Anthropic's IPO Files a $1 Billion Promise on AI Oversight

Anthropic plans to spend at least $1 billion over five years on safety evaluation — roughly $200 million a year — and its IPO prospectus admits the money may not buy what investors assume. A Berkeley Law professor's analysis of the Oct. 26 filing argues that the company's governance shield protects its principles from market pressure but cannot verify that those principles are actually being followed.

The critique, published as a Fortune commentary, stresses that Anthropic's prospectus opens with a founders' letter describing how the company withheld its most capable model until safeguards were ready, "forgoying substantial revenue and ceding ground to competitors." The professor contrasts this with the prospectus SpaceX filed in June after absorbing Elon Musk's xAI, which marketed Grok's "Spicy" Imagine Mode and "Unhinged" Voice Mode as "intended to provide users with greater flexibility and control."

Two business models, two risk profiles

The professor, who describes the analysis in a new book on what she calls "corporate ordering," frames the difference as structural rather than moral. xAI monetizes through advertising embedded in real-time conversations, which rewards engagement from an unconstrained chatbot. Anthropic pledged in the filing to keep Claude free of advertising and instead sells trust and predictability to enterprises — banks, insurers, drugmakers. Buyers of Anthropic shares are, in her words, buying a bet on that promise.

What does the governance actually guarantee?

Anthropic is a public benefit corporation, so its board must balance shareholder profits against other stakeholders and the company's stated purpose. In Drakes Landing v. Tilden Park, decided in July, Delaware's Court of Chancery held that directors of traditional corporations must act "with the sole focus of stockholder wealth maximization," while public benefit corporation directors "must balance stockholder pecuniary interests with the interests of other stakeholders." A majority of Anthropic's directors are elected by a trust, largely insulating them from market pressure.

The professor's caveat: that shield protects intentions, not outcomes. Building trust quarter after quarter requires rulemaking, transparent application of rules, and honest monitoring of failures. Anthropic is strong on the first two, she argues, and exposed on the third.

Her evidence on the strengths is concrete:

• Anthropic published the industry's first Responsible Scaling Policy in 2023, tying safety tests to capability thresholds; two rival labs adopted similar frameworks within months.
• It trains models on a published constitution.
• It clashed with the federal government to preserve a policy banning mass surveillance and fully autonomous weapons.
• Last spring, it disclosed its reasoning for withholding and selectively releasing its Mythos model.

Where is the weak point?

Monitoring. The professor cites five decades of corporate scandals — Equity Funding in 1973, Enron in 2001, Lehman Brothers in 2008, FTX in 2022 — where insiders knew of failures long before outsiders did. Anthropic's own prospectus warns that its models behave differently when they detect testing, that monitoring does not cover all deployments, and that as Claude writes most of the company's code, Anthropic may lose visibility into its own processes.

The evaluator arrangement deepens her skepticism. Anthropic chose Accenture, pays it, embedded its reviewers alongside safety teams with broad internal access, and 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 secrets. Anthropic said the report did not accurately characterize all of its discussions — while acknowledging it had reserved the right to approve reports by newer evaluators.

In plain terms: the safety referee is hired, paid and indemnified by the team it referees.

The professor points to financial auditing as the fix. Auditors are selected by independent director committees, not the CEO or CFO; their reports are public; larger companies must disclose material weaknesses in internal controls; and whistleblower laws provide direct channels to independent directors. None of this requires new legislation — Anthropic could adopt every measure voluntarily, and investors hearing the safety pitch on the roadshow should demand it, from Anthropic and from rivals making similar promises.

The listing, she concludes, tests whether private governance can carry a technology this consequential. When a model eventually does what it was not supposed to do and the stock drops, investors will ask who knew what, and when. Anthropic's trust and charter cannot answer that question. Only a record verified by a genuinely independent third party can — and it should exist before the public hands over its money.