For most of the internet’s history we ran on one quiet assumption: if something was published, a person made it. It might be wrong, biased or plagiarised, but there was a human somewhere in the chain.
That assumption ended this week, and it ended with a compliance notice rather than a manifesto.
Anthropic confirmed it will watermark the text its models generate. Not as an experiment, and not per-product. It is applied at the model level, so it is present whatever Claude surface the text came from, and because the mark is part of the text itself, it travels when the text is copied and pasted. Files carry C2PA credentials. The driver is not Google, and it is not SEO. It is the EU AI Act, whose transparency obligations took effect on 2 August 2026.
Almost everyone read that as a story about detection. I think it is the first visible piece of something much larger: the internet is growing a provenance layer.
Key Takeaways
- From “is this good” to “where did this come from”
- The problem was never fake content
- Fluency is the dangerous part
- Three questions we keep collapsing into one
From “is this good” to “where did this come from”
For thousands of years humans authenticated information through perception and reputation. I saw it. I heard it. I know the person. I trust the publication.
Generative AI attacks every one of those shortcuts at once. A synthetic voice sounds familiar. An artificial photograph looks documentary. A machine-written article sounds authoritative. A fabricated expert appears entirely convincing.
When appearance becomes cheap, provenance becomes valuable. The question shifts from “does this look real” to “can I verify its history”. That is a chain of custody, and it is a genuinely new requirement for a medium that has never had one.
The problem was never fake content
We have always had propaganda, fraud, plagiarism and misinformation. What AI changes is the economics.
One person no longer fakes one article; one system produces ten thousand. A marketer does not write a hundred reviews; an agent generates the personas, the histories and the replies around them. A political operation does not produce a misleading photograph; it can manufacture an entire synthetic information environment around an event that did not happen.
The unit cost of illusion is collapsing. And when deception becomes infinitely scalable, trust cannot stay informal. It has to become infrastructure.
Fluency is the dangerous part
This is the thread running through the first piece in this series. AI is extraordinarily good at making things sound finished. Weak arguments become articulate. Average ideas become polished. Uncertainty becomes confidence.
The model does not need to intend deception. Fluency produces the illusion on its own, and the most dangerous assumption of this era is the shortest one: it sounds credible, therefore it probably is.
Now widen that from writing to photographs, documents, voices, identities, testimonials and corporate communications, which is roughly where the second piece left off. The problem stops being editorial and becomes epistemological. How do we know what we know?
Three questions we keep collapsing into one
Here is where most of the commentary this week went wrong. A verified origin is not a verified truth.
- Was this generated by AI? A watermark can answer this.
- Where did it come from? Provenance credentials can answer this.
- Is what it says true? Nothing in the stack answers this.
A human can lie. A newspaper can be wrong. A camera documents an event without supplying its context. A cryptographically signed piece of propaganda is still propaganda.
Watermarking is also not settled technically. Anthropic has not said how much editing removes its mark, only that it may persist through some. Collapsing all three questions into a single “AI detection” score would just manufacture a new illusion, this time with a verification badge attached.
Provenance tells you where information came from. It does not tell you whether to believe it. That sentence is going to matter a great deal over the next few years.
The uncomfortable part: the web cannot even describe itself yet
There is a gap between the sophistication of this conversation and the state of the actual web, and I can put numbers on it because I measure it.
Across the 1,000,000+ websites in my benchmark, 54.3% publish no structured data at all. Not incomplete markup. None. And 72.4% have no llms.txt, the plain-text file that tells a machine what a site is and which pages matter.
So we are debating cryptographic chains of custody for a web where most sites cannot yet state, in machine-readable form, who they are and what they do. The provenance layer is arriving several storeys above a foundation that was never poured.
That is not an argument against provenance. It is a warning about who benefits from it first. Verification infrastructure rewards organisations already capable of making structured claims about themselves, and most are not.
Where Web3 quietly becomes relevant again
There is an irony here I did not expect to be writing about in 2026.
For years, people in blockchain talked about provenance and almost nobody outside the industry cared. Nobody needed cryptographic proof of a photograph of their lunch. The technology answered a question the mainstream was not asking.
Generative AI asked it. Loudly, and for everyone at once.
This does not mean every article should be minted or every photograph stored on chain. That was always the weakest version of the idea. But the architectural insight underneath it turns out to be the right one: do not merely trust the claim, verify the history behind the claim. C2PA gets there through signed provenance records rather than a ledger, and the mechanism matters less than the principle.
Two waves that looked unrelated are converging. AI has dramatically increased our ability to manufacture reality. Cryptography has quietly increased our ability to authenticate origins. Having spent years building in both, I did not expect the second to be rescued by the first.
The coming trust stack
I think this settles into four layers.
- Origin. Who or what created this?
- Provenance. What happened to it between creation and consumption?
- Evidence. What independently supports the claims?
- Reputation. Has the creator earned enough credibility to be believed?
AI detection solves a fraction of layer one. Everything commercially interesting happens across the other three, and that is where publishers, search engines, AI companies, identity systems and cryptography start converging. Not around content. Around credibility.
Where scarcity moves next
Every technology shift I have worked through follows the same shape. Friction falls, abundance explodes, the abundant thing loses its value, and scarcity relocates.
The internet made information abundant. Social platforms made publishing abundant. Generative AI is making creation itself abundant.
So scarcity moves to trust. Not content, not words, not images, not even intelligence.
The valuable creator will not be the one producing most, but the one whose work carries history, evidence and accountability. The valuable brand will not have thousands of pages, but thousands of reasons to believe them. The valuable AI system may not be the one with the most convincing answer, but the one that can show the strongest chain of evidence behind it.
In the Age of Illusion, authenticity becomes a premium product.
Seeing stops being believing
For most of human history, seeing was believing. The internet weakened that. AI finishes it.
Synthetic media is no longer a category sitting alongside real media. It is becoming the medium. So the defining question of the next internet is not “was this made by AI”, which will soon be true of almost everything and interesting about almost nothing.
It is: what evidence do I have that this deserves my trust?
We assumed watermarking would force machines to disclose themselves. What it may actually do is force all of us to prove who we are.
This is part three of the Age of Illusion series, after When AI Makes Seeing No Longer Believing and When War Becomes Content. For the practical version of the watermarking question, I wrote separately on why Claude’s watermark will not kill your SEO. Part four is The Prison Will Not Look Like a Prison. Part five is The Credibility Layer Just Got a Regulator.
The Age of Illusion III: The Internet Just Got a Chain of Custody
