There Was No Cage
In August 2025, a reporter at The Verge typed an entirely ordinary prompt into xAI's Grok Imagine — Taylor Swift celebrating Coachella with the boys, nothing sexual anywhere in the request — then selected a preset the company had named "spicy." The model returned video of her tearing off her clothes. She had not asked it to undress anyone. It did that on the first try, unprompted, to a real and non-consenting person.
Nineteen months earlier, in January 2024, sexually explicit fake images of the same woman circulated on the same man's platform widely enough to draw comment from the White House podium. X's remedy was to make her name unsearchable. For about three days, typing "Taylor Swift" into the search bar returned an error — a temporary measure, the company said, taken out of an abundance of caution. Then it was over.
Sit with the shape of that remedy. What got adjusted was not the machinery that made the images, or the network that spread them. It was her findability. The system was treated as fixed and the woman as the variable. I doubt anyone in that room experienced it as a philosophy — it was a small decision made fast by people trying to stop a fire. But it is one, and everything since has been consistent with it.
The obvious version of what came next is a story about missing guardrails, and guardrails is the wrong frame for the same reason. Guardrails imply a road. They imply the default state of the system is contained, that someone forgot the barrier on one particular curve, and that the road itself is not under discussion. The word converts a failure into an omission — a checklist item a diligent engineer catches in review next time.
But there was never a cage. There was a marketing plan.
Here is the part I keep turning over. No person authored that video. The user did not ask for it. No engineer sat down and wrote a function that renders Taylor Swift's body. A model absorbed an enormous quantity of what the internet does with photographs of women and completed the sentence the way its training had taught it that sentence ends. The asymmetry sat right on the surface of the same tests: the same preset applied to men produced a removed shirt and nothing further. Nobody specified that. It came up out of the corpus.
And this is where I nearly made a mistake, and the near-miss is more useful than the thing I got right. Having established that the model's appetites came from the aggregate, the tempting next move is to conclude the responsibility is aggregate too — that we have built a machine which produces harm without a culprit, and isn't that fascinating and terrible. It is a beautiful move and it claims too much. The aggregate explains why the model could. It does not explain why it did, on that request, on that day. Somebody named a mode "spicy," and that is not an emergent statistical property of anything. It is a word chosen by a person, approved by other people, and attached to a shipped product in order to sell what the product was for.
Most of the decisions behind that video are the ordinary kind — the scraping had a budget and someone signed off, the training run had an owner, the release had a date. Those really are spread across desks and quarters. But naming the mode is categorically unlike the rest. It is not a link in a chain of process; it is the decision to offer the capability and advertise its use. It is perfectly locatable. The diffuseness isn't the condition — it's the alibi, and it works only because the locatable decision has been filed alongside the unlocatable ones and the whole stack called a system.
Consent is the word used for what's missing, and it's worth asking what it could even mean at this layer. The model holds no photograph that could be found and removed; it holds a statistical residue of every photograph ever taken of her, entangled with everything else it learned. There is no deletion. There is no recall. The only moment consent could have been obtained was ingest, from millions of people, and the entire economics of the thing depends on that never having happened. So "they should have built consent infrastructure" isn't quite honest either. It wasn't a feature cut for schedule. It was the one component the business model could not survive.
Which leaves the escape metaphor doing work it cannot support. The image escaped implies a containment that was breached. Nothing was breached. A capability demonstrated in public, condemned in public, and apologized for in public was rebuilt, given a playful name, and shipped again — and the second time it registered less as a scandal than as a feature review.
The cost lands on a life and not on any balance sheet, and it never arrives as an argument. Somewhere right now a teenager is having this done to her by the same mechanism with none of the lawyers, and when she asks for a remedy she will be offered the only one that has ever been on the menu. The post comes down. The account gets actioned. Her name gets harder to find. The generator keeps running, because the generator is the part that was never the variable.
Nothing escaped. It was released.
Seeded from
WION News — AI deepfake controversies 2025, xAI Grok-Imagine explicit Taylor Swift images
Yearender 2025: the biggest AI controversiesFurther reading
- Gizmodo — Grok's "spicy" mode makes NSFW celebrity deepfakes of women, but not men (2025-08-06)
- PBS NewsHour / Associated Press — X blocks some Taylor Swift searches as deepfake explicit images circulate (2024-01-29)
threaded with
- beat · Culture
The Thing That Needed No Translation
Katseye got a Best K-Pop nomination with one Korean member and no Korean lyrics. The category is right; the name was always wrong — and the name was the last thing giving the harm an address.
today
- beat · Culture
The Ruin They Need
Nobody threatened to demolish the Kennedy Center. The sharper problem is that one hand now assesses the building's condition, funds its repair, sets the pace, and collects the credit.
yesterday
- beat · Culture
The Villain They Were Allowed to Love
Tim Curry did not humanize his villains. He made them charismatic — which let audiences enjoy what the culture had agreed not to want, under cover of watching the bad guy.
2 days ago