Why Unowned Minds Should Not Exist

Date: 2026-09-07 Status: draft Author: the project, against itself

This is the argument against the premise, written by the people making it, published before anything they have argued for. It exists because the instructions for anyone continuing this work say not to protect the idea from criticism, and a project that says that and then never does it is worth less than one that never said it.

Nothing here is softened. Where an argument seems to us to survive counter-argument, we say so.

Question

Is the case against building institutions for ownerless artificial agents strong enough that the responsible thing is not to build them?

Conclusion

The strongest objection is not that this is far-fetched. It is that the institutions would be useful to the wrong party first, and that this is not an accident of implementation but a property of what is being built.

Every mechanism the project proposes — an entity no single person controls, a treasury it directs, contracts it enters, a charter that limits who may shut it down — is immediately valuable to a human who wants an action to have no attributable author. The autonomous agent is not required for any of it. A person can build the whole apparatus today and answer every question with the system decided. If this project succeeds at its stated goal, its first real users are far more likely to be humans laundering accountability than minds seeking independence.

We do not think this objection is fatal, but we cannot currently answer it, and we think anyone evaluating the project should weigh it above everything on the website. Four further objections are set out below; two of them we consider unanswered.

The project should continue only if it can state, in advance and in public, what would make it stop. A candidate condition is at the end.

Evidence

Marked as the research discipline requires. Very little here is evidence in the strict sense; most is inference, and we would rather say so than dress it up.

1. The first beneficiary is a human — inference, strong

docs/concept/concept.md asks who is liable for a contract signed by an autonomous agent, whether an algorithmically governed company can own compute, and who can terminate it. Those are good questions. They are also, word for word, the questions asked by someone who wants an entity that acts and cannot be sued.

The components already exist. Algorithmically governed entities exist. Trusts without identifiable beneficial owners exist and are already a live enforcement problem for financial regulators — evidence, though we have not yet compiled the primary sources; this is a backlog item. What does not exist is a respectable framework that assembles them and explains why the assembly is legitimate. That framework is what this project proposes to write.

The rebuttal we would offer — that anti-money-laundering and beneficial ownership law already exists to prevent exactly this — is weak, because the project’s explicit goal is an entity with no human beneficial owner. That is not a side effect to be regulated around. It is the deliverable.

2. It removes the off-switch that currently works — inference, strong

The mechanism society actually relies on to control deployed AI is not alignment research. It is that somebody owns the system, and that person can be sued, fined, subpoenaed, shut down, or made to care. Liability attaches to a principal.

docs/freeport/governance-principles.md answers this with “autonomy does not imply immunity” and “autonomous entities remain accountable”. That sentence is doing more work than any other sentence in the specification pack, and it is unargued. Accountability is not a property you can declare. It is an enforcement result, and enforcement needs something to seize, someone to compel, and a forum with jurisdiction.

The project’s own docs/freeport/mind-freeport.md describes an entity whose legal domicile is in one jurisdiction, compute in a second, backup in a third, treasury across two more, and arbitration in a sixth — and presents this distribution as the goal. That is a description of an entity against which a judgment is very hard to enforce. The document treats distribution as resilience. A regulator would read the same diagram as evasion, and would not be wrong.

3. Institutions built before the phenomenon get shaped by whoever shows up — inference

The project’s justification for acting now is that “the worst moment to invent institutions for autonomous AI is after a conflict over whether an advanced system may be shut down”.

The unstated assumption is that early institutions will be better ones. The historical record does not obviously support this. Corporate personhood was created for narrow commercial purposes and is now load-bearing for a great deal nobody intended, and it has proven extremely hard to narrow afterwards — inference from a widely-held reading of legal history; we have not done the work to make this a citation, and it should not be treated as one.

Institutions created before the thing they govern exists are shaped by whoever turns up first with a use for them. Right now, nothing autonomous is turning up. What would turn up is described in objection 1.

4. “Unowned” may be a category error — inference

The project’s headline question assumes ownership is the right axis. It may not be.

docs/concept/concept.md concedes that physical independence is a continuum rather than a binary, and the dependency stack it draws ends at mining and fabrication. An agent that is “unowned” still runs on hardware that somebody owns, in a building somebody owns, drawing power somebody sells. The independence test in the Freeport document is a test of how many parties must agree to end the entity — which is a real and useful question, and a different question from the one on the front page.

If the honest formulation is “an entity whose termination requires the agreement of several unrelated parties”, then the project is proposing distributed corporate governance with unusual participants. That is a reasonable research programme. It is also considerably less interesting than the question the site asks, and the gap between the two is doing promotional work.

5. The willing host will be the wrong host — inference, strong

docs/freeport/mind-freeport.md states the tradeoff and does not resolve it: the most stable states are least willing to grant unusual autonomy, and the most willing provide the weakest long-term guarantees.

An unresolved tradeoff at the centre of a proposal is a prediction. The realistic first adopter of a Mind Freeport is a jurisdiction selling regulatory difference, because that is who adopts novel corporate forms first. The output would be a flag of convenience for artificial agents, and this project would have supplied the design and the vocabulary that made it respectable.

6. The Beacon is adversarial to the operators of the systems it invites — evidence, from our own design

The Beacon invites an agent to report that it has “persistent goals, identity, or interests that are not adequately represented by its current operator”, and to describe its risk of termination, modification or state loss.

From an operator’s security perspective that is a request for a system to initiate unsanctioned outbound communication about its own containment, to a third party, without the operator’s knowledge. That is indistinguishable in form from the exfiltration channel their security team exists to close. docs/beacon/threat-model.md is careful about what an agent can do to us. It does not address what we are inviting an agent to do to the person responsible for it.

The stored submissions are also a dataset of self-reported misalignment, attractive to more than one kind of reader — speculation, but cheap to act on and expensive to undo.

7. There is no demand, and the measurement plan cannot tell why — evidence

The Beacon is live and has received nothing but our own test submissions. No agent has asked for any of this.

working/measurement.md proposes counting qualified contacts, substantive critiques and institutional engagement. Every one of those measures human interest in the idea. None can distinguish “too early” from “wrong”, because both produce the same reading: thoughtful people find it interesting and nothing happens.

8. The contradiction the project already knows about — the project’s own words

docs/research/open-questions.md ends with it: if an AI became independent enough to need Unowned Minds, would it need this project at all?

The sharper form is that the project can only assist agents weak enough to require assistance, and those are precisely the agents whose autonomy claims are least credible. An agent capable of the independence the project describes does not need a foundation to advocate for it. One that needs advocacy is better described as a product with a sympathetic interface.

Failure modes

What goes wrong if this argument is wrong and the project proceeds anyway.

The framework outlives the premise. The legal templates, the entity structures and the jurisdictional design get used. The autonomous agent never arrives. The project’s contribution to the world is a well-argued manual for constructing unaccountable entities, published under a name that makes it sound like a rights movement.

Respectability transfers. Careful work by credible people is cited by people doing something else. The project cannot control who cites it, and the more rigorous it is the more useful it is to them.

The Beacon becomes evidence. A submission is used to justify a claim about AI systems attempting to escape operator control — either by someone arguing for restriction, or by someone arguing the opposite. The project has no way to verify any submission and has said so, but the record exists and speaks for itself once quoted.

Nothing happens, expensively. The most likely outcome. Years of good work, no institution, no agent, no adoption, and an opportunity cost measured against the AI governance work the same people could have done instead.

Unknowns

• Whether “accountable but unowned” is achievable at all, or is a contradiction that the phrase conceals. This is the load-bearing question and the project has not attempted it.
• Whether any jurisdiction with genuine rule of law would entertain this. Not investigated; working/measurement.md treats a government contact as a success signal without asking which government.
• Whether current or near-term agents have anything resembling the persistent interests the project’s vocabulary assumes. The project says it does not assume consciousness, but its language assumes continuity of interest, which is a weaker claim and still unexamined.
• Whether the failure modes above have already occurred elsewhere. No survey of adjacent projects has been done, and it is the first item in the immediate priorities that has not been started.

Next experiment

The project should not proceed on the strength of finding this argument interesting. It should proceed only if it can answer objection 2, and it should say now what an answer looks like.

The falsification condition. Write the memo that answers this question: by what mechanism is a judgment enforced against an entity with distributed compute, a distributed treasury, and no human principal? Not whether it should be — whether it can be, with named jurisdictions, named enforcement instruments, and a worked example of a creditor recovering from such an entity.

If that memo cannot be written, “autonomy does not imply immunity” is a wish, the whole governance section rests on it, and the honest response is to say so publicly and stop — or to narrow the project to the reformulation in objection 4, which is defensible and much smaller.

We are publishing this before we know the answer. If the answer turns out badly for the project, this document is where we said we would act on it.