Let’s talk about the week the machine was legally required to stamp its own work — and why, when you actually went to check, most of the stamps weren’t there.
Here is what happened, because most people missed it entirely. On the second of August, two rules switched on in the same month. California’s AI Transparency Act started requiring every big AI company with over a million users to hide a label inside the images, audio, and video their tools produce — and to publish a detection tool so you could actually read that label. Across the Atlantic, the EU’s AI Act Article 50 went live at the same time, forcing AI providers to mark generated content. It is the first time in history that the machine has been legally ordered to tell you it is a machine.
That is a genuinely big deal. For once, the answer to synthetic media is not “teach everyone to be suspicious of everything.” It is provenance. The content carries its own birth certificate. That is the right instinct, and I’ve argued for it in here before. Proof over suspicion. Proof has to scale. Suspicion never does.
Then the real world tested it. And the real world does not read the press release.
An independent review by Indicator and the human-rights group WITNESS checked thirteen major companies that generate images and audio. This is the actual test the new law is supposed to pass. Here’s the number that matters: only seven of them had published the detection tool the law requires. An eighth, the music company Suno, produced one after being called. That leaves HeyGen, Midjourney, Mistral, Synthesia, and xAI — some of the most-used generation tools on the planet — with no tool available at all.
Let me be plain about what that means. The law went live on August 2. Weeks later, five of the biggest AI companies in the world still had no way for you to check whether their output was theirs. Not a technical limitation. Not a legal grey area. They just hadn’t done it. And nobody has been fined. The attorney general’s office, asked whether anyone has been penalized, said they can’t comment on potential investigations. That is not enforcement. That is a suggestion with a deadline.
And here is where it gets worse than “they didn’t bother.” Among the companies that did ship a tool, the test found something almost nobody is talking about. Most tools could only identify their own unedited content. Only Google and OpenAI could catch their own output after it had been edited. Only Adobe and Microsoft could identify another company’s unedited content more than half the time. So your Midjourney image runs through a filter, a crop, a colour grade — and the stamp is gone.
OpenAI said it themselves, almost proudly. Quote: “Metadata is not foolproof. It can be stripped, lost through uploads and downloads, or broken by transformations like file format changes, resizing, or screenshots.”
That sentence is the whole problem in a single box. The single most common way a fake image travels in 2026 is a screenshot. Someone screenshots the AI image and sends it on WhatsApp. The screenshot strips the metadata. The stamp dies in the exact place people actually share content. The tool that was supposed to be the answer quietly admits it does not survive the journey the content actually takes.
Now flip it around, because there’s a second failure that’s harder to see. Meta started tagging ads “created or significantly edited by generative AI” earlier this summer. Last week, Instagram marketers noticed Meta’s filters flagging content made with Canva’s background-remover as “AI info.” Not a deepfake. Not a synthetic person. You removed the background from a photo of your own product, and the machine stamped it as AI.
So put both failures together. One one side, the stamp gets stripped so easily it’s decorative. On the other side, it’s so aggressive it lands on ordinary human work. The system over-flags a background-removal and under-catches an actual clone. That is not a bug in the rollout. That is the shape of the thing. A stamp that fails in the direction that protects the platform, and over-fires in the direction that burdens the human.
And that’s the part people are missing. We keep treating the label as if it were a verdict. “It’s stamped, so it’s fine. It’s stamped, so it’s fake.” But a stamp that half the builders don’t provide, that can’t survive a screenshot, that the public can barely run — OpenAI’s verification tool blocked testers after as few as seven checks in a session, Google and Meta capped theirs at ten to fifteen a day — that is not a verdict. It is a floor. A very low floor.
I want to be fair, because this genuinely matters and I don’t want to be the cynic who sneers at the one good idea. Provenance infrastructure is the only answer that scales. Anthropic is even watermarking text now — statistically, invisibly, inside the word choices of every Claude output, under the EU Code of Practice. That is real engineering. It is the right direction. Text watermarking is the strongest version of this idea we have.
But here is the truth that the next eighteen months will decide. A stamp is only worth something if the person holding it can actually check it, and if the absence of the stamp means something. Right now neither is true. Half the tools don’t exist. The ones that do can be killed by a screenshot. And there is no legal force behind any of it yet — no fines, no enforcement, just “we can’t comment on potential investigations.”
So the sober reading is this: the machine has been told to introduce itself, and it has mostly declined. The infrastructure exists on paper, and on paper it looks like we finally fixed the problem. In practice, it’s a birth certificate written in a language the public doesn’t have a dictionary for, printed on paper that dissolves when it passes through the places we actually live.
The principle that has to survive is not “label everything.” It’s that a mark means nothing until the person who sees it can check it cheaply, and until its absence means something. We don’t need more stamps. We need stamps that survive the screenshot, that the ordinary person can run without a rate limit, and that someone actually enforces.
Because right now, the machine has been ordered to stamp its own work — and it has responded by printing the stamps where nobody can read them.
That is not a fix. It is a floor that’s pretending to be a ceiling. And the people who get hurt while we admire the floor are the same people who always get hurt: the ones who trusted the label because they were told it would protect them.
A stamp you can’t check is not proof. It’s the appearance of proof. And in a world where the machine can fake anything, the appearance of proof is exactly the thing we can no longer afford.