AI Watermarks: Is Your AI-Written Content Marked?

Anthropic has started embedding invisible, machine-readable watermarks in text and files produced by Claude models released after early August, to comply with the EU AI Act. Other major providers have committed to similar things. If your business drafts proposals, blog posts, or client emails with AI, the reasonable question is whether that output now carries a mark, and whether it matters. Short version: increasingly yes, and mostly it only matters if you were being vague about it.

What it actually is

A watermark is not a visible label and it is not a note buried in the file properties. In images, it is usually a pattern woven through the pixels that the eye cannot see. In text, it works by nudging word choices into a statistical pattern that a matching detector can recognise and a reader cannot.

The important part is when it happens. A watermark is added at the moment of generation by the company that made the model. That is a different thing from an AI detector, which looks at finished text afterwards and guesses. One is a signal deliberately placed; the other is an inference. Confusing the two causes most of the bad thinking in this area.

Coverage is patchy, so ask

Source of content	Likely watermarked?
Recent models from major providers	Increasingly yes, verify per vendor
Older model versions	Often not
Smaller vendors and niche tools	Usually not
Self-hosted open-weight models	Generally not

Text watermarking is also newer and more fragile than the image version. So the honest position for a business is that you cannot assume your output is marked, and you cannot assume it is not. If it matters to you, ask the vendor and ask where it is documented.

Why this mostly does not change your work

If your business drafts with AI and a person reviews, edits, and takes responsibility for the result, a watermark is irrelevant to you. That is a normal way to work, it is how most professional output has always been produced with tools of some kind, and nobody sensible objects to it.

The businesses that should feel a little uncomfortable are the ones whose arrangement depends on nobody being able to tell. Charging bespoke rates for generated work. Describing content as written by your team when a model wrote it and nobody read it carefully. Those were always reputational risks, and the ground under them is shifting from guesswork toward a signal someone deliberately put there.

Do not treat it as a test

This is the part worth getting right, because the temptation to use watermarks as a lie detector will be strong, particularly for anyone commissioning freelance work.

Watermarks can be weakened or destroyed by ordinary editing, paraphrasing, translation, or passing text through a second model. Which means absence proves nothing. A supplier whose work shows no watermark may have written it themselves, or may have edited it enough to break the signal. Accusing someone on that basis is a bad idea, and the mirror-image problem of detectors producing false positives has already cost real people real money. Treat a present watermark as evidence of provenance. Treat its absence as nothing at all.

The direction this is heading

Watermarking is arriving because regulation asked for it, chiefly the EU AI Act and its requirement to machine-readably mark AI-generated content, which we covered in the EU labelling deadline. Canada has no equivalent rule today, and the current federal work is a consultation rather than a mandate. But provenance infrastructure does not stay regional for long, because vendors build one product and ship it everywhere.

The practical read is that content provenance is becoming a normal property of digital work rather than a special investigation. That is broadly good for businesses doing honest work, and it slightly raises the value of being visibly human where being human is the point, which is the argument in the authenticity advantage in a flooded feed.

Three things to actually do

Ask each AI vendor you rely on whether their output is watermarked, and get the answer in writing or find their documentation. Check that your public descriptions of how content gets made would survive someone verifying them, which is a lower bar than it sounds and worth confirming. And if you sell into the EU, read the marking requirements properly rather than relying on your vendor to have handled it, because the obligation sits with you.

That is roughly an hour of work. For most Canadian businesses, the correct response to this news is a small amount of diligence and no change to how you operate. The exception is if reading this made you slightly nervous about how you describe your content, in which case the fix is saying so plainly rather than hoping the signal is fragile.

Frequently Asked Questions

What is an AI watermark?

It is a signal deliberately embedded in AI output so the content can later be identified as machine-generated. In images it is usually a pattern in the pixels invisible to the eye. In text it works by nudging word choices in a statistical pattern that a matching detector can recognise but a reader cannot. It is not a visible label and it is not metadata you can strip by copying and pasting. The important distinction is that it is added at generation time by the vendor, rather than guessed at afterwards by a detector.

Is my AI-written content watermarked?

Increasingly likely, depending on the tool and when the content was produced. Anthropic has begun embedding machine-readable watermarks in output from Claude models released after early August 2026 to meet EU AI Act requirements, and other major providers have signed up to similar commitments. Coverage is uneven, though. Text watermarking is newer and less robust than image watermarking, older model versions may not carry it, and smaller vendors and self-hosted open-weight models generally do not. Assume nothing either way without asking your vendor directly.

Should this change how we use AI?

Not really, if you were already being straight about it. If your business drafts with AI and a human reviews and takes responsibility for the result, a watermark changes nothing about that arrangement. The people who should be uncomfortable are the ones relying on nobody being able to tell. What it does change is the risk profile of quietly passing off generated work as bespoke human effort, because the detection question is shifting from guesswork toward a verifiable signal put there deliberately.

Can watermarks be removed?

Sometimes, and that is exactly why nobody should treat them as proof. Heavy editing, paraphrasing, translation, or running text through another model can weaken or destroy the signal, and text watermarks are more fragile than image ones. Which cuts both ways. Absence of a watermark does not mean content is human, and a business should be very careful about drawing conclusions from a negative result. Treat these as evidence of provenance when present, not as a test that produces a reliable answer either way.

What should a Canadian business do about it?

Three things, none of them urgent. Ask your AI vendors whether their output is watermarked and where that is documented, so you know rather than guess. Make sure your disclosure practice would survive someone verifying it, meaning do not describe content as written by your team when it was drafted by a model. And if you publish into the EU or serve EU customers, look at the AI Act marking requirements properly, because those are actual obligations rather than vendor policy.