A Receipt, Not a Signature
On 2 August the EU begins requiring AI content to be marked as machine-made, and the rule answers the easy question, which tool touched the file, while stepping around the one that matters, whether a person stands behind it.
On 2 August the European Union switches on the transparency duties of Article 50 of the AI Act. From that date a chatbot has to tell you it is a machine, synthetic audio, images, video and text have to carry a machine-readable mark, and a deepfake has to be labelled as artificially generated. Non-compliance is expensive: fines run to €15 million or 3% of worldwide turnover. Brussels calls this transparency, and after three years of synthetic media flooding every feed the instinct to demand a label is entirely reasonable.
Read what the label actually certifies, because that is where the idea starts to slip. The mark records provenance. It answers a question about origin: was a machine involved in making this. That is the easy question, and it is getting easier to answer trivially, because the honest reply is almost always yes. A machine drafted it, or cleaned it, or graded the exposure, or suggested the phrasing. The rule stamps the tool. It says nothing about the thing a reader actually needs, which is whether a person exercised judgment over the output and can be held to it. A watermark is a receipt, not a signature. It tells you what passed through the till. It does not tell you who signed for the goods.
The statute knows this, and the tell is written into the law itself. Article 50 exempts AI systems that perform an assistive function for standard editing or do not substantially alter the input. Spell-checking is out. Grammar correction is out. The moment you concede that some machine involvement does not count, you have admitted that "machine-made" is not a clean binary but a dial, and that the interesting reading sits somewhere in the middle, where most real work now lives. A law that has to carve out the hybrid case is a law built on a category that does not hold.
Then there is the engineering, which is less forgiving still. Robust marking of this kind is not a solved problem. Text watermarks degrade under paraphrase and light editing. C2PA metadata is stripped by a screenshot or an ordinary social re-upload. Invisible marks like SynthID survive more but not everything, and spoofing can stamp the "artificial" signal onto a genuine photograph. So the receipt fails in both directions at once. It washes off the synthetic content it was meant to flag, and it can be forged onto the real content it was meant to protect. A signal that can be both erased and counterfeited is worse than no signal, because it teaches people to trust the absence of a mark they should never have trusted.
None of this makes disclosure worthless, and the good version of the rule is easy to state. The chatbot that announces itself is genuinely useful, because it corrects a specific deception, a machine pretending to be a person. That is a labelling problem the technology can actually meet. The trouble begins when a narrow, achievable disclosure gets stretched into a universal one, and the label is asked to carry the weight of trust that only accountability can bear.
Because the real transparency question was never about the pixels. It is about the person. When something misleads you, the useful fact is not that a model was in the loop. It is whose judgment stood behind the claim, and who answers now that it is wrong. Provenance is a chain of custody. Accountability is testimony. The law is regulating custody because custody is tractable, and leaving testimony to look after itself, which is the harder and more important half. Robodebt did not fail because nobody labelled the algorithm. It failed because everyone could point to it, and it could point to no one.
So the practical lesson for anyone shipping work made with a machine, which by August is most of us, is not to wait for a stamp to do the work of a conscience. The label on your output will tell your reader which tool you used. It will not tell them whether to believe you. Your name does that. Sign your work, own the judgment inside it, and treat the machine as what it is here, a way to extend how far that judgment can reach. The receipt is for the tool. The signature is still yours to give.