One of the most important experiments in AI corporate governance is about to go down: Anthropic, a company ostensibly committed to building safe AI for the long-term benefit of humanity, is on the verge of going public.
Anthropic insists that its long term benefit trust (LTBT), an independent body tasked with holding the company to its mission, will help shield it from profit pressures. The company is also counting on its efforts to “fine-tune” its governance structure over the past five years, including introducing a responsible scaling policy, a frontier safety roadmap and its “transparency hub,” to counter market forces on its mission.
Anthropic’s public benefit corporation (PBC) structure, as it stands now, also makes a strong bet on its seven founders prioritizing safety. So far, that bet has arguably been sustained (although there are some strong counterarguments).
Yet, Anthropic says its post-IPO safeguards can be changed: calling into question the effectiveness of its LTBT and, by extension, the company’s fidelity to its safety-oriented mission.
In particular, the future may impose especially weighty decisions on these seven people. Anthropic plans to give the founders a majority of the voting power among shareholders, which is not unheard of but diverges from how many people think about the governance of publicly held companies. So while billions of people — users and non-users, investors and non-investors — may have a direct or indirect financial and social interest in Anthropic’s corporate decisions, just that handful of founders may be able to dictate the company’s most important decisions. So it’s worth asking: Who protects the mission from the founders if their interpretation of it changes?
Anthropic’s governance experiment
Because of the commitment its founders made to aligning AI development with society’s long-term interests, Anthropic constructed an elaborate private governance architecture for pre- and post-IPO.
Anthropic’s founders left OpenAI out of concern that it was not taking AI’s downsides seriously enough, and, in 2021, formed the company as a Delaware-based PBC. As such, Anthropic’s directors are legally required to balance its stated public benefit with the interests of shareholders as well as the interests of those materially affected by the company’s activities.
When Anthropic goes public, it will remain a PBC and its board will effectively be controlled by the LTBT. Public shareholders will also have limited opportunities to force the company to bend to financial or populist pressure that may conflict with what the founders deem to be Anthropic’s vision. In many ways, the IPO will maintain the company’s system of checks and balances. It will also introduce new rights for founders and employees as well as an additional route by which shareholders may upend the LTBT without the consent of its trustees. But, as Anthropic readily acknowledges, this is all an experiment. These checks may not operate as intended or endure for as long as expected.
The Delaware PBC statute shields the balancing decisions of directors so long as they are informed and not swayed by personal gain, as well as not ones that “no person of ordinary, sound judgment would approve.” In theory, this corporate form allows the board to say, “We are delaying this model even though doing so will cost the company billions.” In practice, the odds of enforcement in the event Anthropic diverges from its mission is unlikely. There’s relatively little case law testing the efficacy of the constraints imposed on PBCs such as Anthropic. Delaware provides no public regulator tasked with determining whether a PBC has faithfully pursued its public benefit, and does not require an ongoing independent audit for maintaining PBC status.
The statute also does not permit many actors to contest the company’s adherence to its mission. Only shareholders holding at least 2% of the company’s shares — or, for a listed company, the lesser of 2% or $2m worth — can bring an action to enforce the balancing duty. The LTBT, which is separate from the board, was established in 2023, in part, to fill that accountability gap.
The trust itself is a Delaware “purpose trust” intended to be governed by five financially disinterested trustees (one of the five seats is currently vacant) — most of whom are experts in complex governance questions, such as financial regulation and national security issues — and it will have significant power post-IPO. It’s meant to stay out of day-to-day commercial decisions and focus on “extreme events,” such as ensuring leadership is “incentivized to carefully evaluate future models for catastrophic risks.” When Anthropic’s Series C closed in 2023, it amended its charter to add a single share of Class T stock, held by the trust. After the IPO, that stock can be transferred only with the same approvals that can amend the trust’s terms. That single share carries the right to elect and remove a majority of the company’s board, subject to special procedures for one seat on which the lab’s employees may — in very specific circumstances — have limited input.
Anthropic’s publicly filed prospectus — a version of the disclosure document a company files with the SEC before going public — creates a new actor in the company’s complex corporate governance scheme. According to the filing, Anthropic’s seven co-founders will form a “Founder LLC” when it goes public, contending that the founders are “distinctly equipped to be stewards of our mission.” They will decide among themselves, generally by majority vote, how to cast a single Class F share. That lone share commands 50.1% of the vote on key shareholder matters. Class A stock, sold to the public, carries one vote per share. This suggests that the seven founders may have expansive power over key decisions.
The founding seven also wield substantial influence over the post-IPO board. The board will have seven seats. The LTBT would elect four. Class F (again, the share held by the Founder LLC) and Class A would elect the remaining three together, with Class F possessing the controlling share of the vote. In the traditional telling, dispersed shareholders channel market discipline by voting out directors who stray. That common story will not apply here.
Employees also get a very limited check. Non-founder employees will receive Class E shares that lack economic rights but come with some opportunities to inform governance changes. Their most notable role concerns one of the four trust-selected seats — the “Specified Class T Director.” The trust cannot nominate, redesignate, or remove that director without cause unless the board’s nominating committee approves. If the committee declines or fails to approve within 15 days, it goes to a Class E vote. ( Employees are set receive a single Class E share as standard, but get additional shares for each standard Class A share they pledge to charity.)
But there’s a bit of a quirk: those shares will not be immediately available upon IPO. Anthropic has until March 31, 2027 to set the terms for issuing them. Until then, the committee’s decision is final. The trust can remove that director for cause without either approval. The committee consists of one trust-selected director, Netflix co-founder Reed Hastings, and two directors elected by the founder-controlled vote: co-founder Daniela Amodei and VC Yasmin Razavi. Approval takes two of three votes.
The Founder LLC also has rules for its own membership. Founders can exit the LLC if they resign, die, sell too much stock, or are expelled for “cause.” Once the group dwindles to two founders or successors for more than 90 consecutive days, the Class F share’s outsized vote will end following a 450-day transition period. The founders could also end that arrangement by voluntarily surrendering the share or dissolving the LLC, in which case the change would take effect almost immediately. Any of those scenarios —-a sunset, in the parlance of the prospectus —would result in the board operating under a new regime. The LTBT would still elect a majority, but Class A would have the ability to vote on the remaining members.
The filing also resolves several details that were previously unclear. The Founder LLC agreement addresses who can be expelled for “cause” and which successors count toward the three-member minimum. A sunset also cancels employees’ Class E shares without compensation and moves directors to staggered three-year terms. Most importantly, the filing details when and how the powers afforded to the LTBT and the composition of the board can be altered.
Who can rewrite the rules?
The division of power between the LTBT-selected board members and the remaining three raises a possibility worthy of external scrutiny: What happens when the founders lose a 4–3 vote and decide the governance structure itself is the problem? Could the Founder LLC use its 50.1% to eliminate Class T, shrink its director-election rights, change the board’s size, or rewrite the Trust’s governing documents?
The 2023 announcement offers a partial answer. Anthropic built in provisions that “allow changes to the trust and its powers without the consent of the trustees if sufficiently large supermajorities of the stockholders agree.” A post by the lawyers who helped design the trust — Yale’s John Morley and Wilson Sonsini’s David Berger and Amy Simmerman — helps fill in some blanks. They describe three amendment paths: the trustees acting together with the shareholders; the trustees acting together with the directors, an option meant to last only until the trust could elect a board majority; and a shareholder supermajority acting without the trustees. The new filings keep only the last two, each in revised form.
The first surviving path is the trustee-and-director route, which now remains available at any time. Such changes require approval from 80% of the LTBT trustees and 80% of Anthropic’s authorized board seats, counting vacant seats, each rounded to the nearest whole number. By way of example, with seven authorized board seats, six directors must approve. With four LTBT trustees in place, three must approve; with five, four.
The second is the shareholder route, which does not require the trustees’ consent. The designers frame that failsafe as protection for shareholders against the trustees. Until the earlier of the founders’ sunset or the IPO’s third anniversary, the route requires three separate approvals: the Founder LLC, a majority of the votes cast by employees’ Class E shares, and a majority of the votes cast by Class A shares. From the third anniversary until the sunset, the employee threshold rises to 75% of votes cast. After the sunset, the route requires 85% of all outstanding Class A shares. In the IPO’s first three years, then, the founders cannot use the failsafe alone, but they need only simple majorities of the public and employee votes cast, and the trustees cannot block it.
When the trustees and the board agree, the trust agreement may be altered without any shareholder vote. The upshot is that trustee selection is a big deal and that process is left largely to the discretion of existing trustees, subject to one check. They must give the board’s nominating committee 30 days’ notice.
Under the current charter, the Class T stock’s “protective provisions” already give the trust a veto over charter changes that adversely affect its rights. After the IPO, that veto can be bypassed: Any change that adversely affects Class T’s rights, as well as any change to the company’s PBC status and the board’s statutory powers, will require approval by six of seven board members plus the trust, or the applicable shareholder approval threshold.
Delaware law supplies another backstop against subverting the LTBT. Under Section 242(b)(2) of the Delaware General Corporation Law, holders of a class of stock vote as a separate class on any charter amendment that would “alter or change the powers, preferences, or special rights of the shares of such class so as to affect them adversely.” In plain English: you can’t rewrite the rules of the game to hurt one team without that team getting a vote on the new rules. However many votes the other shareholders muster, the class whose rights are on the chopping block gets to weigh in separately. Unlike some other class votes, this one cannot be waived in the charter. Standing alone, that default would give the trust a veto over any amendment stripping Class T of its director-election rights.
The filings now explain how the failsafe interacts with that statutory veto. Under the stockholders’ agreement, once shareholders clear the voting threshold, they can instruct the trust to vote its Class T share as they see fit — including in favor of amendments that strip Class T of its rights. Should the trust refuse, Anthropic’s corporate secretary may cast the vote on its behalf. Put differently, the trust keeps its veto in name, but the shareholders who clear the threshold decide how it is used.
Those shareholders also have a more direct option. The charter allows them to redeem the trust’s share — that is, to have Anthropic cancel it without compensation. Because a redemption carries out the charter’s existing terms rather than amending them, Section 242(b)(2) never comes into play, and neither the board nor the trustees must sign off. Once the share is gone, the trust-selected directors’ terms generally end and the trust itself terminates.
The filings also address three other potential ways around that statutory vote.
First, board expansion. The charter specifies that the trust elects a majority of however many seats exist, and only approval by 80% of the board can change the board’s size. Adding seats would nevertheless ensure the trust selects a majority.
Second, a merger. A Delaware court, in Warner Communications v. Chris-Craft Industries, held that a class’s right to vote on adverse charter amendments does not automatically extend to mergers. A merger that cashes out the Class T share, or converts it into stock with weaker rights, could therefore serve as an end run. Anthropic could foreclose that avenue through drafting: in Elliott Associates v. Avatex, the Delaware Supreme Court held that class protections covering changes made “whether by merger, consolidation or otherwise” do secure a class vote on a merger. Anthropic’s post-IPO charter includes that language. Consequently, a merger that strips Class T’s rights requires the same approvals as any other effort to do so, and no longer offers a separate end run.
Third, the trust agreement itself. Through the failsafe, shareholders can amend the trust agreement directly, such as the provisions for trustee selection. That would leave Class T’s rights untouched, and so trigger no class vote, while changing who wields them. Unlike a charter amendment or a merger, this route does not require the board’s approval at all — and, unlike the first two end runs, nothing in the filings forecloses it.
Though the prospectus has received considerable attention, these details have so far not been headline news. Yet the strength of Anthropic’s most important safety-governance institution may ultimately depend on these governance choices.
The real test of this governance experiment will come later and perhaps without warning: instances in which that elaborate governance design will be stress tested, such as when quarterly returns disappoint. If the safeguards outlined here operate as intended, Anthropic will have handed other companies, and the lawmakers still searching for a way to govern them, a working template for keeping a public mission alive inside a public company. If they don’t, the key will be identifying the source of the failure as soon as possible and devising the necessary corrections.
