A Watermark Answers the Wrong Question

In August 2026 Anthropic explained how Claude stamps the content it makes. It does it in two ways. It weaves an imperceptible watermark into the words, and when it writes a file like an image it attaches signed provenance metadata using C2PA, the open content-authenticity standard. The headlines that followed said what they always say when a marking scheme turns up: AI content can now be detected.

I want to slow down on that, because the marks prove a good deal less than the headline suggests, and the gap between that and what people assume is where the interesting work sits.

Read the hedge

Start with the text watermark. The careful wording is that it "may persist through some editing". That one word, some, is doing all the work. Paraphrase the sentence, translate it, run it back through a different model, or edit it hard, and the mark is gone. So the signal is one-sided in a way that matters.

If it's present, it tells you one narrow thing: a Claude model was somewhere in the loop. If it's absent, it tells you nothing at all, because anyone who wanted it gone could strip it in a keystroke. And whoever is motivated to hide that a machine did the writing is exactly the person who'll remove it. Everyone else doesn't care.

A test you can pass by pressing rewrite isn't worth much. It answers "did an AI touch this", sometimes, and it's silent on who published the words, whether they're true, and what you should do with them.

The file mark is more serious, and still narrow

The provenance metadata is the better of the two, and it deserves credit. Look at what Anthropic actually did. Instead of inventing a proprietary mark, it reached for C2PA, an existing open standard the whole industry can read. That's the right instinct, and it's the founding principle of the Gathering. Never reinvent. Fill the gap the standards leave open.

But watch the edge of what C2PA certifies. It records who processed the file. A Claude model touched these pixels, on this date, in this way. It's a manifest about the asset's history. It says nothing about the claim inside. It won't tell you the price in the picture is right, the policy it quotes is real, or the dosage it shows is safe. It attests the bits, not the meaning.

"Made by AI" is the wrong axis

Here's what both marks share. They answer "what tool made this". And that question is quietly going useless, because soon the honest answer for almost everything is yes, a machine was involved.

The question that actually earns its keep is a different one. When a machine decides what to cite, recommend, or buy from on your behalf, it needs to know what this thing is, who stands behind it, whether it's current, and if anyone authoritative says the same. A watermark can't answer one of those.

That's the whole reason I work on MX and REGINALD, and the difference in posture is the part I want to spell out. A watermark is a hidden mark that one side tries to detect and the other works to erase. It's adversarial by design. MX runs the other way. The publisher opts in, declares in the open what the file is, puts their name to it, and any client can check it. Nothing's hidden, and nothing's being hunted.

REGINALD then answers what the watermark leaves on the table. Who published this, and when. Whether it's the current version. Whether the publisher still stands behind it. And whether an authoritative source, somewhere else, states the same claim on its own. Readable, so a machine understands it. Attributable, so it knows who to hold to account. Corroborated, so the claim someone will act on isn't the only place it can be checked.

One honest limit, so I'm not selling detection either

None of this catches a liar, and I'll say so plainly, because it'd be easy to read the whole argument as "so MX is the real detector". It isn't. MX won't unmask a dishonest publisher any more than a watermark will. What it does is make the honest one legible and accountable, and give a machine a reason to prefer the file that declared itself and staked a name on it over the one that turned up with nothing.

That's the shift I want people to make. Stop asking whether a machine made the words. Start asking whether you can trust them. And ask if the next machine down the line can confirm that on its own. The watermark was never going to tell you that. It was answering a different question, and a smaller one.

See it on your own site

If you want the same story from the machine's side, the companion post on how ChatGPT decides which page to cite walks the retrieval end of it: a mark tells a model what made the thing; MX shows it what the words mean and whether to trust them enough to quote.

And to see what an agent actually gets when it reads your own site, start with an AX (agent experience) audit. It fetches your pages the way these systems do, shows you the plain text they receive, and points at the spots where a machine is left guessing about you.