The Power to Decide What Good Means

A founder wants the company’s purpose to survive its next investor. A newsroom wants to investigate the people who advertise in it. An AI lab wants the freedom to delay a release that everyone else wants shipped.

These are good reasons to protect an institution from outside pressure. Some worthwhile decisions become almost impossible if the people making them can be removed the moment the decision gets expensive.

But there is something else inside that protection.

A mission cannot appoint a director, approve a release, or decide which compromise has gone too far. People do those things. They interpret the purpose, decide what it requires, and exercise the authority built around it.

When we protect a mission, we also protect someone’s interpretation of it.

That is an unavoidable feature of governance. The question is whether we design for it honestly.

The same protection works when you are wrong

Consider a founder who refuses to abandon a research programme after two disappointing years.

The investors want it closed. The founder believes the work needs another three years. Special voting rights allow the founder to hold the course.

If the programme succeeds, the structure will look far-sighted. If it consumes another three years and produces nothing, the same structure may look like the reason nobody could stop it.

At the moment of disagreement, both futures remain possible. Governance has to work before anyone knows which story they are in.

Google made this bargain explicit when it prepared to go public in 2004. Its founders argued that a dual-class share structure would preserve independence and the ability to pursue a long-term approach. They also explained that the arrangement would leave outside investors with less influence over strategic decisions. Investors were being asked to back a particular team’s judgment, alongside the business. (Google’s 2004 founders’ letter)

That is a defensible bargain. Building something difficult can require years of conviction that looks unreasonable from outside.

But the voting rights cannot distinguish persistence from attachment to a failing idea. They protect the decision-maker in either case.

The quality of the judgment has to come from somewhere else.

A purpose still leaves choices

Moving authority from a founder to a trust changes who makes the judgment. It does not remove the need for judgment.

Patagonia is a useful example outside technology. Its Purpose Trust owns 2% of the company and all its voting stock, while the Holdfast Collective owns the remaining 98% and all the nonvoting stock. The trust’s role includes protecting the company’s values through powers over decisions such as board membership and changes to its charter. (Patagonia’s ownership explanation)

That gives an environmental purpose institutional weight. It also leaves people with difficult choices.

Suppose a business could grow substantially by making a product less environmentally demanding than the alternative, while still increasing its own total footprint. Should it expand? How much should it reinvest in the business rather than make available for environmental work?

Those are illustrative questions, not allegations about Patagonia. They show how much decision-making remains after a purpose has been protected.

Two people can be equally committed to an environmental mission and disagree about the right answer. They might have different forecasts. They might also agree on the forecasts and disagree about priorities.

A trust can preserve the authority to make those choices. It cannot make the choices cease to be contested.

Anthropic’s proposed IPO arrangements illustrate the same issue in AI. Its October 26, 2026 filing describes a founder-held special share carrying 50.1% of general shareholder voting power, alongside a separate Long-Term Benefit Trust that would generally elect and remove a majority of the board. These are distinct mechanisms for allocating authority, with conditions and exceptions set out in the filing. (Anthropic’s IPO filing)

The argument for insulating AI decisions from financial pressure is serious. A board may need to delay a profitable release because the evidence suggests unacceptable risk.

Yet people who share a commitment to public benefit can disagree about whether restriction or deployment better serves it. One worries about misuse. Another worries about concentrating access in a handful of organisations.

The mission supplies a reason to take the decision seriously. It does not supply a unique answer.

Some disagreements cannot be solved by more expertise

This becomes particularly important in technology because a technical decision can carry a value judgment a long way.

Consider the threshold in a fraud system. Lower it, and more fraudulent transactions may be caught. More legitimate customers may also be blocked.

Engineers can measure that trade-off. They can improve the model and reduce some errors. Someone still has to decide what balance is acceptable, what happens to the person wrongly blocked, and whether that person can get the decision reconsidered.

A commitment to protecting customers does not settle whose inconvenience or loss should count most.

That distinction matters when an institution appoints experts to guard its purpose. Expertise can improve predictions about consequences. It does not, by itself, resolve whose interests should take priority.

An objection may expose a factual mistake. It may also reveal that the people making the decision have overlooked an interest they do not personally experience.

Both deserve a route into the institution.

Follow the objection

When I want to understand protected authority, I start with someone who thinks it has made a consequential mistake.

An employee has evidence that a strategy is failing. A user believes a platform’s rule is harming the people it was meant to protect. A contributor thinks a technical project has drifted away from its stated purpose.

Where does the objection go?

A feedback channel is a beginning. The harder part is whether it reaches someone with access to the evidence and authority to require a response. If the objection holds up, can the decision change? Can the people making it be replaced?

ICANN, which has a central role in coordinating the internet’s naming system, offers a useful example of making some of these powers explicit. Its Empowered Community has defined mechanisms, subject to formal procedures, for rejecting budgets, challenging certain changes to its rules, removing directors, and recalling the board. Those powers do not establish that every participant has equal influence or that the mechanisms always work well. They do establish that accountability includes specified actions the community can take. (ICANN’s Empowered Community)

That is the level of detail I would want to examine. Who can initiate a challenge? What support must they gather? What information can they obtain? What can happen at the end?

A power may look substantial on paper and be prohibitively difficult to exercise. Its practical cost belongs in the assessment.

There is a real danger in making an institution too easy to redirect. A company can lose its purpose to the next investor. A technical standard can be captured by its largest participants. A public-interest organisation can become reluctant to upset its funders.

But a protected group can also come to treat its own continued authority as necessary to the mission. A challenge to strategy becomes a challenge to purpose. Removing a leader starts to sound like abandoning the institution’s reason for existing.

A mission should be able to survive a change in the people interpreting it.

The question applies to the builder too

This is a question I have to bring back to Sanctity.

If you are building a values layer for AI, you are entering the same problem. Someone chooses which values to name, which questions to ask, and which disagreements the available answers allow people to express.

Those choices shape what can emerge before the first person participates.

At Sanctity, we write the candidate questions and lines. Public answers help determine which lines are proposed. The team retains the decision to adopt them, with the decision and its reason published. It is an early process for a proposed public AI constitution; no AI follows it yet. (How Sanctity’s process works)

That division of authority matters. Choosing among options we wrote is a form of participation. It leaves us responsible for the options, the omissions, and the decision about what enters the constitution.

The harder test comes when someone thinks the question itself is wrong.

Perhaps we have put two different concerns into one answer. Perhaps an option that seems reasonable to us excludes something essential to someone living under very different conditions. Perhaps the majority is comfortable with a rule whose cost falls almost entirely on a minority.

More responses will not automatically repair a badly framed question. We have to be willing to examine the frame.

For Sanctity, that means treating our interpretation as something people should be able to inspect and challenge. As the work develops, that is the test I want Sanctity held to: whether a serious objection can change the questions, the proposed rules, or the process itself.

The concern grows as these choices move from a published constitution into systems that act. A value written into software can govern decisions far beyond the room where it was chosen.

I want to know whose judgment entered that rule, what authority they had to put it there, and how someone affected could get it reconsidered.

Building Sanctity puts me inside those questions. It gives me a responsibility to answer them.