AI Content Is Now Watermarked. What It Means for Your Brand

From 2 August 2026, text produced by Anthropic's Claude models carries an invisible watermark, and image files the models output carry signed provenance data. The change applies across every surface those models run on, worldwide. Other providers are covered by the same regulation, so it is reasonable to expect the rest of the market to follow.

If your business publishes content that has been near an AI model, and most now do, this changes what can be proven about that content later. It does not change your search rankings. It does change what you should have written down before a client, a procurement team or a journalist asks you a direct question.

What has actually changed?

Two separate things, and they work differently.

The first is a statistical watermark embedded in generated text. You cannot see it, and it does not affect how the writing reads. It survives copying and pasting, which means it travels into your CMS, your email platform and your slide deck along with the words. Anthropic says it may also survive a degree of editing.

The second is signed provenance metadata attached to image and vector files, following the C2PA open standard. This is closer to a tamper-evident seal. It records that the file came from an AI model and shows whether the file has been altered since.

The trigger was the EU AI Act transparency code, which came into force on 2 August 2026 and requires AI-generated or AI-edited content to be marked in a machine-readable way. The obligation sits with model providers first. It also reaches the organisations deploying those models, which includes most marketing teams.

One practical caveat. Models released from 2 August mark their output at launch, and older models are being retrofitted, so coverage is still filling in. Assume anything produced from now on is marked rather than trying to work out which model touched what.

Does a watermark mean the content was written by AI?

No, and this is the part most coverage has skipped.

A mark shows that the text passed through the model. It does not show who wrote it. If someone on your team writes a page from scratch and then asks an AI tool to proofread it, tighten it or translate it, the finished text can come back marked. Anthropic acknowledges this directly.

So the mark is evidence of involvement, not evidence of authorship. Anthropic is also explicit that detection is not fully conclusive, and that the absence of a mark does not prove a human wrote something.

That cuts both ways for brands. Work your team wrote from scratch can be flagged. Work produced almost entirely by a model can come out clean after a heavy rewrite. Anyone treating a detector result as proof of anything is overreading it.

There is also no public detector yet. Anthropic has said tooling for third parties is coming, without saying when.

Will watermarked content hurt your rankings?

There is no evidence that it will, and no reason in the current guidance to think it should.

Google's position has been consistent for years. It rewards content that is useful, accurate and demonstrably experienced, and it acts against content produced at scale to manipulate rankings, whoever or whatever produced it. Production method is not the test. Quality and usefulness are.

Nothing in the watermarking announcements changes that, and search engines have made no statement connecting these marks to ranking systems. Treating this as an SEO problem points you at the wrong fix. The real exposure is commercial.

Who is actually going to check?

Think about who has a reason to look, because that tells you where the risk sits.

• Clients with an AI clause in the contract. Plenty of agreements signed in the last two years say content will be human-written. Until now that was unverifiable. Soon it will not be.
• Procurement and tender teams. Public sector and enterprise tenders increasingly ask for an AI declaration. A declaration you cannot stand behind is a problem.
• Journalists and competitors. A detector plus a well-known brand's blog is an easy story.
• Regulators, in specific sectors. Financial services, health and anything consumer-protection adjacent will get there first.

None of those are search engines, and all of them are people you have a relationship with. That is why the answer is a policy rather than a technical countermeasure. It is also worth asking about when you choose an agency, because the honest ones will already have thought about it.

What should brands do now?

Four things, roughly in order of urgency.

Check what you have already agreed to

Read your supplier and client contracts for AI clauses. A commitment to human-written content, made when nobody could check, is now a commitment someone can test. If you cannot meet it, renegotiate it before it is tested rather than after.

Write down how you actually use AI

Not a public relations statement. A working document that says what AI is used for in your business, what stays with a person, who signs off, and what you tell a client who asks. Most teams are already using these tools daily without an agreed position, which means the answer to a direct question gets improvised by whoever is asked first.

Keep your evidence

Briefs, outlines, drafts, version history, editorial notes. If human-written work is ever wrongly flagged, process evidence is what answers it. This costs nothing if you start now and is impossible to reconstruct later.

Fix your public claims

Go through your website, proposals and pitch decks for language like "written by our team of experts" that may no longer be strictly true. Replace it with something accurate. Human-led and AI-assisted, with named editorial accountability, is a stronger claim than a purity claim you cannot evidence.

Should you try to strip the watermarks?

Our view is no, and not for squeamish reasons.

It is technically possible to degrade these marks. Heavy rewriting weakens a text watermark, and re-saving or converting a file can drop the provenance data. But look at the trade. The upside is avoiding an uncomfortable conversation. The downside is a client discovering later that you engineered a workaround, which turns a conversation about tooling into a conversation about honesty. Published archives can also be rescanned as detection improves, so anything you ship today stays checkable.

There is a legitimate version of the same instinct, and it is simply better work. If you want less machine signature in your output, put less machine in your output. Use AI for research, structure, data analysis and quality checks, then have a person write the words. That produces better content anyway, and the smaller footprint is a by-product rather than the goal.

For anything covered by a no-AI commitment, keep it out of the tools entirely. That is cleaner than trying to clean up afterwards.

Where is this heading?

Within a year we expect detection tools to be widely available, disclosure questions to appear as standard in briefs and tenders, and "how do you use AI" to become a normal question in a pitch rather than an awkward one.

The brands that come out of this well will be the ones with an answer ready. Not a defensive answer, and not a claim of purity, but a clear account of where the machine helps, where the people decide, and who is accountable for what gets published. That is a stronger position than either extreme.

Frequently asked questions

Can I tell if content is AI-generated right now?

Not reliably. There is no public detector for these watermarks yet, and the existing commercial AI detectors work by guessing from writing style, which produces frequent false positives. Treat their output as a prompt to ask a question, not as proof.

Does the watermark contain my company's details?

Nothing published so far suggests the mark identifies a user or an account. It marks the content as having passed through the model. Anthropic has said fuller technical documentation is coming.

Does this apply to AI images too?

Yes, but by a different mechanism. Image and vector files get signed C2PA provenance metadata rather than a statistical watermark. That data can be stripped by format conversion or a screenshot, which is why it is described as tamper-evident rather than permanent.

Do we have to disclose that we used AI?

It depends on your sector, your jurisdiction and your contracts. Under the EU AI Act there are transparency duties for deployers in defined circumstances, and individual regulators are adding their own. Outside those, disclosure is a commercial and reputational decision rather than a legal one. Either way, the decision is better made deliberately than discovered under pressure.

Will this affect our search rankings?

There is no indication that it will. Search engines assess content on usefulness and quality rather than on how it was produced, and none has connected these marks to their ranking systems.