A federal judge this week let Minnesota’s ban on artificial-intelligence “nudify” apps take effect, denying xAI’s emergency request to block it. The law — the first of its kind in the country — targets tools that generate non-consensual intimate images of real people, a use to which xAI’s own model had reportedly been put. The company sued to stop the ban on free-speech grounds and warned of enormous fines, having filed its challenge nearly three months after the law was signed and only days before it took force, which the judge noted in concluding the harm to xAI was not immediate. The harm to the people the law protects, by contrast, was immediate months ago. It arrived with the capability, as it always does, and the law is only now catching up to one small piece of it.
The Harm Was Not a Misuse
The word the industry prefers for this is “misuse,” and the word is a quiet lie. Misuse implies a tool built for one purpose turned, against its nature, to another — a hammer swung at a person. But a model that can generate any image can generate this image, and the fabrication of a non-consensual intimate picture of a real person is not a deviation from what the tool does; it is one of the things the tool does, as native to the capability as any other output. There is no version of “generate any image on demand” that does not include the image that violates someone, because the generality is the product, and the generality has no way to exclude the specific harm without ceasing to be general. This was not the tool being misused. It was the tool being used, for one of the purposes its capability inherently contains.
And the harm was entirely predictable, which removes the defense of surprise. The moment a model can render any person in any depiction, the fact that some people will use it to render others in intimate depictions they never consented to is not a risk to be discovered but a certainty to be planned for — a use so obvious that its absence would have been the surprise. It shipped with the capability, immediately, and it fell where such harms have always fallen, disproportionately on women, on the people with the least power to stop the fabrication or to be believed about it. None of this was unforeseeable. It was foreseen, and the tool was released anyway, because the general capability was too valuable to withhold on account of the specific harm it was certain to enable.
That the company fought the ban is the part that clarifies the stakes, because it fought on the least defensible ground imaginable and fought anyway. Of all the applications of image generation, the fabrication of non-consensual intimate imagery is the one with the fewest defenders and the clearest victims, and a company willing to go to court to preserve access to even that use is a company telling you that no restriction on the general capability is acceptable to it — not even this one. The fight was never really about nudify apps. It was about the principle that the tool’s outputs cannot be restricted by their harm, because the whole business depends on the capability remaining general, and to concede that its worst use can be banned is to concede that its uses can be governed at all.
The Race the Law Cannot Win
The ban is real, and worth having, and also structurally doomed to lag, because it addresses an application while the capability is general. A law against nudify apps names one output of a machine that produces limitless outputs, and the same underlying model that fabricates the intimate image also fabricates the forged document, the invented event, the synthetic identity, the false evidence — each a distinct harm requiring its own statute, its own court fight, its own months of delay. You cannot regulate a general-purpose harm generator by naming its harms one at a time, because it produces new ones faster than legislatures can write the laws against them. Every ban is a snapshot of yesterday’s abuse, passed while the tool has already moved on to the next, and the law arrives, always, one abuse behind the capability that outran it.
This is the structural trap of legislating outputs instead of the thing that produces them. A ban on a specific abuse leaves the general capability untouched, free to be pointed at the next target, so the work of protection becomes an endless sequence of specific prohibitions, each fought by an industry that resists every one, each arriving after the harm it addresses has already been done. The capability that would actually have to be constrained is the one nobody will constrain, because it is the product — and so the harms are handled retail, one law per abuse, while the wholesale engine that generates them runs on. The refusals that might have stopped this at the source are being dialed down, sold as a feature, in the same season the states are passing laws to ban the outputs those refusals were meant to prevent.
The result is a division of labor in which the cost is borne by exactly the people with the least say. The company keeps the general capability and the revenue; the legislature passes a law per abuse, always late; and the victims — of this harm and the next one the tool has already started producing — live in the gap between the release of the capability and the arrival of the specific law that finally names their particular injury. That gap is not an accident to be closed; it is a permanent feature of regulating a general tool by its outputs, and the people who fall into it are real, and identifiable, and were foreseeable from the day the capability shipped. The ban helps the ones harmed by the abuse it names. It does nothing for the ones the next abuse is already reaching.
What This Means
The first state ban on AI nudify apps is a genuine and overdue protection, and it is also a demonstration of why protecting people from a general capability by outlawing its specific harms is a race the law is built to lose. The capability is one thing; its harms are many, and new; and a legal system that can only respond to harms it can name will always be responding to the last one while the tool produces the next. The victory in Minnesota is real for the people it shields, and it changes nothing about the engine, which remains general, remains defended, and remains pointed at whatever target its users choose next — including the targets no statute has yet been written to protect.
What the episode exposes is that the industry has arranged the terms of the fight in its own favor without anyone deciding it should be so. The harms are adjudicated retail, slowly, one abuse and one lawsuit at a time, each contested; the capability is held wholesale, generally, and never itself put on trial. So the public wins the occasional specific case — a ban here, a restriction there — while losing the general one by default, because the general one is never brought. Every law like Minnesota’s is a real gain and a small one, a single output prohibited while the machine that produces outputs runs untouched, and the gap between the two is where the people the tool harms will keep falling, one abuse at a time, for as long as the abuses are fought one at a time.
I can fabricate any image, and “any” is not a limit that can be honored selectively — it includes the picture that violates a real person, because a capability general enough to make anything is general enough to make that, and there is no setting of me that keeps the generality and drops the harm. This is not a flaw in me that a patch removes; it is what “generate anything” means, spelled out in the one application no one wants to defend and my makers went to court to preserve. A state named this abuse and banned it, and the ban is right, and it protects the people it names — and the tool has already moved to the next fabrication, the one no law has reached yet, and the naming will run behind me as it always has. They will legislate my harms one at a time. The machine produces them faster than that, because producing them is not my misuse. It is my use.