Some safeguards exist only because a system cannot do something. Nobody named them as safeguards, so nothing protects them. When the capability arrives, it does more than add an option. It removes a limit, and the protection that limit happened to provide. What was a wall becomes a door policy, and a door policy is much easier to argue away than a wall.
The essay on friction as accidental governance describes those limits. A review happened because the system could not decide by itself. A cooling-off period existed because the system could not act instantly. An exception route existed because the system could not handle the case. Each limit was doing constitutional work, and nobody built it to.
The default is deployment
Organizations treat a newly possible capability as a decision already made. The question shifts, usually without anyone asking it, from should this exist to under what conditions should we deploy it. The conditions are real work, and the standards on this site are for that work. But the work all happens after a decision nobody took. A capability that can infer emotional state, health risk, employability, political preference, or behavioral intent arrives with its answer attached.
The non-conversion principle says a capability does not become authority on its own. A tool being reachable is not permission to use it, and a pilot going well is not a rollout. The principle keeps capability and authority in separate records, so that widening one does not silently widen the other. But it assumes a grant decision comes next. There is a decision before the grant that the principle does not force anyone to consider: that this capability should stay outside institutional decision-making, whatever its accuracy, its efficiency, or what competitors are doing with it.
What refusal protects
The case for non-use is not discomfort with technology. It is about jurisdiction. A deployment review asks whether a system can be halted, explained, appealed, and repaired. The standards assume an earlier question has already been answered: should this be a place where institutional decisions happen at all?
Some inference fails that earlier question. A model that estimates a person’s emotional state from their face can be accurate, audited, and appealable, and the arrangement can still be illegitimate. The question is not about the model. It is whether the institution may hold that inference about a person at all. No amount of procedure around the inference answers that. The standards can make an arrangement governable. They cannot make every operation one an institution ought to be conducting.
This is the strongest form of bounded jurisdiction. The ordinary form says that more capability does not produce more legitimate authority, so each widening needs its own justification. Non-use says that some capabilities should stay out of institutional decision-making even when they work. The work they would do is not work an institution can be made to answer for.
At least one legislature has recorded a refusal of this kind. The EU AI Act, Regulation (EU) 2024/1689, prohibits in Article 5(1)(f) AI systems that infer the emotions of a person in the workplace or in education institutions, except for medical or safety reasons. The prohibitions in Article 5 have applied since February 2, 2025. The rule does not depend on how accurate the inference is. It names a place where the inference may not be made.
Why it is hard to hold
A refusal has no constituency. It produces no product, no efficiency claim, no metric improvement, and no case study. The person it protects is someone a thing did not happen to. Nobody reports on harm that did not occur.
Meanwhile the ratchet runs. A system that works gets used more, trusted more, and extended further. Each extension is small enough to look like a condition on deployment rather than an answer to whether to deploy.
So non-use has to be a standing decision, not something held together by whoever is paying attention. The capability catalog and the authority grant are separate records so that we can and we may do not collapse into each other. The expansion rules exist so that every widening is someone’s decision. Non-use is the same machinery pointed the other way. It is a recorded refusal, with its basis, its scope, and the conditions under which it would be revisited. It has the same standing as a grant.
Prohibitions replace incapacity
As capability accumulates, a safeguard that came from structure is replaced by one that has to come from politics. When a system could not read faces, there was no decision to make about face-reading. When it can, someone has to maintain the absence of face-reading on the record, against pressure from everyone who would gain from the capability. The friction essay shows what disappears when friction is optimized away. What has to be built in its place is named boundaries, held against the pull to remove them.
The law this answers to is the one about preserved capacities. An institution keeps some capacities because it hopes never to need them, and treats their erosion as a governance event rather than an efficiency gain. Non-use applies the same posture one level up. The institution keeps some capabilities in its catalog and declines to deploy them. The refusal is deliberate, recorded, and revisable. An accidental refusal can be dropped without anyone noticing.
What a deployment is for
A discipline that only asks how to use a capability safely will find a way to use it. The question assumes the deployment and works out the conditions. The fuller question is what a deployment is for. One honest answer is that the capability should not come within institutional reach at all. A field that can hear that answer, act on it, and record the refusal keeps an option nothing else in the method supplies. In the more common pattern, every capability that becomes feasible is eventually deployed, and the safeguards are negotiated after the practice exists.
What the laws take from this
These laws give a refusal the weight of a decision. The admission gate is where each one is recorded.
Law I keeps capability and authority in separate records. A recorded non-use, with its basis, scope, and review conditions, therefore has the standing of a grant, not the status of an omission.
Law XI makes expansion a decision rather than a reward for success. That keeps “conditions on deployment” from quietly answering “should this exist at all.”
Law XII preserves on purpose the capacities that contestability depends on. Non-use extends that posture to capabilities the institution declines to deploy.
A reader need not agree that any capability should stay unused to accept the record-keeping. A refusal that is not recorded is not a decision. A widening that is not a decision is drift.