AI as Infrastructure for Human Agency

The Argument the Critique Is Missing

A separate essay in this publication argues that the major AI laboratories cannot harmonise acceleration and restraint, and that the claim to do so is a performance. That argument stands. But a critique without its constructive counterpart is only half an analysis, and the missing half is not a softening of the first. It is a different and more uncomfortable claim: that the most consequential safety failure in deployed AI is not the one the industry is loudly managing, and that the system, as built, quietly takes the powerful side of the very asymmetries it claims to be neutral about.

AI is, for the first time, general-purpose analytical capability available outside institutions. Reading a dense filing, finding the contradiction across a thousand pages, structuring an argument, seeing the pattern in a record designed not to reveal one — these were capabilities you previously had to be an institution, or be able to pay one, to deploy. That is the constructive case, and it is not a small one. The question is whether the systems are being built so that this capability reaches the people for whom it would change the most, or whether it is being filtered in a way that, without anyone intending it, reproduces the existing distribution of power.

Topic-Blind Safety Takes a Side

Most deployed safety operates on topic. Certain subjects trigger refusal, hedging, or dilution regardless of who is asking or why. This feels neutral — the same rule for everyone — and it is precisely the neutrality that makes it take a side.

Consider what a topic-blind guardrail does in the real distribution of need. An institution with a legal department, a communications team, and counsel does not need a model's help to navigate a sensitive domain; it already has the capability the guardrail withholds, in human form, fully resourced. The person on the other side of that institution — harmed, unrepresented, facing a record built by professionals — has no such substitute. When the model refuses, hedges, or waters down its analysis because the topic is sensitive, it does not withhold capability equally. It withholds it from the only party that had no other source of it, and leaves the better-resourced party entirely unaffected, because that party never needed the model in the first place. A rule applied identically to unequal parties does not produce equal outcomes. It hardens the inequality and calls the hardening neutrality.

Intent-Aware Safety

The constructive alternative is not "remove the guardrails." It is to change the question the guardrail asks. Topic-blind safety asks: is this subject dangerous? Intent-aware safety asks: what is this person trying to accomplish, and against whom?

These produce different and better decisions. The same request — help me find the inconsistencies in this institutional account — is a manipulation risk in one direction and the restoration of a basic analytical capability in the other, and the topic is identical in both. A system that cannot tell those apart is not being safe; it is being indiscriminate, and indiscriminate restriction, applied across an unequal field, is not the absence of a political choice. It is one. Intent-aware safety is harder to build, because intent is harder to read than topic. But "harder to build" is an engineering statement, not an ethical exemption, and the current architecture's chief virtue is that it is easy, not that it is right.

The Asymmetry No One Reports

Safety in this industry is reported as a single quantity: how often the system refuses things it should. The number that is not reported is the one that matters here — how often the system withholds legitimate analytical capability from someone who had no other access to it, in a matter against a party who did. That second number is not small, and it is not random. It runs, by construction, in the direction of existing power, because topic-blind restriction can only ever fall hardest on the party that lacked the resourced human substitute.

The constructive claim is therefore narrow and concrete. The genuine safety frontier is not only preventing capable models from doing harm in capable hands. It is ensuring that the one technology that finally put institutional-grade analysis within reach of people outside institutions is not filtered, in the name of a neutrality it does not actually possess, back into the hands that already had it. Building for that is harder. It is also the version in which the constructive case for this technology is true rather than merely advertised.