How AI watermarks are changing L&D

If your team makes training with AI, the files you ship are starting to carry a hidden note that says so.

Since 2 August 2026, the EU AI Act has asked companies that offer AI tools to mark what those tools create. The mark has to be one that software can read and detect.^[1] Anthropic, Google, Meta, Microsoft and OpenAI have all signed up to do it.^[2]

We make one of those tools. So I've spent the last few weeks reading the law, the European Commission's guidance and the tech behind it. Here's what I think it means for you, from trust and compliance to certificates and the vendors you pick.

What an AI watermark is

There are two kinds of marks, and they work best together.

The first is a signed label inside the file. The open standard for it is C2PA, also called Content Credentials. The coalition behind it compares it to a nutrition label for digital content.^[3] It records what made the file and when. Its steering committee includes Adobe, Amazon, the BBC, Google, Meta, Microsoft, OpenAI and Sony.^[3]

The second is an invisible watermark woven into the picture, sound or text itself. Google's SynthID is one example. It marks images, audio, text and video made with Google's AI. Google says its image and video marks are built to survive cropping, filters and compression.^[4]

You need both because labels fall off. Anthropic lists format changes, re-saving and screenshots as ways a file loses its label.^[5] A watermark stays in the content itself. That's why the EU's Code of Practice asks for both: digitally signed metadata and an imperceptible watermark.^[6]

What the law asks, and of whom

Article 50(2) of the AI Act covers companies that offer AI systems making audio, images, video or text. Their outputs must be "marked in a machine-readable format and detectable as artificially generated or manipulated".^[1]

The rule has applied since 2 August 2026. Tools already on sale before that date have until 2 December 2026.^[7]

In June 2026 the Commission published a voluntary Code of Practice on how to do it.^[6] By the end of July, about 190 organisations had signed. Signers of the section for AI providers include Anthropic, Google, Meta, Microsoft, OpenAI and Synthesia. Degreed, the learning platform, signed it too.^[2]

Here's some of what providers who sign the Code commit to.^[6]

• A signed label and an invisible watermark. Text gets a watermark once it's longer than 200 tokens.
• A free way to check content, open to the public whenever the public sees that content.
• A line in their terms that bans removing or tampering with the marks.
• Marks that survive everyday handling, like compression.

For video, the Commission's guidelines say both the picture and the sound must be marked.^[8] Fines for breaking Article 50 go up to €15 million or 3% of worldwide turnover. For small companies, it's whichever is lower.^[9]

California is moving the same way. On 30 September 2026, the governor signed SB 1000. It drops the old million-user threshold from the state's AI Transparency Act. Now any company whose AI tool is open to Californians must hide a disclosure in the pictures, video and audio it makes.^[10]

What's yours to do

Most of this is your vendor's job. The marking duty sits with whoever provides the AI tool.^[1]

Your duty, as the team using the tool, is narrower. If a video shows a deep fake, you must say it's made with AI.^[1] One of the Commission's own examples is close to home. A realistic synthetic avatar of a company CEO, congratulating employees on the year, counts as a deep fake.^[8]

Text gets a lighter rule. It covers AI text published to inform the public on matters of public interest. The Commission says pages on an internal corporate network don't count as published.^[8] So the policy summary on your intranet needs no label under this rule, while a realistic avatar video might. Check your own cases with your legal team.

Your learners will be able to check

Once marks are common, checking takes seconds. Google's SynthID Detector lets anyone check an image, video or audio file for marks from Google and partners like OpenAI and NVIDIA.^[4] Anthropic has a free checker for files made with Claude.^[5]

That changes how people notice AI in training. Today a learner has to guess. With Gemini, Search or Chrome, they can already ask Google whether a file was made with its AI.^[4]

Oliver Schilke and Martin Reimann ran 13 experiments with 4,093 people on how AI disclosure affects trust.^[11] Again and again, people who said they used AI were trusted less. Students trusted a professor less when the professor said AI helped with grading.

One of their findings matters most here. The drop in trust was bigger when someone else exposed the AI use. Saying it yourself costs less than being found out.^[11] Watermarks make being found out much more likely.

So my advice is simple. Assume every AI-made asset can be traced, and tell your learners first. Say what AI made, what a person checked, and who signed it off. Then make sure that last part is true.

Quizzes and certificates

Watermarks won't catch a learner who uses AI to answer a quiz. Marks in text are the weakest kind.

Google says SynthID's text detection drops sharply when text is thoroughly rewritten or translated.^[12] In 2024, OpenAI said its own text watermark was easy to get around. Translating the text or rewording it with another model was enough.^[13] Anthropic says plainly that a missing mark doesn't mean AI wasn't used.^[5]

Hanlin Zhang, Boaz Barak and four colleagues went further in a paper at ICML 2024. They proved that, under natural assumptions, a determined attacker can erase a watermark without spoiling the content. Their attack stripped three published text watermarks with only minor loss in quality.^[14]

So don't build your assessment rules on a detector. Put your effort where you have control. That means practice inside the lesson, and a final quiz with a pass mark and a set number of tries.

Certificates are the other half of the same idea. A watermark tells you where a file came from. A certificate should tell you, just as reliably, that a person really passed.

A PDF on its own proves little, because anyone can edit a PDF. In May 2025 the W3C made Verifiable Credentials 2.0 a web standard. It's a way for an issuer to make claims that others can check. The W3C's own examples are "a driver's license or an education certificate".^[15] 1EdTech's Open Badges 3.0 follows it, so a badge can be checked with cryptography.^[16]

I expect auditors and hiring managers to want the same from training certificates. That means a way to check each one with whoever issued it.

What to ask your vendors

If you buy AI tools for training, add these to your checklist.

• Ask whether their videos, voices and pictures carry a signed label and an invisible watermark.
• Ask whether they've signed the provider section of the EU Code of Practice. The Commission publishes the list.^[2]
• Ask for the free tool anyone can use to check their files. Signers of the Code commit to offering one.^[6]
• Ask whose marks they rely on. The Commission lets a vendor rely on its AI model provider's marks, but the vendor still has to show it complies.^[8]
• Ask whether their presenter counts as a deep fake. If it does, you're the one who has to label it.
• Ask how a learner's certificate can be checked, and by whom.

Where Veurone Studio fits

Here's where we stand today.

Our presenters are characters, not synthetic people. The Commission's guidelines list AI voices for animated characters, where nobody is misled about who's speaking, among things that aren't deep fakes.^[8] In our reading, a lesson presented by one of our characters isn't a deep fake, so it needs no deep fake label. I wrote more about that in our post on branded characters.

Every course can end with a certificate in your brand. Learners get one when they pass, and anyone can check it's real with its code. You can set an expiry, withdraw one when you need to, and learners can add it to LinkedIn.

Our AI agent also says it's an AI at the start of every new chat. It asks you to check what it makes before you share it.

AI watermarks won't decide whether your training is good. They'll make it easier for everyone to see how it was made. I think that's healthy for L&D, as long as you're the one who tells people first.

If you'd like to try it, sign up for free and build a short course with a final quiz.