AI Watermarks and Content Credentials: Do AI Labels Really Work?

If you’ve published something which passed through an AI tool recently, it could carry a mark you can’t see and didn’t choose to add.

That’s increasingly likely as the industry responds to new AI transparency requirements. 

Transparency rules under the EU AI Act came into application on 2 August 2026, and Google, Meta, Microsoft, Anthropic and OpenAI have all committed to greater transparency around AI-generated content. 

On the surface, that sounds straightforward: AI makes something, it gets an AI-generated content label, and everyone knows where they stand. 

Except that’s not quite how it works.

Some marks are visible, some sit quietly in metadata and others need specialist tools to detect them. More importantly, detecting that AI touched something isn’t the same as proving AI created it.

So what do all these new AI labels, watermarks and credentials actually tell your customers?

In the spirit of transparency, the research and editing of this article was supported by AI.

What Types of AI Labels Exist?

It helps to understand the main ways AI-generated content can be marked.

They’re not all doing the same job. Some labels show where content came from, some hide a detectable signal inside it and others simply tell the person looking at it that AI was involved.

What Are Content Credentials?

The Coalition for Content Provenance and Authenticity, or C2PA, uses Content Credentials to create a tamper-evident record containing information about how a digital file was created and edited. 

The standard is backed by major technology and media companies, including Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI and Sony. 

Rather than simply asking “Was this made with AI?”, Content Credentials can help answer a broader question: “What happened to this content before it reached me?”

This might include information about how it was created, what tools were used and whether it was edited afterwards.

What is AI Watermarking?

An AI watermark doesn’t necessarily look like the traditional watermark you’d see stamped across an image. Instead, it can be embedded into the content in a way that’s invisible to the person viewing it.

For text, that can mean subtly influencing the word choices a model makes. You won’t see it when you read the text, but a compatible detection system may be able to identify it later.

That means detection isn’t always open to everyone. If verification requires a specific key or API, the company that created the watermark still controls who can reliably check for it.

OpenAI explored text watermarking before shelving its plan in 2024, after research suggested that almost 30% of ChatGPT users would use the product less if watermarking was introduced. Not a small number. It will be interesting to see how Anthropic – which recently announced the introduction of watermarking on Claude – will perform over the coming months.

Transparency probably sounds simple to providers: until it starts affecting how people use the product.

What’s the Difference Between the Two?

The easiest way to think about them is this:

• Content Credentials record provenance. They aim to show the history of a piece of content and how it was created or changed.
• AI watermarks embed a signal. They can help indicate that a particular AI system generated or processed the content.
• Neither automatically tells you who authored the idea. That distinction becomes especially important with text.

Content Credentials are closer to a digital history. AI watermarks are closer to a detectable fingerprint.

Both can support content authenticity, but neither gives the whole story on its own.

What Other Labels Exist?

Platforms may add:

• Visible AI labels such as “Made with AI” or “AI-generated”.
• Creator disclosures where the person uploading content declares that AI was involved.
• Automatic platform labels triggered when a service detects AI-generated media or recognises embedded provenance data and applies its own label.
• Metadata that stores information about AI use within a file without necessarily displaying it prominently to the user.

These labels are arguably the easiest forms of AI disclosure for customers to understand because they don’t require specialist tools. But they also rely on either the platform detecting AI correctly or the person publishing the content being honest about how it was made.

That’s one reason we’re increasingly seeing several methods used together.

How Are Companies Marking AI Content?

There’s no universal approach to AI labelling or content authenticity, and regulation doesn’t currently necessitate a specific method. Different companies mark different types of content in different ways, with very different levels of visibility to the person actually consuming it.

And regulation isn’t the only reason they’re doing it.

As more AI-generated content appears online, AI companies also need ways to identify their own output and prevent future models being trained indiscriminately on content created by earlier models. Labelling can help platforms recognise what their own AI systems have produced.

So, the technologies below aren’t only about helping customers identify AI. They’re useful to the companies building the models too.

With that in mind, let’s look at how three of the biggest AI companies are approaching it.

1. Google and SynthID

Introduced in 2023, Google’s SynthID can watermark AI-generated images, audio, video and text. 

Importantly, it doesn’t necessarily add a visible AI label. The signal can sit inside the content itself.

By May 2026, Google reported watermarking more than 100 billion images and videos, plus 60,000 years of audio. 

Verification is also becoming more consumer-facing through Gemini, Search, Lens and Circle to Search, while businesses can access an AI Content Detection API through Google Cloud.

Google’s approach also extends to YouTube, where creators have been asked to disclose realistic AI-generated content since 2024. In 2026, YouTube made those disclosures more prominent and began automatically applying them in some circumstances.

So Google is effectively using several layers at once:

• SynthID as an embedded watermark
• Visible labels through platforms such as YouTube
• Detection tools through Gemini and Search
• APIs for businesses that need deeper verification

2. OpenAI Content Credentials and Verification

OpenAI takes a slightly different mix of approaches.

It uses C2PA Content Credentials to provide provenance information on generated media, and has partnered with Google to add SynthID to images. Its verification tools expanded to include audio and API access in July 2026.

Again, that means the information can exist at several levels: embedded provenance, watermarking and tools that allow someone to inspect the content afterwards.

But there’s an important distinction running through all of this.

A file carrying provenance information isn’t the same as a customer understanding that information.

If the data is there but your customer never sees it, how transparent are you actually being?

3. Anthropic and Text Watermarking

Anthropic introduced text watermarking for models released on 2 August 2026. It uses a version of SynthID-Text, the method Google DeepMind published in 2024.

But its effectiveness varies depending on what you’re creating.

• Short copy is difficult to mark reliably. There simply aren’t enough word choices in a strapline or product name to create a strong signal.
• Factual writing is marked sparsely. When there are fewer genuine choices about how something can be phrased, there’s fewer opportunities to add the pattern.
• Code is barely marked. Functional output often needs to be exact, so any marking is mainly found in comments.

Anthropic has promised a detection API, but it isn’t available yet. Older Claude models are due to be covered over the coming months.

That’s a useful snapshot of how quickly AI content labelling is developing: policy arrived first, the mechanism followed and the means for everyone else to check it is still catching up.

Don’t AI Detectors Solve This Problem Already?

This is where we need to separate two ideas which sound similar but work very differently.

Watermark detection looks for a signal deliberately put there. AI detection tries to determine whether AI was involved by analysing the finished content.

And that second approach can be wildly unpredictable.

In August 2026, Search Engine Journal’s Andy Betts tested an article he’d written himself using several leading AI detectors. The results ranged from 100% AI-generated to entirely human-written. Even articles published years before ChatGPT existed were flagged as containing AI-generated material.

Other research raises similar concerns. Stanford researchers found that detectors can disproportionately misclassify writing by non-native English speakers, while 2026 research found that relatively small human edits could allow genuinely AI-generated content to bypass detection.

So, an AI detector giving something a 90% score isn’t the same as finding a deliberate watermark or verified Content Credential.

An AI Watermark Doesn’t Mean AI Wrote It

Imagine you write an article yourself, then ask an AI tool like Grammarly to proofread it. Or perhaps you ask DeepL to translate, ChatGPT to summarise or Claude to convert the article into another format.

Depending on the system, the resulting text could end up carrying an AI label even though you wrote the original copy.

And that problem isn’t limited to one type of AI label:

• AI watermarks can indicate that a particular model generated or processed content. With Anthropic’s text watermarking, for example, a positive result can establish that Claude processed the text, but not that Claude authored the original ideas or copy.
• Content Credentials and metadata can record information about the tools used during a file’s creation or editing. That’s useful provenance, but the presence of an AI tool in that history doesn’t necessarily tell you how significant its contribution was.
• Visible AI labels can sometimes be applied by platforms based on disclosures, detected signals or provenance data. Again, the label tells the user that AI was involved, but may not explain whether that involvement was generative or making a relatively minor edit.

That distinction is crucial.

“AI was involved” and “AI created this” aren’t interchangeable statements.

Interestingly, the regulation itself recognises some of that grey area. Article 50 of the EU AI Act includes exemptions for systems used for standard editing that don’t substantially alter the meaning of the text. Published text that has undergone genuine human review or editorial control can also be treated differently.

So, we could end up in a situation where a technical signal indicates AI involvement, but the publisher has no legal obligation to label the content as AI-generated.

The technology and the legal responsibility won’t always tell the same story. And as AI becomes embedded in more everyday creative tools, understanding the difference between AI involvement and AI authorship is only going to become more important.

What Does This Mean for UK Businesses?

If you’re reading this from the UK, the obvious question is whether any of it applies to you. The answer is a bit more complicated than a simple yes or no.

The UK Doesn’t Have an Equivalent AI Labelling Law Yet

The UK has no direct equivalent to the EU AI Act’s general transparency framework. The ASA has clarified that the CAP and BCAP Codes contain no AI-specific rules and that there’s no blanket requirement to disclose AI use in UK advertising.

Existing rules still apply, particularly around misleading claims; AI-generated advertising which doesn’t accurately represent what you’re selling can be in breach. 

The UK’s approach is also still developing, so don’t assume today’s position will hold.

But the EU Act Could Still Apply to You

Being based in the UK doesn’t automatically put your business outside the EU AI Act. The legislation has extraterritorial reach in certain circumstances, including where AI output produced outside the EU is used within the Union – which could bring UK organisations into scope.

Businesses operating across the UK and Europe should establish whether the rules apply to their use of AI. Breaches of the Article 50 obligations carry penalties of up to €15 million or 3% of global annual turnover. Reason enough to find out.

Can Automatic AI Labelling Be Circumvented?

There’s another challenge with AI watermarking and other forms of automatic AI labelling: people are inevitably going to look for ways around it.

That shouldn’t be particularly surprising. As AI becomes part of more businesses’ everyday workflows, there’ll always be people who see disclosure requirements, watermarks or embedded metadata as a roadblock rather than a transparency measure.

And there are several ways these signals can become less reliable:

• Editing the output: Text watermarks rely on patterns across the words a model chooses. Change enough of those words and the original signal can weaken or disappear.
• Processing it through another AI tool: Asking another model to rewrite AI-generated copy could alter the original watermark, although you may simply replace one model’s output label with another’s.
• Changing the file: Content Credentials and metadata can behave differently when files are edited, converted, re-saved or moved between platforms.
• Deliberate circumvention: As watermarking becomes more widespread, we’re likely to see more tools designed specifically to interfere with or remove these signals.

That creates problems in both directions. Finding an AI watermark doesn’t necessarily prove AI authored something, while failing to find one doesn’t necessarily prove a human did.

It’s a problem we’ve looked at before. In Designing an AI-Free Logo, we asked a deceptively simple question: how do you prove something wasn’t made using AI? A “human-made” badge might make that claim, but without a reliable way to verify it, it ultimately depends on trust and governance.

Automatic AI labelling approaches the same problem from the other direction. Instead of asking creators to prove that AI wasn’t involved, platforms are attempting to leave evidence when it was. But if those signals can be altered, removed or misunderstood, neither approach gives us absolute proof of authorship.

At TH3, we don’t think the answer is to build your creative process around avoiding detection in the first place.

AI is genuinely useful for research, interrogating a subject and pressure-testing a structure. We use it for those things ourselves.

But the draft is where authorship lives, and that’s the part worth keeping human. Not because a watermark might catch you out, but because your thinking, perspective and voice are what your audience is actually there for.

What Should Businesses Do About AI Disclosure?

We don’t think the answer is to panic and stop using AI. But businesses do need a clear and defensible position on AI transparency.

Decide What AI Use You’re Comfortable With

Start internally. 

Where can AI support your team, and where does human authorship matter? You might want stricter controls around customer communications, thought leadership, photography or brand design than you do around internal research or administration.

You can’t communicate your position until you’ve decided what that position actually is.

Separate AI Assistance from AI Generation

Don’t treat every interaction with an AI tool as equivalent.

There’s a meaningful difference between asking AI to proofread a human-written article and asking it to produce the entire article from a prompt.

Your internal policies and external disclosures should recognise that distinction.

Keep Records of How Important Work Was Created

For high-value creative work, keep a sensible record of the process.

This could include original files, drafts, version histories, prompts, source photography or design development. 

Provenance is far easier to prove when you’ve documented it as you go.

Check What Your Contracts Say

For agencies in particular, AI watermarking creates some interesting questions around authorship and warranties.

Anthropic’s position is that a watermark helps establish whether Claude might have produced or processed something. It doesn’t change who owns the output or who’s legally responsible for it.

But that’s still worth checking against your own agreements. What have you promised clients about authorship, AI involvement and IP assignment?

Understand What Your Tools Are Adding

Find out whether the AI platforms you use attach Content Credentials, SynthID or another form of watermark or metadata.

And don’t assume that information will survive indefinitely. Re-saving files, converting formats, taking screenshots and uploading content to another platform can all affect metadata.

Knowing what your tools add is becoming part of understanding your own content workflow.

Be Transparent Where it Actually Helps Your Audience

More labelling doesn’t automatically mean better transparency. 

Ask what your customer or user genuinely needs to know. 

If AI materially created or altered something, disclosure may be appropriate or legally required. If an AI tool corrected one spelling mistake, an enormous “MADE WITH AI” banner arguably gives them a less accurate impression of how the content was made.

The goal should be meaningful transparency, not disclosure for disclosure’s sake.

Are Content Credentials the Answer?

AI disclosure is becoming more sophisticated, but sophistication doesn’t automatically equal clarity.

Content credentials and AI watermarks can create a useful record of provenance, and for images, video and audio we’re moving towards systems where that information can follow a piece of content and be checked. 

Text remains much harder to verify. A watermark can tell us a model interacted with some words without proving who wrote them, a missing watermark may not prove the opposite, and much of this sits in metadata the average customer will never inspect. 

That brings us back to the purpose of AI transparency in the first place. If businesses are going to label AI-generated content, those labels need to help people understand how AI was actually involved.

For businesses, technology can only take us so far. Clear policies, sensible disclosure and transparency about how you use AI still matter, particularly as the technology and regulation continue to develop.

Content Credentials might become an important part of that. AI watermarking might too. But neither removes the need for businesses to make thoughtful decisions about where AI belongs in their creative process and what their customers genuinely need to know.

Because ultimately, transparency isn’t about proving that you’ve followed a labelling system. It’s about giving people an honest understanding of how the work in front of them was made.

Sources

• Adrian Parham (22/06/2026). EU AI Act Chatbot Disclosure and Deepfake Labeling: July 22 Signatory Deadline. Tech Times.
• Aisha Down (17/08/2026). Claude to start watermarking AI-generated text – but will it make quality worse? The Guardian.  
• Alan Baker & Leo Salem (11/02/2026). The EU AI Act – what does it mean for UK organisations that use or provide AI systems? Farrer & Co.
• Aminu Abdullahi (21/07/2026). YouTube Tightens Monetization Rules for AI-Generated and Low-Quality Content. TechRepublic.
• Andy Betts (11/08/2026). The AI Detection False Economy: Fueling FOW (Fear of Writing). Search Engine Journal. 
• Anon (09/05/2024). Partnering with our industry to advance AI transparency and literacy. TikTok Newsroom. 
• Anon (19/05/2026). Advancing content provenance for a safer, more transparent AI ecosystem. OpenAI.
• Anon (27/05/2026). Improving AI labels for viewers and creators. YouTube Official Blog. 
• Anon (29/05/2025). Disclosure of AI in Advertising: Striking the Balance Between Creativity and Responsibility. ASA. 
• Anon (11/06/2026). AI and Deepfakes: Four Things Advertisers Need to Know Before They Hit “Run”. ASA.
• Anon (28/07/2026). C2PA Welcomes TikTok to Steering Committee, Advancing the Adoption of Content Credentials at a Global Scale. C2PA (Coalition for Content Provenance and Authenticity).
• Anon (30/07/2026). Use of Content Credentials to identify synthetic and non-synthetic content. C2PA.
• Anon (31/07/2026). EU transparency code for AI-generated content signed by 190 signatories. Cade Project.
• Anon (11/08/2026). Introducing a New Label for AI-Generated Artist Identities on Spotify. Spotify Newsroom.
• Anon (08/2026). How Claude marks AI-generated content. Claude Support.
• Anon (14/08/2026). How Claude’s text watermark works. Anthropic News.
• Anthony Ha (15/08/2026). Anthropic shares more details about how Claude’s new watermarks will work. TechCrunch. 
• Cabinet Office. AI to power change at the heart of government as Lord Vallance appointed chair of new PM AI Taskforce. GOV.UK.
• Frank Landymore (16/08/2026). People Horrified That They’ll Be Busted Now That Anthropic Is Watermarking AI Content. Futurism. 
• Jess Weatherbed (19/05/2026). Google is trying to make deepfake detection more accessible for everyone. The Verge. 
• Jess Weatherbed (17/08/2026). Anthropic explains how Claude’s invisible text watermarks will work. The Verge. 
• Laurie Richardson & Pushmeet Kohli (19/05/2026). Making it easier to understand how content was created and edited. News from Google. 
• Markus Reinisch (28/07/2026). Meta is Signing the EU AI Act Code of Practice on Transparency of AI-Generated Content. Facebook News.
• Matt G. Southern (11/08/2026). Anthropic To Mark Claude Text & Files Under EU AI Act Code. Search Engine Journal.
• Roger Montti (13/08/2026). Why Anthropic’s Claude Watermark May Be A New Text-Marking Method. Search Engine Journal. 
• Roger Montti (15/08/2026). Anthropic Reveals What The Watermark Is And How It Can Be Defeated. Search Engine Journal. 
• Sydney Bradley, Katherine Li & Dan Whateley (31/07/2026). From Substack to YouTube, here are the social platforms cracking down on AI slop. Business Insider. 
• Thibault Spirlet and Henry Chandonnet (18/08/2026). 3 arguments for and against AI watermarks. Business Insider.