AI Watermarks Explained: How to Check for Them and How to Remove Them

An AI watermark is not one thing, and that confusion is why most advice about it is wrong. There is the visible logo some tools stamp into the corner of a video, which you can crop away in about four seconds. Then there is the invisible layer sitting inside the same file, which survives the crop, the screenshot and the re-upload, and which most people do not know is there at all.

That second layer stopped being a research curiosity this month. Since 2 August 2026, Article 50 of the EU AI Act has required providers of generative AI systems to mark their output in a machine-readable format, and the major labs have shipped. This guide covers what each kind of watermark actually is and which tools stamp which. It also covers how to check any file yourself, using free tools that already exist, and what removing one does and does not accomplish.

The Key Takeaways

• Two different technologies: a visible logo you can remove, and invisible signals such as SynthID and C2PA that you generally cannot.
• Removing the logo changes nothing underneath. OpenAI embeds both C2PA metadata and a SynthID watermark in generated images, and the watermark is designed to survive cropping and compression.
• You can check files yourself today. openai.com/verify reads OpenAI images for free, and the Gemini app will tell you whether Google AI made a file.
• Claude now watermarks text, ChatGPT does not. Every Claude model launched on or after 2 August 2026 carries it, worldwide. OpenAI still describes text marking as a goal.
• The law targets providers, not you. Article 50 obliges the companies building these systems. Editing your own output breaks no rule.

What Is an AI Watermark?

An AI watermark is a signal embedded in AI-generated content so that software can recognise where it came from. It covers two very different things. There is the visible logo some tools stamp on images and video. Then there are invisible marks such as Google's SynthID and C2PA Content Credentials, which sit inside the file or inside the words themselves and survive ordinary copying.

Keeping those two apart is the whole game. Almost every page that ranks for this topic treats them as one, which is how you end up with guides confidently promising to strip something they have not understood.

So take them one at a time.

Visible watermarks: the logo in the corner

This is the one everybody pictures. A generated video arrives with a bouncing logo in the lower corner, or an image carries a small wordmark. It is a branding decision and a disclosure decision at the same time, and it is applied at the surface of the picture.

Because it lives in the pixels you can see, it comes off the way any pixel does. Crop the frame, reframe for a different aspect ratio, or generate on a tier that does not apply it. There is nothing clever happening here, and nothing to defeat.

Invisible watermarks: SynthID and C2PA

The invisible layer is two separate mechanisms that often travel together, and they fail in opposite ways.

C2PA Content Credentials are signed metadata attached to the file. They can record which tool made it, when, and what happened to it since. Because it is metadata, it carries rich detail, and because it is metadata, a platform that strips metadata on upload will remove it without meaning to.

SynthID works the other way. Google DeepMind's system embeds the signal into the content itself: into pixel values for images, the waveform for audio, and token choices for text. It carries far less information than metadata, and it is much harder to lose. According to Google DeepMind's own documentation, the image and video watermark is designed to survive cropping, filters, frame-rate changes and lossy compression, while the audio version survives added noise, MP3 compression and speed changes.

That is why the serious providers now ship both. Metadata gives you the detail, the embedded watermark gives you the durability, and each covers the other's weakness.

The third meaning people are actually searching for

There is a third sense of the phrase that has nothing to do with either: using an AI tool to erase a watermark somebody else put on their own photo. That is a copyright question rather than a provenance one, and it is not what this guide covers.

What Each AI Tool Stamps on Your Output

Coverage varies by product, model, export path and file type, and it changes month to month. Here is where the major tools stood at the end of August 2026.

The row worth staring at is ChatGPT text. A great deal of anxious writing assumes OpenAI has been quietly marking prose for years. It has not. Its own help documentation describes expanding provenance signals to all modalities including text as a goal, not a shipped feature, which puts it behind Anthropic on exactly the modality people worry about most.

How to Check Any File for an AI Watermark

This is the part almost nobody writes down, because the tools are new and the affiliate pages have no reason to mention free ones. Work through it in order.

Checking an image made with ChatGPT

Go to OpenAI's verification tool, currently labelled a research preview, and upload the file. It accepts PNG, JPG and WEBP, and reports whether it finds C2PA metadata, a SynthID watermark, or no supported signal. It is built to recognise images from ChatGPT, Codex and the OpenAI API, and it will not identify content from other companies.

Two practical notes from OpenAI's own instructions, both of which change your results. Upload a single image rather than a collage, and if you are working from a screenshot, crop it tightly around the picture itself. Extra chrome around the edges gives the detector less to read.

Checking Google and Gemini content

Google's dedicated SynthID Detector portal exists, but it is gated. DeepMind describes collaborating with journalists and media professionals to test it, and ordinary users get a waitlist form. Several guides list it as though you can simply open it, which is not the case today.

The route that does work is free and immediate: upload the file in the Gemini app and ask whether it was created or altered by Google AI. That covers images, video and audio across Gemini, Veo, Lyria and NotebookLM. It is a strange way to run a verification service, and it works.

Checking text from Claude

You cannot, yet. Anthropic has said it will offer a watermark detection API and is still working out the implementation. Until that ships there is no way for a reader, a teacher or a journalist to test a passage for Claude's mark, and anything advertising itself as a Claude watermark checker today is guessing.

What a "no signal found" result actually means

Less than people assume. OpenAI lists several reasons a file that really was AI-generated comes back clean. It predates provenance signals. It came from an unsupported product or export path. The metadata was stripped during upload or conversion. The watermark degraded through compression, cropping or format changes. An audio clip can also be too short to carry a readable signal.

So a hit is meaningful and a miss is not. This is the opposite of how people read the statistical AI detectors that guess from writing style, where a confident accusation is often wrong in the other direction. If you are assessing an image rather than prose, our walkthrough on how to tell if a photo is AI generated covers the forensic checks that sit alongside provenance.

How to Remove an AI Watermark

Here is the honest version, which the tool pages have every commercial reason not to give you. Work down it in order, because the first option is usually the one people skip.

Start with the official route

Check what your plan offers before you reach for anything else. Providers generally attach visible marks to free and lower tiers and offer clean exports higher up, and the terms differ between products, so read the ones covering the tool you are using. A clean file at generation time beats a reconstructed one, and it costs you nothing in quality.

Cropping and reframing

Visible marks usually sit in a fixed corner. If you are cutting a landscape clip to a vertical format anyway, the mark often falls outside the new frame with no processing and no quality loss. This is the least destructive option available and it is free.

Why removal does not hide the AI origin

Cropping or painting out the logo leaves the invisible layer exactly where it was. An OpenAI image carries C2PA metadata and a SynthID watermark, and SynthID is built to survive precisely the transformations a removal workflow applies. You have changed what a human sees and nothing about what software reads.

That matters most when the file leaves your hands. A video that looks clean to you still answers honestly when a platform, a newsroom or a verification tool asks it.

If you are picking a tool partly on this basis, our roundup of the best AI video generators tested and ranked notes how each handles export. The guide to Nano Banana's free limits covers what Google's image models attach on the free tier.

What editing does to a text watermark

Anthropic has published its own durability notes, and they are more candid than anything the detector industry offers. Light editing is likely to leave the mark intact. A complete rewrite removes it, though as Anthropic points out, at that stage it is arguable whether the text is still AI-generated in any meaningful sense.

Two details cut against intuition. Translation does not launder anything, because Claude chose every word of the translated output, so the result is freshly marked. And code is barely marked at all, since correct code leaves little room for arbitrary word choice, so the signal concentrates in comments and naming.

What the EU AI Act Actually Requires

Article 50 of the EU AI Act took effect on 2 August 2026, and it is the reason this all landed at once. Per the European Commission's own guidance, it requires providers of systems generating synthetic audio, images, video or text to mark those outputs in a machine-readable format, detectable as artificially generated. The marking has to be effective, interoperable and robust as far as is technically feasible.

Two limits are worth knowing, both set out in that same Commission guidance. Systems already on the European market before 2 August 2026 get until 2 December 2026 to comply with the marking duty specifically, and content generated before that date does not need labelling retroactively. The rules also carve out standard editing functions that do not substantially alter the input.

The point that most coverage skips: this obligation sits on providers, meaning the companies shipping the models. It is not a rule about what you may do with your own files. Someone rewriting their own Claude output is not breaching the AI Act, and any page implying otherwise has misread who the law is addressed to. That is also why Anthropic applied its watermark worldwide rather than in Europe alone, saying it does not yet have a durable way to scope it by region.

Does Claude Watermark Text?

Yes, worldwide. Anthropic's support documentation states that Claude models launched on or after 2 August 2026 support marking at launch. That covers the Claude app, the API, Claude Code, Cowork and Claude Tag, and older models are being added over time. Anthropic's technical post describes an adaptation of DeepMind's SynthID-Text method, which nudges word selection among candidates the model was already indifferent between. Nothing is forced, so quality holds, and the signal lives in the choices themselves rather than in hidden characters or metadata. That is why it survives being pasted into a plain text box.

What it does not do is more interesting than what it does. Anthropic states the watermark carries nothing identifying about the user, their organisation or their conversations. It cannot distinguish human-written text, it cannot tell you which other AI system produced something, and it cannot confirm that Claude was the sole author. Anthropic makes that last point itself: people use Claude to proofread, translate, summarise and convert files whose ideas came from somewhere else entirely, and all of that output carries the same mark. It answers one narrow question, which is whether Claude was involved.

Anyone reading it as a quality verdict has the wrong idea entirely. A watermark is a provenance signal, not a judgement, which is worth remembering given how much AI writing is now edited, checked and rewritten by people before it goes anywhere. If the privacy dimension of all this interests you, our breakdown of what OpenAI collects covers what else travels with your files.

The Bottom Line on AI Watermarks

Treat the visible logo and the invisible signal as separate problems, because they are. If a mark is spoiling a video you made, crop it or generate it clean on a tier that does not apply one. That is a legitimate production decision and nothing more.

If your actual question is whether a file is AI-generated, the answer now starts with provenance rather than with a style-guessing detector. Run the image through OpenAI's verifier, ask Gemini about Google content, and read a negative result as "no signal found" rather than "not AI". For Claude text, wait for the detection API, and treat anything claiming to do it today as noise. The tooling is young and uneven, but for the first time the honest answer to "how do I check this" is a set of steps rather than a shrug.

FAQ