What does the Claudish watermark mean for businesses?

Anthropic recently announced that Claude now adds an invisible watermark to the text it generates. The watermark can withstand light editing so it may end up in a board paper, marketing copy or client proposal without anyone noticing it on the page (it's not something we can see with the naked human eye...). Quite frankly I hope it’s solid wake up call for the lazy Claude copy paste increasingly seen over email and social media that needs a clean up.

This change follows Article 50 of the EU AI Act, which came into force on 2 August 2026 and impacts AI model providers, causing them to mark synthetic content in a machine-readable format. Anthropic now applies the watermark out of its Claude models.

Interestingly a watermark indicates that an AI system probably supplied some of the wording. It doesn't rate the work, test whether the writer understood it, or decide whether a company policy was breached. It can tell us something about where the work came from (its provenance), but nothing about whether the work is any good or will be impactful (the resulting outcome or impact).

I see that distinction as important because businesses could easily use this new signal to reach the wrong conclusion. A watermarked document may be treated as evidence that someone was lazy, concealed their AI use or contributed little of their own thinking to the work themselves.

But none of those things can be read from the watermark.

It records provenance and part of the execution. It doesn't judge the outcome as to whether the work is accurate, useful, well reasoned or safe to make decisions off the back of.

The question businesses should in-fact be asking is therefore bigger than how to detect AI-produced work.

It is how to distinguish responsible AI assistance from poor work, and how to recognise the advantage created when people use these tools well.

Unpacking these ideas starts with understanding what the watermark measures and how it works.

The watermark isn't a hidden label pasted into the file that we as humans can see and scratch at. It's created in the way that Claude chooses the words it regurgitates, which can only be unpacked through statistical patterns.

Consider this sentence:

Claude could fill the gap with words like "introduce", "adopt" or "roll out". All three would work in that sentence.

In a simplified version of watermarking (the exact 'system' is secret), the rule nudges Claude towards one of those equally viable options of words it could include. The preferred choice of word sequencing changes with the surrounding text (its all probabilistic maths), so there's no fixed list of suspicious words (basically Claude isn't trying to 'trip anyone up').

That small nudge happens repeatedly through a passage:

Thinking about useful comparisons, a weighted coin came to mind.

In the same way, no single word exposes the watermark. The pattern emerges across many ordinary word choices and across paragraphs. The reader sees normal text whereas the watermark detector would see an unusually high number of choices matching Anthropic's secret pattern.

Light editing may leave most of those choices in place, so changing the heading and cleaning up a few sentences still leaves hundreds of original words for the detector to count. Whereas a complete rewrite would replace the word sequence choices and remove the identifiable patterns, which Anthropic states directly in its documentation: "Light editing probably won't remove the watermark completely; a complete rewrite where every word is replaced will".

The result supports one conclusion: Claude probably generated a meaningful amount of the finalised text.

It can't reconstruct how the document was produced because the watermark detector never sees the input information to generate the output: the meeting that generated the idea, the spreadsheet used to check the numbers, or the rejected drafts discarded. It also can't see whether someone read the final document before approving it.

Let's bring this to life with two board papers to make the limitation obvious:

The detector may give both papers a similar result even though the first contains a week of human work and the second contains almost none. One manager has used AI to express and strengthen their thinking, whereas the other has used it to avoid thinking at all.

The watermark cannot distinguish between them, which is where we risk individuals jumping to conclusions about what the watermark means, and how they should treat the output they receive if a watermark is identified.

The reverse problem can also exist.

This is why I find this shift and expectation from the market so interesting, in that this is the central limitation. The watermark measures provenance, but it does not measure performance.

So what does that all mean?

Watermarked work becomes suspect to humans the instant it shows up. Whereas unmarked work is assumed to be human and, to many, therefore, better. Employees might learn that visible AI use carries risk, while invisible AI use does not, even at the expense of the quality of the output.

That would create exactly the wrong incentive.

**People would not stop using AI. They would rewrite the output until the signal disappeared, move between tools or find other ways to conceal how the work was produced.**The business would reward the appearance of doing everything manually rather than the responsible use of AI, all wasting resources with people attempting to 'clean up' their work to avoid the AI tarnish.

**It would also risk labelling people as lazy without examining what they actually did.**Using AI to avoid research, judgement and accountability is a problem. Using it to analyse more material, explore alternatives, improve an explanation or challenge an initial position may produce better work. The visible involvement of AI tells us which tool was used, not which of those behaviours occurred and the effort involved.

**The same distinction applies beyond text.**A watermark in an image cannot tell whether the person accepted the first generation or worked through dozens of concepts to find the right one. A mark in a video cannot tell whether the claims were checked, the story was considered or the finished piece achieved its purpose. Provenance travels across media, whereas what is more and most important is the judgement applied throughout, which sits outside the mark.

The advantage will not come from producing work with no detectable trace of AI. That will become a feature of how people work moving forward, rather than a bug. It will come from using AI to improve the work while retaining enough human judgement and accountability to stand behind the result, with clear explainable proof from the author that they understand the intricacies and nuances of the output.

A team that uses AI well can examine more information, test more options, produce earlier drafts faster and spend more time refining the final decision. That team should be able to outperform one that prohibits the tools or drives AI use underground, all for the sake of combatting a culture that breeds negative viewpoints about using AI to produce work.

But that advantage depends on trust and can't be bought.

Customers, managers and colleagues need to know that faster production has not replaced responsibility. Businesses need a way to confidently say "AI contributed to this work, a person checked it, and we are accountable for the outcome". Taking that position as an organisation, and being explicit about it will create more comfort than doubt. The more individuals and businesses attempt to masquerade their AI usage to appear 100% human in the outputs produced, the greater the scepticism.

I have never hidden the fact that AI plays a significant role in helping me create content, challenge my thinking and automate parts of my work. I also recognise that people have valid reasons for choosing when and where they use it.

But in a business context, I find it increasingly difficult to imagine organisations remaining competitive while insisting that every part of the work must be produced without AI assistance. The opportunity is not to remove people from the process but to use these tools while remaining clear about where human judgement and accountability sit in the process.

In that sense, the watermark could become useful. It provides one piece of evidence about how something was produced. Combined with sources, working notes, version history and named approval, it can contribute to a more transparent and trustworthy record of production.

The watermark itself is not the advantage, but the advantage is an organisation capable of using AI openly and well in alignment with their brand could well be, when marketed right.

Many AI policies still concentrate on whether employees are allowed to use a tool and whether the resulting content must be disclosed. Watermarking makes those questions easier to police, but it also exposes how incomplete they are with very little critical thought going into the purpose of such a policy.

A policy built only around AI producing work treats the presence of AI as the most important event. A useful policy also considers what the person was responsible for before, during and after using it.

That means businesses need to engage their people on five practical questions:

The answers will vary with the consequences of the work. An internal meeting summary does not need the same controls as financial advice, a board recommendation or a public claim. The standard of review and transparency should change with the risk profile of the activity, regardless of whether a watermark is present or not. The watermark just brings this conversation to the forefront with proper risk to perceptions of an organisation if not managed appropriately.

Businesses also need a fair response when a watermark detector flags something. The result should start a conversation, not end one.

The person should be able to explain how AI was used, show the sources and working material behind the output, and identify what they checked, challenged or changed. A detector score alone should not become proof of laziness, deception or poor performance.

There is also a broader lesson here for teams using AI to produce a first draft. Most focus on capturing the finished output but lose the valuable material generated along the way: the judgement calls, rejected options, prompts, corrections, source material and back-and-forth that shaped the result.

In almost every business we work with, little attention is given to preserving this context. Yet this “data exhaust” contains important evidence of how the organisation makes decisions, applies its expertise and turns information into action, accumulating into the business’s know-how.

As organisations begin building an evidence bank around AI-assisted work, the benefit will extend beyond proving how one document was produced.

That accumulated context can be reused to help AI understand the organisation’s standards, language, decisions and ways of working. Each project then contributes to the next, creating a compounding advantage: AI that produces work increasingly aligned with how the organisation’s best people think and operate.

The arrival of reliable watermarks will make the involvement of AI more visible. But it will not resolve the harder questions businesses have to answer.

Those questions concern the outcome rather than provenance of the output, and no hidden pattern of words from Claude or another model provider can answer them.

Businesses should use provenance data for what it is: information about how something was produced. They should not confuse it with a measure of effort, intelligence or value. The watermark can identify AI involvement, but it cannot identify laziness, judgement or quality.