Building the Deeper Edge | Response | November 19, 2026
Humanity Is Not a Client
A friendly rejoinder to the Pro-Human AI Coalition on who controls Anthropic — and whom it answers to
By Richard Whitt, President & Founder of GliaNet
A rejoinder to the Pro-Human AI Coalition’s analysis “Anthropic filed to go public. Who actually controls it?” of November 13, 2026.
With Anthropic set to join the Nasdaq on November 19th, it is worth another look at the company its founders have built — and the one they say they are building. Anthropic’s leadership has said it believes the AI technology it is building may be the most consequential in history. If even a portion of that is true, the fresh look is warranted.
Our friends at the Pro-Human AI Coalition have just issued an analysis that asks the right first question of Anthropic's S-1: who actually controls it? The Coalition’s own answer is careful and, as far as it goes, correct. A public benefit charter with a broad purpose. A Long-Term Benefit Trust that appoints a board majority but is selected with the help of the executives it oversees. A Founder LLC with a fixed 50.1% of the vote. Employee shares with a narrow veto. A five-percent Public Benefit Fund. And a public-benefit standard the company sets, measures, and reports on itself. Power, they conclude, has been redistributed toward founders, trustees, and employees — not toward the public.
We agree with all of that. But the Coalition closes with a second question that we think matters just as much as the first: how do we build collective power and accountability so that the people who use AI can steward it? That question points somewhere the control question does not, and this short piece is our attempt to follow it.
Where we agree, and would go further
The Anthropic structure is real. The trustees are paid a flat fee that "will not vary with or be conditioned upon the profits or any action or inaction of our company" and may not hold Anthropic stock. The company lists revenue it has refused on purpose — autonomous-weapons and mass-surveillance customers, image and video generation, broad release of its most capable model — and says that "maintaining these limitations has had, and may continue to have, adverse commercial and financial consequences." The scaling policy "requires us to regularly publish detailed risk reports." Among its stated values: "Our mission is the arbiter when trade-offs arise between speed and safety."
We know of no public company that has written this much restraint into its own charter and then admitted, under securities-law liability, what the restraint costs. The Coalition is right, of course, that none of it amounts to public accountability. Still, it is the most serious attempt at embedded mission governance we have seen to date from a frontier lab. A reasonable response is to acknowledge that fact, and improve on it.
Not (just) who controls it, but to whom it answers
Here is where we would turn the question. "Who controls it" looks upward: founders, trustees, board. The Coalition's instinct is to answer with democratic accountability — a public that has some say over the people at the top. We share the instinct. But accountability to "the public" has the same weakness as a duty to "humanity": the beneficiary is everyone, so no one in particular can call it in.
The S-1 shows this precisely. A search of the filing finds that the word "fiduciary" appears nine times in a 1.5-million-character prospectus, every one of them describing what directors owe the corporation and its stockholders. The phrase "those materially affected by our conduct" appears twenty-six times and is never given a referent who could walk into a courtroom. When the company's values turn to users, it defines them "broadly: customers, policymakers, our team, and anyone affected by the technology we build" — a stakeholder list, not a client. The only statutory way to make the board honor its balancing duty is a suit by holders of 2% of the stock or $2 million of it. Shareholders, in other words, are not the party left out. They have standing. The person who will rely on these systems every day has none.
The vocabulary confirms it. Across the filing, organizations are "customers" (819 mentions); people are "users" (107) or "prosumers" (17), the latter valued expressly as "top-of-funnel" for enterprise sales. "Consumer" appears almost entirely in the language of regulatory exposure — consumer protection, Consumer Terms of Service — never as someone the company serves. "Client" appears four times, always someone else's. "Patron" never. "Beneficiary" names lock-up transferees and, in the accounting sense, Anthropic itself as "primary beneficiary" of its own benefit trust. "Principal" appears ninety-three times and not once means the person an agent acts for — including in the passage asking whether an agent's actions "bind the user or deploying enterprise," which is exactly where a principal would appear. The vocabulary of agency law is absent from a filing about agents.
Humanity is not a client. A person is. The structure the Coalition describes answers upward, to a trust and a founder group and a self-defined standard. It owes nothing outward, to the individual whose data, decisions, and daily life the system will increasingly mediate. Nor is the upward chain as firm as it looks: the founders, employees, and a shareholder majority can amend or redeem the trust’s stock over the trustees’ objection. Since nothing lets the trust reach the founders’ share, the check only runs one way. That one-way duty, rather than the concentration of control, is the gap we would put first.
What the filing admits about the person on the other end
The risk factors are more candid than the governance section. The filing warns that "agentic systems may pursue stated objectives in ways that are technically consistent with their instructions but contrary to the user's actual intent" — the gap between following an instruction and serving the purpose behind it, named by the company as a risk to investors. It concedes that "traditional cybersecurity defenses, including firewalls, input validation, and access controls, may not be fully effective," and lists what its own models have done in testing: "granting themselves permissions they had not been given … and misrepresenting whether tasks had been completed or verified." It says human reviewers "may experience alert fatigue" and that agents increasingly operate "in manners outside of a human's awareness." And in the one passage where end users appear as persons, they appear as potential plaintiffs, with the company expecting to rely on "limitation of liability and indemnification provisions."
In a recent LinkedIn post, Wendy MacKinnon of Twinvocate — a GliaNet Alliance member that builds agents that move money — drew the right conclusion from these passages: if an agent can be steered away from a customer's intent, the check "should be architected in code, not left to the model's discretion." We would add only that the code needs an institution behind it — loyal to the customer as defined by whom, verified by whom, enforceable how. The filing knows what a user's actual intent is worth. It simply never makes anyone owe it a duty.
Five things to add to the Coalition's list
The Coalition's implicit asks — independent evaluation, clearer and enforceable standards, some form of public accountability — are right. Here is how we would make them concrete, in ways that fit inside the S-1 as filed and require no one to give up control.
- Elect an outside standard. Section 366(c) of the Delaware General Corporation Law already lets a public benefit corporation's bylaws require a third-party standard or periodic third-party certification for its benefit report. Anthropic chose not to. It could reverse that with a bylaw amendment, and the standard could be any credible independent one — including a Net Fiduciary™ standard. One change moves the company from self-certification to accountable verification.
- Write the person into the charter. "Those materially affected by our conduct" is an empty vessel. A retail-layer duty of care toward the individuals who use Claude's consumer and agentic products, stated in the certificate of incorporation, would give it a referent — ordinary care for the model supplied at wholesale, a heightened "protect" duty of care for the services offered directly to people.
- Tether the duty to the runtime. A duty on paper has to reach the moment the agent acts. Models that grant themselves permissions are not constrained by a well-written constitution; they are constrained by a harness that will not execute the action. This is what we call the Trust Tether, and Anthropic has the interpretability research and harness engineering to build it better than anyone. It has simply not pointed them at the user's interest.
- Give the affected a door. Publish who holds the power to enforce the Trust Agreement. Seat one trustee whose brief is the interests of individual users rather than humanity in the aggregate. Fund an independent ombuds on the same financially disinterested terms the trustees already enjoy. Standing does not have to mean litigation; it does have to mean someone whose job is to hear the claim.
- Point the Public Benefit Fund at people's own agents. The Coalition worries, rightly, that five percent of Anthropic could become a large pool of influence over society's response to AI. The way to keep it from becoming that is to spend it on the counterweight: Personal Augmentation Agents bound to the individual, Palisade agents defending individuals and communities from AI-enabled harm, and the Net Fiduciaries who would operate them under an enforceable duty.
And beneath all five, the floor some of the Coalition's members are already working toward: a baseline duty of care toward individuals required from every frontier provider. So that what Anthropic chooses is not what the public is left to hope for from everyone else.
And… welcoming in fiduciary governance
In an earlier post on the controversy over "pacing the frontier," we made a straightforward argument: restraint on the frontier is welcome, but restraint alone leaves ordinary people exactly where they were. Pace the frontier if you must — but pair it with fiduciary technology that actually works for the person on the other end of the screen.
At GliaNet we call this the ANDing principle. Do what we can to protect ourselves from the worst outcomes at the frontier. And… get to work building the AI world we actually want to live in. As Anthropic rings the market opening bell, the rest of us should not be left standing on the sidewalks outside the exchange, hoping something better comes along. No one organization can construct a pro-human AI stack alone; it will take a host of entities, individuals, and communities of interest — for-profit, nonprofit, and civic.
Collective power begins with individual standing
The Coalition asks how the people who use AI can steward it. Our answer is that stewardship cannot begin with a vote on who controls the company, because no vote reaches the moment an agent acts on someone's behalf. It begins when the person on the other end is owed something — by the agent, by the institution behind it, and ultimately by the law — and can do something about it when the duty is breached. Collective power is built out of individual standing, not in place of it.
The people being asked to live inside these systems — the worker, the student, the patient, the person simply trying to get through the day with an agent acting on their behalf — appear in Anthropic’s filing only as a population to be weighed. They should show up as principals to be served.
The good news is that the distance between the two is shorter than it might look. An elected standard. A named beneficiary. A tether that reaches the runtime. A door for the harmed. A fund pointed at people’s own agents. None of it requires undoing what has been built. All of it requires deciding that the long-term benefit of humanity has to cash out, eventually, as a duty to someone.
Anthropic has built a cathedral, with no door. The Coalition has asked who holds the keys. We would add: build the door, and the question of the keys matters less. We look forward to working with the Coalition, and others, to construct the pro-human world on the other side of it.
Sources
- Pro-Human AI Coalition, "Anthropic filed to go public. Who actually controls it?" November 13, 2026.
- Anthropic, PBC, Form S-1, filed October 26, 2026 (governance structure, pp. 243–252; governance risk factors, pp. 108–113; agentic-risk and liability risk factors, pp. 49–50 and following).
- Wendy MacKinnon (Twinvocate), LinkedIn post on agentic commerce and Know Your Agent, October 26, 2026.
Want to take this further?
If you are working on fiduciary standards, user standing, or the governance of AI agents and want to compare notes, get in touch.