AI Text Is Being Watermarked. Your CV Pile Won't Notice.

From August, the text coming out of Claude carries an invisible watermark. OpenAI, Google, Meta, Microsoft and Mistral have signed up to do the same.

If you screen CVs for a living, that probably sounds like the cavalry arriving.

It isn't. And the gap between what watermarking does and what employers will assume it does, is where the expensive mistakes are going to happen.

I build AI screening tools, so I've spent a fair while inside this problem. Here's the honest version.

What actually got announced

Anthropic published the details on 14 August 2026. Claude models released on or after 2 August now embed a watermark in the text they generate.

This isn't a hidden character or a tag bolted onto the end. Nothing is added to the text at all.

The technique comes from Google DeepMind's SynthID-Text work, published in Nature in 2024. When a model writes, it constantly picks between words that would work equally well. Overcast or grey. Began or started.

Those choices are normally settled by a random number. Watermarking swaps the source of that randomness for a secret key, so the pattern of choices across a long passage can later be checked against the key.

The output reads identically. It costs no more, runs no slower, and carries nothing that identifies you, your company or your chat.

The reason all of this is happening at once is the EU AI Act. Article 50 obligations kicked in on 2 August 2026, and roughly 190 companies signed the Code of Practice on Transparency of AI-Generated Content in July. Anthropic is applying it worldwide because there's no reliable way to apply it only inside Europe.

So this is not one vendor's feature. It's the beginning of an industry-wide provenance layer.

Why a CV is the worst possible document to watermark

Now the part that matters for hiring. A CV is close to a worst-case document for this technique, for four separate reasons that stack on top of each other.

They're too short. Watermark detection is statistical. It needs a decent run of word choices before the pattern rises above noise. Anthropic says plainly that detection "doesn't work well on small samples." A CV is a page of fragments and bullet points, not an essay.

They're too factual. The watermark can only live in choices where either option is equally good. Job titles, employer names, dates, systems, numbers: none of those have a spare word to play with. Anthropic's own example is that you can't watermark the title of Newton's Principia, because only one word is correct. A CV is mostly that kind of text.

They're edited, not written. This is the big one. Most candidates don't ask an AI to invent a career. They paste in the CV they already have and ask for a tidy-up. Anthropic addresses this directly: when Claude proofreads someone's writing, "nearly all the words are the person's," so there's very little for a watermark to attach to.

A rewrite clears it. Anthropic concedes that a full rewrite removes the watermark entirely. Candidates retype content into their own template all the time without any intent to hide anything.

The trap: it proves processing, not authorship

Suppose you have a detection system, you run it across your pile, and one CV comes back positive.

What have you actually learned?

Anthropic answers this in one sentence: a watermark "cannot distinguish 'Claude wrote this' from 'Claude heavily edited this.'"

Sit with what that means in a hiring context.

The candidate who wrote every word of their own CV, then asked an AI to fix the grammar because English is their second language, may well carry a mark.

The candidate who had a model invent three achievements that never happened, then pasted the result into a Word template and reworded a few lines, probably won't.

You'd be screening on tool choice, not honesty. Worse, you'd be screening in a direction that quietly disadvantages careful, less confident and non-native-English applicants, who are exactly the people most likely to run their writing through a checker before sending it.

For Australian employers there's a compliance edge to this too. The privacy rules covering automated decision-making in hiring take effect on 10 December 2026. Rejecting candidates on a signal that the vendor itself describes as inconclusive is a difficult thing to defend if anyone asks you to explain the decision.

Will it make CVs any better?

This was the question I found most interesting, and I think the answer is no, for a reason that has nothing to do with the technology.

Watermarking changes no incentive for the candidate. It doesn't make a CV a more truthful account of what someone can do, and it doesn't make the applicant think harder about the role. It's a label on the manufacturing process, applied after the fact.

The CV was already a weak signal before any of this. It's a document written by the candidate, about the candidate, to be judged by a stranger, with no verification step anywhere in the chain. AI made it cheaper to polish. Watermarking now records that polishing happened, which is a different thing from making the underlying claims true.

There's one plausible way this improves things, and it's cultural rather than technical.

If provenance marking becomes universal and unremarkable, using AI to help write an application stops being a secret worth keeping. Right now candidates hide it, which is what drives the arms race between optimisers and screeners. Make disclosure boring and the incentive to conceal drops away.

We're a long way from that. But it's the version of this future where the CV pile gets calmer, and it doesn't arrive through detection.

Where watermarking genuinely helps

None of this makes the technology useless. It's aimed at a different problem than yours.

It works at document scale, where there's enough text for the statistics to bite: reports, articles, submissions, long-form policy documents. It's meaningful for provenance in journalism and academia. The C2PA file credentials are a real step for images.

There's a decent argument it will do more in your business for inbound content than for hiring. Supplier proposals, consultant reports, marketing copy you've commissioned. Longer documents, more generated text, higher stakes if it's slop.

Google is also working with Apple, OpenAI, ElevenLabs, Nvidia and others toward interoperable marking, so the multiple-keys problem should soften over time.

For a two-page CV, though, the physics don't change. Short, factual, edited rather than authored.

What to do about your CV pile instead

The answer hasn't moved, and it's a bit boring, which is probably why it keeps getting skipped in favour of a detection tool.

Stop trying to work out how a CV was produced, and start measuring things a candidate can't outsource to a language model.

Ask for specifics a real practitioner can answer and a polished document can't. Who else was on that project, and which part was yours? What was the number before and after, and over what period? A candidate who did the work answers in seconds.

Define what good actually looks like in the role before you advertise, drawn from the people already succeeding in it, so every applicant is measured against the same standard rather than against how well they write.

Use structured interviews with the same questions for every candidate, and a work sample where the role allows it. Behaviour under a realistic task is the one thing no CV, human-written or otherwise, can fake for you.

There's a useful footnote in Anthropic's own explainer here. They point out that detection software works differently, by looking for the tells in AI phrasing, and they name two: the "this isn't X, it's Y" construction, and a strange fondness for the word "quietly."

Which tells you something about where this is heading. When the model makers are publishing the giveaways, the giveaways stop being reliable, because everyone writing prompts reads the same article.

Where GrowMyTeam fits

We screen against a behavioural benchmark built from your own high performers, then generate interview and reference questions from each candidate's actual history rather than from how their CV reads.

That approach was chosen before any of this was announced, for a simple reason. We assumed from the start that the document would keep getting more polished and less informative, so we built to measure the things a document can't carry.

Watermarking doesn't change our roadmap. It's a useful transparency measure aimed at a problem that isn't the one sitting in your inbox.

Try the behavioural profiling on one of your own roles, no cost, about ten minutes.