⚠️ Affiliate Disclosure: This post is not an affiliate associated review. It may become one in the future and this disclosure will reflect that. This does not affect my reviews and I only recommend tools I genuinely believe in.
What Is AI Watermarking?
AI watermarking is how a file carries a record of the fact that AI made it. Sometimes that record is a visible logo stamped across a picture. More often now it is invisible — a signal woven into the pixels, the audio waveform or the word choices, which you cannot see or hear but software can read.
Until recently this was a niche concern. It is not any more. New EU rules requiring AI content to be marked came into force on 2 August 2026, and Google has begun letting anyone check an image’s origin directly from Search. If you use AI tools for work, this now affects you.
The Two Kinds of Watermark
Almost every confusion about this topic comes from mixing these up.
1. Visible watermarks — the free-plan logo
This is the badge or logo an AI tool stamps on your download when you are on a free plan. It exists for one reason: to encourage you to upgrade. It is a pricing decision, not a safety measure.
2. Invisible watermarks and content credentials — the provenance record
These are designed so you never notice them. There are two main systems, and they work quite differently:
- SynthID — Google DeepMind’s technology, which embeds an imperceptible signal inside the content itself. Because it lives in the pixels or the waveform rather than in a file’s metadata, it is built to survive ordinary handling like cropping and compression.
- C2PA Content Credentials — an open industry standard that attaches a signed history alongside the file, recording what created it and what edited it since. Think of it as a receipt travelling with the image.
The scale is larger than most people realise. Google reports having watermarked over 100 billion images and videos, and 60,000 years’ worth of audio, with SynthID.
Who Uses AI Watermarking?
Far more of the industry than you would guess, and the list grew sharply in 2026.
| Who | What they do |
|---|---|
| SynthID across its generative models; verification built into Search, Gemini and Chrome | |
| OpenAI | Bringing SynthID to more of its AI-generated content |
| ElevenLabs | Adopting SynthID for AI-generated audio |
| Meta | Labelling camera-captured media with Content Credentials on Instagram |
| YouTube | Labels identifying AI-generated content |
| Canva | Named among Google’s content-transparency partners |
| Google Pixel | Pixel 10 was the first phone to add Content Credentials in its camera app |
| NVIDIA | Watermarking video from its Cosmos models, with Google |
Note what Pixel is doing there, because it inverts the usual logic: it marks photographs to prove they came from a real camera. As AI output becomes indistinguishable from the real thing, proving something is authentic matters as much as flagging what is synthetic.
How to Check Whether Something Is AI-Generated
You no longer need a specialist tool. In the Gemini app, and now rolling out across Google Search and Chrome, you can upload or point at an image and simply ask “Is this made with AI?”. Google says the feature has already been used 50 million times worldwide.
One important limit: this only detects watermarks that are actually there. A model that does not watermark its output will not be caught, and an absence of a watermark is not proof that something is real. It answers “was this made by a participating AI tool”, not “is this genuine”.
Can You Remove an AI Watermark?
This is the question everyone actually arrives with, so here is a straight answer — but it depends entirely on which kind of watermark you mean.
One distinction worth keeping clear before anything else: editing a visible corner logo out of an image you hold the rights to is an editing job — any photo editor can do it. Trying to defeat an invisible provenance mark like Google’s SynthID is a different act entirely. You are not tidying a graphic; you are erasing the record of where the image came from, and as the rest of this guide explains, both the technology and the law treat those two things very differently.
The visible logo: yes, by paying
If you are looking at a badge on a free-plan export, the intended route is to upgrade. That is the whole point of it — it is a nudge towards a paid tier, and paying makes it disappear.
Cropping or editing it out is a bad idea for a reason people often miss. On most free tiers you do not hold commercial rights to the output in the first place. We found exactly this while reviewing Suno, Udio and Nafy AI: the free plan is for personal use only, and some platforms explicitly reserve the right to use anything you make there in their own marketing. Removing the badge does not grant you rights you were never given. If you need clean output for client or commercial work, the paid plan is not just the easier path, it is the only lawful one.
The invisible watermark: not reliably, and increasingly not legally
Invisible watermarks are a different matter, and we are not going to publish a method for defeating them. What is worth understanding is why the question is becoming moot.
SynthID is embedded in the content itself and is specifically designed to survive the ordinary things people do to files — cropping, resizing, compression, filters. Researchers have shown that no watermarking scheme is perfectly robust, and that remains an active area of study, but it is not a matter of ticking a box in an editor.
C2PA Content Credentials are more fragile, because they travel alongside the file rather than inside it. In practice they are often lost by accident: plenty of platforms strip metadata when they re-encode an upload. That is a genuine weakness of the standard rather than a technique, and it is one reason the industry is pairing credentials with embedded watermarks instead of relying on either alone.
The Legal Position Changed This Month
This is the part that turns watermarking from a technical curiosity into a business consideration.
Article 50 of the EU AI Act — the transparency provisions — took effect on 2 August 2026, with the European Commission publishing its final guidelines on 20 July. In broad terms, providers of AI systems that generate synthetic image, audio, video or text must ensure those outputs are marked in a machine-readable way and detectable as AI-generated. Systems that talk to people, such as chatbots, must make clear that they are AI.
Generative systems already on the market before that date have until 2 December 2026 to meet the machine-readable marking requirement. Penalties for breaching Article 50 run to €15 million or 3% of worldwide annual turnover, whichever is higher.
What This Means If You Use AI Tools for Work
Nothing here should stop you using AI tools. It should change how you handle what comes out of them.
- Assume it is detectable. Plan on the basis that anything from a major AI tool can be identified as AI-generated, now or later. Build that into what you promise clients.
- Pay for the tier that gives you rights. The watermark is usually the visible half of a licensing question. Free almost always means non-commercial.
- Say so. Disclosure is becoming the norm across platforms, and being upfront costs far less than being found out.
- Do not treat a missing watermark as proof. Plenty of AI output carries no mark at all. Detection tools tell you what they found, not what is true.
- Keep your originals. If provenance ever matters to your work, the unaltered file with its credentials intact is what proves your case.
Frequently Asked Questions
Can Google detect AI-generated images?
Yes, where the image carries a SynthID watermark or C2PA Content Credentials. You can ask “Is this made with AI?” in the Gemini app, and the capability is rolling out through Google Search features such as Lens, AI Mode and Circle to Search, and into Chrome. Images from tools that do not watermark will not be flagged.
Does AI watermarking apply to text as well as images?
It can. Google has open-sourced its SynthID text watermarking technology, and the EU’s transparency rules cover synthetic text alongside image, audio and video. Text watermarking is harder than image watermarking, though, and is less widely deployed in practice.
Is it illegal to remove an AI watermark?
There is no single UK law that says so in those words. But it will normally breach the tool’s terms of service, and if you are stripping a mark in order to pass AI work off as human-made, you are in the territory the new transparency rules were written to address. Using AI is fine. Actively concealing it is where the risk sits.
Do I have to tell people I used AI?
It depends on the platform and the context rather than on one blanket rule. YouTube requires disclosure of realistic AI-generated content, many publishers now ask for it, and the EU rules push firmly in that direction. As a working principle, if a reasonable person would want to know, tell them.
Does watermarking mean AI content ranks worse on Google?
No. Google’s position is that it rewards helpful, original content regardless of how it was produced, and penalises low-value content made purely to game rankings. Watermarking is about transparency, not a ranking signal. Our guide to Google Preferred Sources covers how Google surfaces trusted publishers.
Which AI tools add visible watermarks to free exports?
It varies, and it changes often, so check the current terms before you commit. The pattern we see repeatedly in testing is that free tiers restrict downloads, add branding, or withhold commercial rights — sometimes all three. We flag this in every review, including ElevenLabs, Canva Magic Studio and Midjourney.
Read more about how we test AI tools.
James is an SEO and content specialist with over 12 years’ experience, and tests AI tools so everyday people don’t have to. WhatAITools.co.uk covers honest, no-jargon reviews to help you find the right tool for your needs.
Leave a Reply