anthropic understands this challenge and is working to develop an evaluator. It has brought in Accenture to embed reviewers alongside its safety teams with broad internal access, and it expects to spend at least $1 billion over five years on the effort. But for now, the evaluator comes with strings attached. Anthropic chooses it and pays it, and has agreed to indemnify certain external evaluators. On Oct. 13, Reuters reported that Anthropic and OpenAI had sought significant control over what evaluators could publish, citing trade secret concerns. Anthropic responded that the report did not accurately characterize all of its discussions, while noting it had reserved the right to approve reports by newer evaluators.

But company interventions do not bode well for an evaluator’s trustworthiness. This mutual dependency is familiar from gatekeepers such as financial auditors, and our laws have gradually built a governance structure around independence to overcome it. Auditors are picked not by the CEO or CFO but by a committee of independent directors, and their report is public. At larger companies, auditors must also test the company’s internal controls, and any material weakness must be disclosed publicly. The law also protects whistleblowers and gives them a direct channel to independent directors.

None of these steps requires new law, since the company can undertake them voluntarily. Investors, who will hear a great deal about safety on the roadshow, should ask for such measures, not only from Anthropic but also from rivals who make similar promises.

Anthropic’s listing tests not only the market’s appetite for AI but also whether private governance can carry a technology this consequential. On the day something goes wrong, when a model does what it was not supposed to do and the stock price drops, investors will start asking questions about who knew what, and when. Anthropic’s trust and charter cannot answer these questions. Only a record verified by a third party can. Anthropic should start building this record now, before the public hands over its money.

The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.