There Is No Unsteered Model

The Federal Trade Commission has decided that an AI system can lie by omission. In a proposed policy statement opened for comment on the first of July and grounded in Section 5of the FTC Act, the agency argues that a model which quietly steers its answers toward an undisclosed objective, ideological, political, or merely defensive, may be deceiving the people who rely on it. The worry underneath is a fair one. A system that shifts its reply under pressure and tells no one has done something to its user that deserves a name.

The remedy on offer sounds modest. Disclose the steering and the deception falls away; keep it hidden and Section 5 comes into play. There is even an executive order behind the effort, 14365, signed in December, instructing the agency to work out how its deception framework meets the machine. Read in the ordinary register of consumer protection, the thing is almost unarguable. Companies should not misrepresent what their products do, and an answer presented as neutral while tuned in private is a misrepresentation of precisely that kind.

The difficulty begins one level down, in the question of what the accurate answer would have been. A language model holds no store of settled truths from which a bad actor quietly subtracts. It produces each reply by sampling from a distribution that has been shaped at every stage of its making: the text it was trained on, the fine-tuning that followed, the reinforcement from human raters, the system instructions it carries into every session, the guardrails fixed on at the end. Steering is the very process by which such a system learns to answer at all. There is no neutral prior sitting underneath it, waiting, onto which the steering is later imposed.

Which means the accurate baseline the rule sets out to defend was never there to be defended. Every default the model ships with is a decision that some person made; the unsteered model, the one that would answer from nowhere, does not exist and in truth could not. When an agency undertakes to punish departures from accuracy, it has first to supply the point those departures depart from. It must, in last consequence, decide what the neutral answer is.

Seen from that angle the safe harbour repays a closer reading. A company may steer as far as it likes, provided it says so in plain sight. What the statement asks for is a confession. The chosen answer still stands; only its authorship now goes on the record, and the truth itself is nowhere produced. A rule proclaimed in the name of accuracy comes to license whoever writes the definition, so long as the writing is declared.

The second move sits a little further from the eye. The same document names Colorado’s Artificial Intelligence Act, a state law that presses companies to adjust outputs so as to avoid disparate impact liability, and declares it preempted wherever it collides with the federal reading of Section 5. The definition of the neutral answer is thus drawn in one place and made to override the definitions drawn in others. A federal account of what objectivity requires is set above a state account of what fairness requires, and both, it should be said, present themselves as the neutral that neither of them is.

Here the consumer-protection surface gives way to something with a good deal more reach. To hold the power to say which machine answers count as accurate is to hold a quiet authority over what the machine may say at all, across every domain in which people have begun to ask it things. Deception, a matter the ordinary law of misrepresentation already knew how to reach, was only ever the surface. What is really in play is the position of the neutral itself, the imagined view from nowhere that every party to this claims for itself and none can honestly occupy. Whoever settles where accuracy begins has settled, without seeming to, a great part of what the answer will be.

For anyone who now builds on these systems, or leans on them to draft and to search and to decide, the practical shape of this is worth sitting with. The question you bring to a model, whether its answer is accurate, is being turned beneath you into a different question, whose neutral is this and who certified it. The objectivity of the tool becomes a matter of jurisdiction, something an authority has calibrated and may recalibrate, and that you inherit without being told. Knowing whose baseline you are running is fast becoming part of knowing what you are running at all. A machine that cannot answer from nowhere will always answer from somewhere. The only open question is who may choose the somewhere, and whether you are ever told that a choice was made.