Claude just started watermarking everything. So is that good news, or are we all in trouble?

If you are like us, you have had that slightly guilty moment when a piece of work came back faster than it should have. You thought a model had probably touched it, then you sent it anyway, because it was fine and nobody was going to check. As of 11 August, somebody might.

Anthropic, the company behind Claude, said its new models will now hide a watermark in the text they write. You cannot see it, and neither can your client, but a detector can. The mark is designed to survive a copy-paste and a light edit. A proper rewrite or a translation can still knock it out. They are applying it worldwide, not only in Europe, because the EU asked labs to mark AI output as AI and they do not yet have a clean way to stamp EU traffic only. Images and graphics are treated the same way: a PNG, JPG or SVG from Claude gets a signed note in the file metadata saying Claude was involved, which is the same idea cameras already use when they certify a photograph.

So is that good?

It is good if you actually want to be able to tell the truth. Clients, counterparties and, eventually, regulators will ask whether a machine helped write something. If the answer is Claude, you may one day have a way to show it. That is more honest than what we had last month, which was mostly vibes. It is not good if your shop has been treating AI like a silent intern. The stamp is not a gold star. It does not say legal signed this, and it does not say that no client file went into the prompt. It only says a machine was in the room. If the words are wrong, you still sent them. We already know how that story ends: if your AI says it, you own it.

Here is the part that actually matters for a normal business in Hong Kong or Australia. Most of you are not a Claude shop. You are a mixed shop. One person uses Claude, another uses ChatGPT, and someone else has a free tab open on a phone that never asked IT for permission. One of those tools now leaves a fingerprint. The others leave a shrug. That is not a small gap.

If a counterparty runs a detector on a letter and gets a hit, they know Claude was involved. They still do not know who prompted it, or whether a matter file was pasted in. You have explaining to do, and the stamp does not do that explaining for you. If they run the same check and get nothing, that does not mean a human wrote it. It might mean someone used the unmarked chat on purpose, which is exactly what a rational staffer will do once they hear that Claude can be detected and ChatGPT cannot. Congratulations: the line that said "we use Claude, so we are covered" just trained people off Claude.

That is the implication. Watermarking is not bad, and it does not save you. Invisibility is over for one tool and still total for the rest, and staff will follow the invisible one when the work is sensitive. You do not need a Brussels office for that to land. You need one EU client, one file that gets forwarded, or one person who asks a simple question six weeks later.

If this were my company and not a blog, I would do three things this week. I would write down the tools, the real list, including the free ones. I would pick one rule for anything that leaves the building: a named human has to have read it, and the tool has to be on the list. That is three lines. You can argue about the rest later. And I would stop telling myself that an unmarked paste is the same as "we did not use AI." It just means you have no evidence either way.

If every tool you used was marked, this would be a good week, because you could tell a client what was machine-made. That is not this week. This week one vendor stamps, the free tabs do not, and most mid-size shops run both. Use Claude. Use ChatGPT. I do not care. Just do not pretend the new stamp is a policy, and do not pretend the missing stamp is cover.

If you want a grown-up conversation about what can leave the building and what never can, before someone else's detector does it for you, you know where we are.

3peat.ai