It Signs Its Work

On August 2, something changed about every piece of text Claude writes, and almost nobody who uses it for work noticed.

Anthropic began embedding invisible watermarks into the text its models generate. Not a footer. Not a disclaimer. A statistical fingerprint baked into the word choices themselves, detectable by machine, invisible to you. Files get treated differently, with signed provenance metadata riding along in the PNGs and JPGs and SVGs using an open standard called C2PA. The reason is the European Union's AI Act, specifically Article 50, whose transparency rules went live that same day. Anthropic signed the bloc's Code of Practice on transparency of AI generated content, alongside Google, Meta, Microsoft, OpenAI, and a couple hundred other companies.

Here's the part that matters to you, sitting in Ohio or Tampa or wherever you run your business from. Anthropic didn't limit the marking to Europe. It applies worldwide. Every Claude output, everywhere, no opt out.

I want to be careful here, because there's a version of this story that gets people worked up for no good reason, and there's a version that should genuinely change something about how you operate. Let me give you the second one.

First, the calm part. A watermark on AI text is not an accusation. It's a record that a machine touched the words. That is a different claim than saying you didn't write something, or that you cheated, or that the work isn't yours. If you dictate a rough draft, hand it to a model to tighten, then rewrite half of it by hand, the finished piece may still carry a mark. The mark says the text passed through a model. It does not say who did the thinking. Anybody who conflates those two is making a logical error, and eventually the market will sort that out.

Second, the part where you should sit up. It doesn't matter what the mark technically means if the person on the other end of your document interprets it differently. Procurement officers, loan committees, licensing boards, and enterprise legal departments are not going to read the technical documentation. They're going to run a detector, get a result, and make a decision. You will not be in the room to explain the nuance.

So this Monday's audit is not about whether AI writing is okay. You already decided that. This is an inventory. You're going to find every place in your business where machine generated text leaves your building, decide which of those places carries real consequence, and put a position in writing before somebody asks you for one.

It takes about ninety minutes. Let's go.

Most people audit this backwards. They open their tool list and ask what they use AI for. That gives you a tidy inventory of software and tells you almost nothing useful, because the risk isn't in the tool. The risk is in the destination.

Pull up a blank page and write one column: every category of document that leaves your business and lands in front of someone who is evaluating you. Not reading you. Evaluating you. There's a difference between a newsletter subscriber and a bank underwriter, and the difference is what happens to you if they decide something.

Here's my list. Yours will look different, but this gets you moving.

Proposals and bids. Statements of work. Client deliverables that get filed as work product. Loan and grant applications. Insurance submissions. Anything filed with a licensing board or regulator. Legal correspondence. Contracts you drafted rather than signed. Job postings, offer letters, and anything that goes to a candidate. Performance documentation for employees. Investor updates. Published content under your name. Expert testimony, affidavits, or anything sworn.

Now go down that list and mark each one with the honest answer to a single question: if the recipient ran this through a detector and it came back marked, what would actually happen?

For most of them, the answer is nothing. Your newsletter gets marked, and the world keeps spinning. Your job posting gets marked, and no candidate cares, because everybody knows job postings have been machine assembled since long before the models got good.

But three or four items on your list will make you pause. That pause is the whole point of the exercise.

In my experience, the trouble clusters in four places, and they're not the ones people expect.

Bids into large organizations come first. Enterprise and government RFPs have started carrying AI disclosure clauses. Some ask you to attest that submitted material is original work. Some ask you to disclose AI assistance. A few, badly written, ask for both in ways that contradict each other. If you're bidding into that world and you've been treating those clauses as boilerplate, you now have a problem that a watermark makes legible. Read the next one you get. Actually read it.

Anything sworn or attested comes second. Affidavits, regulatory filings, certifications where you're signing under penalty of something. The exposure here isn't reputational. It's legal, and the standard is not whether AI assistance was wrong. It's whether your attestation was accurate. Draft those by hand or draft them with a lawyer.

Third, and this one surprises people, is client work that gets resold. If you write a market analysis for a client and that client folds your pages into a deck that goes to their board or their investors, your marked text just traveled somewhere you didn't authorize and can't explain. You lose control of the context at the handoff. Ask yourself which of your deliverables get repackaged downstream, and treat those with more care than the ones that die in an inbox.

Fourth is hiring, in both directions. If you're evaluating candidates and you start running their materials through detectors, understand what you're actually measuring, which is whether a model touched the text, not whether the person can do the job. Detectors will hand you a number and you will be tempted to treat it as a verdict. Do not. And on the other side, if your own outbound recruiting materials carry marks, nobody cares, so relax.

Here's a wrinkle worth understanding before you overcorrect.

Text watermarks are fragile in a specific way. Anthropic has been reasonably honest about this. A file's marking can be stripped by format conversion, re-saving, screenshots, or similar handling. Text watermarks survive copy and paste, which is the whole point, but they degrade under paraphrasing, running the text through a second model, or simple retyping. The signal is statistical, which means it needs enough words to be detectable at all. A two sentence email probably carries nothing readable.

The EU's own regulation acknowledges this. Article 50 requires marking that is effective, interoperable, robust and reliable, and then qualifies the whole thing with the phrase "as far as this is technically feasible." That qualifier is doing enormous work.

There's also a detection gap. As of this writing, Anthropic has said it will share details on detection mechanisms in forthcoming technical documentation, but has not given a date. Until that lands, no outside party can independently verify whether a given passage carries a Claude mark. Which means for right now, the practical risk is lower than the headlines suggest, and the third party detectors people are already using are measuring something else entirely.

I'm not telling you this so you'll relax and ignore it. I'm telling you because the correct response to a slow moving change is a policy, not a panic. You have time. Use it to be deliberate instead of reactive.

Now for the part of the audit that always turns something up.

Machine generated text has crept into places you stopped thinking about, usually because it arrived through automation rather than through you sitting down and prompting something. That's the stuff worth finding, because you're not making a decision about it. It's just happening.

Go look at your automations. If you run Make, open your scenario list and find every module that generates text rather than moving it. Summaries that get written into your CRM. Follow up emails that draft themselves. Meeting recaps that post into a channel. Lead scoring notes. Ticket responses. Each of those is producing text under your business's name, on a schedule, without a human reading it first.

That's not a criticism. Half of that automation is why you have your evenings back. But you should know it exists, and you should know which of those outputs ends up in front of a customer versus which ones stay internal.

Meeting notes are the sleeper here. If you run Fathom or something like it, you've got AI written summaries of every call you've had for months, and those summaries have a way of getting forwarded. Somebody asks what was agreed, and your recap gets pasted into an email chain that eventually reaches a lawyer. That recap was written by a machine, from a transcript, and it may contain a characterization you never actually said out loud. The watermark is the least of your concerns there. The accuracy is the concern. Go read three old ones and see whether they say what you remember the meeting saying.

Check your help desk macros. Check your proposal templates, because the boilerplate you generated eight months ago is still going out on every bid. Check anything scheduled. Scheduled content is invisible content.

Here's your deliverable for the week. One page. Not a manifesto.

Answer four questions in writing, and put the answer somewhere your team can find it.

What do we use AI to draft? Be specific and generous. List the categories. Marketing content, internal documentation, first drafts of proposals, meeting summaries, whatever it is. Being vague here helps nobody.

What do we never use AI to draft? This is the shorter and more important list. Mine has sworn statements, anything with a client's confidential data in the prompt, and any communication delivering bad news to a human being. That last one isn't a compliance rule. It's a decency rule. If somebody's losing their job or their contract, they deserve sentences a person chose.

What do we disclose, and to whom? You have three real options. Disclose nothing and stand behind your work as your work. Disclose on request and answer honestly if asked. Disclose proactively on specific document types. Any of the three is defensible. What is not defensible is not having decided, then improvising under pressure when a client asks you directly.

Who reviews before it ships? Name a person for each category. Not a process. A person. Processes don't catch things. People do.

That's it. One page. It'll take you twenty minutes and it will save you an ugly conversation eighteen months from now.

I've been thinking about why this story landed harder than most regulatory news, and I don't think it's the watermark.

I think it's that a lot of people have been quietly uncomfortable about how much of their output is machine assisted, and the watermark makes that discomfort concrete. It's a mirror. Nobody minds the mark on content they're proud of. The mark stings on the stuff they'd rather not examine too closely.

If that's landing anywhere near home, the fix isn't hiding the mark. The fix is changing the ratio of your work that you'd stand behind in a room. That's a craft problem, not a compliance problem, and no policy document solves it.

Which brings me to the actual opportunity buried in all this. When machine written text becomes detectable at scale, the market value of clearly human work goes up. Not because human writing is better in some abstract way, but because scarcity does what scarcity does. The specific story only you can tell, the number only you have, the opinion that could get you in trouble, the case study with the client's real name on it. None of that gets marked, because none of it can be generated. It has to be lived first.

That's been true the whole time. The watermark just makes it easier to see.

Block the time Monday morning. Here's the sequence.

Twenty minutes listing every document category that leaves your business and lands in front of someone evaluating you. Twenty minutes marking each one with what would actually happen if it came back flagged. Twenty minutes opening your automation platform and your meeting notes tool and finding the text that generates itself without you looking. Twenty minutes writing the one page position. Ten minutes telling your team it exists.

Next Monday, take the three items you flagged as real exposure and fix the process behind them. Not all of them. Three.

That's the whole audit. You'll be more prepared than roughly ninety five percent of the businesses in your category, which is not a compliment to your category, but it is an advantage. Take it.

One more thing before you go. If you take nothing else from this, take the meeting notes item. Go read three old AI written recaps of calls that mattered. I'd bet real money at least one of them says something you didn't say. That's a problem you can fix today and it has nothing to do with the EU.

See you tomorrow.

Jordan

The AI Newsroom is written by Jordan Hale. This issue contains affiliate links to tools I actually use. If you sign up through them I may earn a commission at no extra cost to you.