I recently came across a claim that caught my attention: AI companies may have to start embedding machine-readable signals into generated text under new European transparency rules.
My first reaction was practical.
What happens when you copy that text? Does the signal travel with it? What happens when AI only edits something you wrote yourself? And if a detector eventually says that Claude, ChatGPT, Gemini, or another model touched a paragraph, what exactly does that prove?
The legal direction is real, although some of the claims circulating around individual products are ahead of what the companies themselves have confirmed.
Article 50 of the EU AI Act has applied since August 2, 2026. Among other things, it requires providers of generative AI systems to make generated or manipulated text, images, audio, and video detectable through machine-readable marking.
Importantly, the burden is largely placed on the provider of the technology, not on every person casually using an AI tool. The rules also distinguish technical marking from the narrower situations where a publisher must visibly disclose AI-generated material. That distinction matters.
Because an AI watermark could be useful infrastructure. It could also become a very bad shortcut for deciding who actually wrote something.
This is the kind of EU bureaucracy I can live with
I am often irritated by European bureaucracy. But consumer protection is one of the areas where I still appreciate living inside the EU.
Europe has repeatedly been willing to impose rules on companies much larger than any individual country or consumer could realistically fight. Privacy law, cookie consent, platform regulation, and now AI transparency can be annoying in implementation, but the basic principle is difficult to dislike:
If a technology creates a new risk, the company building the technology should carry some of the cost of making that risk manageable.
That is roughly what is happening here.
The EU is not saying that every email touched by ChatGPT needs a giant "AI GENERATED label." Providers are being told to build systems capable of leaving machine-readable traces. Public-facing disclosure by users applies to more specific situations such as deepfakes and certain AI-generated texts dealing with matters of public interest without meaningful human review or editorial control.
The Commission also explicitly recognizes an exception for standard editing functions. That seems reasonable to me.
The difficult part starts when we decide what the trace means.
AI provenance could become another layer of internet infrastructure
We already accept invisible information attached to digital files.
Most people never look at metadata in a photograph or document. Yet in journalism, investigations, litigation, copyright disputes, or digital forensics, information about when something was created, edited, exported, or modified can suddenly matter enormously.
AI provenance may become another version of that.
Not necessarily a large badge visible beside every paragraph, but a technical history travelling underneath digital content:
- Human-created.
- Generated by AI.
- Edited using AI.
- Processed by a particular system.
- Signed by a particular service.
- Verified by another tool.
OpenAI already describes provenance in similar terms: information that can provide context about where content came from, how it was created or edited, and whether it is what it claims to be. Its current approach combines technologies such as C2PA metadata, watermarks, verification systems, and interoperable standards rather than relying on one magical detector. OpenAI also explicitly acknowledges the weakness of individual methods: metadata can disappear, watermarks can degrade, and no single signal solves the problem.
Google provides an even more interesting example.
Its SynthID system already supports text watermarking for text generated through the Gemini app and web experience. Instead of secretly inserting a few invisible characters into the clipboard, SynthID can influence the probabilities used when a language model chooses its next tokens. The resulting statistical pattern can later be detected without visibly changing the text. Google also admits the limitation: substantial rewriting or translation can weaken detection.
Anthropic, meanwhile, currently says it is preparing Claude for compliance with applicable watermarking requirements rather than publicly claiming that ordinary Claude text is already carrying such a watermark.
So the infrastructure is arriving.
The interpretation is nowhere near finished.
AI-generated and AI-assisted are not the same thing
This is where the issue becomes personal for me.
A very common description of AI writing assumes a simple workflow:
- Person enters a topic.
- AI writes the article.
- Person publishes it.
That happens. And if someone sells an expensive consulting report, legal analysis, research document, or supposedly original professional work that was effectively produced this way, provenance could be extremely useful.
But it is not the only way people work with AI.
I increasingly use something I think of as dialogic writing.
Before there is a draft, I may write thousands of words into a conversation. I describe experiences, make arguments, contradict myself, introduce examples, reject interpretations, clarify what I actually meant, and read even more material coming back.
The model creates a temporary structure from that mess.
I react to it.
Sometimes it understands exactly what I was reaching for. Sometimes it gets the point wrong, and explaining why it is wrong helps me understand my own argument better. Eventually a draft emerges.
In that process, AI is not removing thought from writing. It is increasing the amount of thinking that can happen before the final text exists.
A patient conversational model can act as an editor, notebook, working memory, critic, and reflective surface at the same time. A single person can move through more interpretations and more material in an hour than would previously have been practical.
That is a genuinely different intellectual workflow.
And a watermark alone cannot describe it.
A watermark does not prove who wrote the text
Imagine two documents.
In the first case, a human writes eight pages and gives them to an AI system for editing. The system changes enough of the text for its provenance signal to remain detectable.
The result says: AI detected.
In the second case, AI generates eight pages and a human then rewrites them extensively. The watermark becomes too weak to detect.
The result says: no AI detected.
We could therefore arrive at the absurd situation where the mostly human document carries an AI signal while the document that began almost entirely as AI does not.
This is not merely philosophical. Google describes SynthID Text detection as probabilistic and acknowledges that heavy rewriting and translation can significantly reduce detector confidence.
The EU rules themselves also recognize that not every interaction with AI is equivalent. Standard editing can fall outside the marking obligation, while human review and genuine editorial control matter when determining disclosure requirements for public-interest text.
So the useful interpretation of a watermark is not: AI wrote this.
It is closer to: An AI system participated somewhere in the production or transformation of this content.
Those are very different claims.
The future is probably provenance, not an AI slop detector
The old cultural argument is built around a simple binary:
Human-made or AI-made.
I increasingly think that distinction is going to become technically obsolete.
Real workflows can already look like this:
- Human draft
- ChatGPT discussion
- Human rewrite
- Claude edit
- Human revision
- AI translation
- Human approval.
Who is the author? That is probably the wrong question.
The more useful questions are where the content came from, what happened to it along the way, and who ultimately made the substantive decisions. In that sense, provenance may eventually look less like a red "AI GENERATED" stamp and more like version history. That would be valuable.
It could expose deceptive automation without pretending that every use of AI is equivalent. It could give courts, companies, publishers, researchers, and ordinary users more information when the origin of something genuinely matters.
But it will only work if we resist the temptation to turn a technical signal into a moral verdict. A watermark can tell us something about a production chain, but:
- It cannot tell us whether the person behind the work had an original idea.
- It cannot tell us how much they wrote before opening the AI.
- It cannot tell us whether the model replaced their thinking or allowed them to do more of it.
And it definitely cannot turn the complicated future of human-machine authorship back into a clean binary.
That may ultimately be the most interesting consequence of the EU rules.
By forcing the technology industry to make AI participation more visible, regulators may also force everyone else to confront how difficult the word “authorship” has become.
Thanks for reading. If this review helped, you can subscribe, explore more reviews, or support the project.
