You’ve probably come across an image online that looked a little too perfect. Maybe it showed a celebrity doing something unusual, or a photograph of an event you couldn’t find anywhere else. Was it real, edited, or generated with AI?
Sometimes there are obvious clues. Other times, looking at the image isn’t enough. Modern image generators can produce convincing results, and the same problem extends to videos, voices and written content.
That’s where AI watermarking comes in.
The idea is to leave a detectable signal in AI-generated content so that compatible tools can recognize it later. Google uses a technology called SynthID, while OpenAI uses a combination of invisible watermarks and Content Credentials for supported media.
There’s a catch, though. These systems don’t work like a universal AI detector. They’re designed to recognize particular signals, and those signals aren’t necessarily present in every AI-generated file.
Let’s look at what an AI watermark actually is, how it differs from content metadata, and what you can realistically check when you’re unsure about an image or video.
What Is AI Watermarking?
A watermark is usually something visible: a photographer’s name across an image, a television channel’s logo, or a company mark added to a document.
AI watermarking often works differently. Instead of putting a visible label on the content, the generating system embeds a pattern that people normally can’t see or hear.
A compatible detector can then look for that pattern and report whether it recognizes the watermark.
For an AI-generated image, the signal might be introduced into the image data during generation. Audio watermarking uses an inaudible signal. Text watermarking is different again because it relies on statistical patterns in the words or tokens selected by a language model.
The important thing to understand is that a detector needs to recognize the watermarking system involved. There’s no guarantee that a tool built to detect one provider’s watermarks will recognize another provider’s output.
How Do Invisible Watermarks Work?
Imagine generating a photograph of a mountain landscape using an AI image generator.
While the image is being created, a watermarking system can introduce subtle changes into its data. Those changes are designed to preserve the appearance of the image while leaving a pattern that a specialized detector can recognize.
You wouldn’t normally notice the pattern just by opening the image in a browser.
Now suppose that image gets resized or compressed before it’s shared online. A robust watermark may still be detectable afterward, depending on how it was designed and how much the image changed.
That’s one advantage over a visible logo in the corner, which can often be removed simply by cropping the image.
But invisible doesn’t mean permanent. Extensive editing, heavy compression or other transformations may weaken the signal enough that the detector no longer recognizes it.
Google SynthID: Watermarking Images, Video, Audio and Text
One of the better-known examples is SynthID, developed by Google DeepMind.
SynthID is designed to embed digital watermarks into supported AI-generated images, videos, audio and text. The implementation differs depending on the type of content.
For images and video, the watermark is embedded in the generated visual data. For audio, it uses an inaudible signal. Google has designed these watermarks to withstand certain common changes, although detection isn’t guaranteed after every kind of edit.
Text is more complicated because there are no pixels or audio samples to modify.
How Does AI Text Watermarking Work?
Language models generate text by selecting tokens one after another. At each step, the model considers a range of possible next tokens and how likely each one is.
A text watermarking system can subtly influence those choices to create a statistical pattern across a passage.
The result still reads like ordinary text. It doesn’t require adding a hidden phrase or an unusual symbol that somebody could easily spot.
Detection becomes harder with shorter passages or text that gives the model little freedom in its wording. Rewriting or translating a passage can also weaken the pattern.
Google explains the technology and its limitations in the SynthID Text documentation.
What Are C2PA Content Credentials?
Another term you’ll see in discussions about AI-generated images is Content Credentials.
These are based on the C2PA standard, which provides a way to attach verifiable provenance information to digital media.
Provenance simply means information about where something came from and what happened to it.
For example, Content Credentials might record which application created an image, whether AI generation was involved, or which editing steps were recorded by participating software.
The information can be cryptographically signed, allowing compatible verification tools to check its integrity and the identity of the signer.
This is not the same thing as embedding an invisible watermark into an image.
C2PA records provenance information. An invisible watermark embeds a detectable signal into the content itself.
And C2PA isn’t exclusively for AI. Cameras, editing applications, news organizations and other publishers can use it with ordinary photographs and videos.
You can read more about the standard at C2PA.org.
AI Watermarks vs Content Credentials: What’s the Difference?
The two technologies are easier to understand when you compare what each one is designed to do.
| AI Watermark | C2PA Content Credentials | |
|---|---|---|
| How it works | Embeds a detectable signal into content | Associates signed provenance information with a file |
| What it can indicate | Involvement of a supported AI system | Recorded origin, creation or editing information |
| Effect of editing | Some edits may weaken detection | Metadata may be removed or disconnected from the file |
| Only for AI-generated content? | Usually designed for supported AI outputs | No |
| Does it prove a claim is true? | No | No |
These technologies are often used together because they have different strengths.
Content Credentials can provide more detailed information about a file’s recorded history. An embedded watermark may remain detectable even after some of that metadata has been removed.
Neither provides a complete history of everything that happened to the content.
How to Check if an Image Was Generated with AI
Suppose you download an image and want to know whether it came from an AI generator.
I’d start by checking the original file for recognized provenance information rather than relying immediately on a website that claims to detect AI images from their appearance.
1. Check OpenAI’s Verification Tool
OpenAI provides a public content verification tool for supported images and audio files.
The tool checks for recognized provenance signals associated with OpenAI-generated content, including supported SynthID watermarks and C2PA Content Credentials.
- Open the verification website.
- Upload the original image or supported audio file.
- Check whether the tool identifies a supported provenance signal.
- Read the result carefully before drawing a conclusion.
If an OpenAI-associated signal is detected, that’s evidence that the file was generated or processed using supported OpenAI technology.
But the tool isn’t designed to identify every AI image created by every company. It also doesn’t verify written text.
For the most reliable result, use the original exported file rather than a screenshot, heavily edited copy or file repeatedly converted between formats.
2. Check for Google SynthID
Google offers SynthID verification through supported Google experiences, including options described on its official SynthID website.
For example, supported verification experiences can examine an uploaded image, video or audio clip for recognized SynthID watermarks.
This is useful when checking media associated with Google AI tools or other participating providers.
It is still a watermark check, not a universal authenticity test. A file can be AI-generated without carrying a signal that a particular detector recognizes.
3. Look for Content Credentials
If an image contains C2PA Content Credentials, a compatible verification tool can display the available provenance information.
Pay attention to who signed the credentials and whether the verification result reports them as valid.
A valid credential can help establish the source of recorded information, but it doesn’t mean that the image depicts a real event or that the accompanying caption is accurate.
Also remember that some platforms and editing tools remove metadata. If a downloaded copy contains no credentials, that doesn’t necessarily mean the original file never had them.
Can AI Watermarks Be Removed or Lost?
Yes. No watermarking system should be treated as impossible to defeat or guaranteed to survive every modification.
Some systems are designed to remain detectable after ordinary edits such as compression or resizing. More substantial changes can weaken the signal.
For text, the problem is especially noticeable. A passage can be shortened, substantially rewritten or translated while preserving much of its meaning but changing the wording the watermark depends on.
Metadata has a different weakness. It can disappear during file conversion, export or sharing even when the visible content remains nearly identical.
This explains why providers increasingly combine watermarking with Content Credentials instead of relying on just one approach.
Can AI Detectors Reliably Identify AI-Written Text?
Not with the level of certainty that many detector websites suggest.
There’s an important difference between checking for a known watermark and using software to guess whether a passage looks AI-written.
A watermark detector searches for a specific statistical pattern introduced by a compatible generation system.
A general AI text detector usually looks at characteristics of the writing and estimates how likely it is to have come from a language model.
Those are not equivalent methods, and neither should be treated as definitive proof of authorship.
Short passages, technical explanations, predictable wording and edited text can make detection difficult. Even a watermark detector may miss content that originally carried a signal.
That matters when somebody is using a detector to judge a student’s assignment, an employee’s writing or a published article. A probability score alone isn’t enough to establish who wrote something.
What’s New in AI Watermarking?
AI provenance has become more important as generated content has spread from images and chat responses to audio, video and everyday publishing tools.
OpenAI expanded its image provenance work in May 2026, followed by support for watermark verification of eligible audio in July.
On October 5, 2026, OpenAI announced textGrain, its text watermarking approach, in connection with the EU AI Act’s transparency requirements.
The initial approach allows API customers to opt in to watermarking for selected models, while eligible ChatGPT and Codex text output in the European Union is covered by a phased rollout.
OpenAI’s text watermark detector isn’t being offered as a general public tool at launch. Access is initially restricted to approved researchers and organizations studying its reliability.
You can read the details in OpenAI’s text provenance announcement.
Google has also continued expanding SynthID and its verification options. These developments make watermarking more useful, but they still don’t create a single detector that covers every AI model and every kind of content.
A Watermark Doesn’t Tell You Whether Something Is True
Consider a photograph shared online with the caption claiming it shows something that happened today.
If a provenance checker finds an AI watermark, that gives you useful information about how the image may have been created.
But if it finds no watermark, you still don’t know whether the photograph is authentic.
It might be an AI-generated image from a system the detector doesn’t recognize. It might be a real photograph from several years ago. Or it might be a genuine image presented with a misleading caption.
Even valid Content Credentials don’t establish that the claims made about a photograph are true. They provide evidence about its recorded origin and editing history.
When the claim matters, look for the original source, compare it with reliable reporting and examine the surrounding context rather than treating one detection result as the final answer.
The NIST report on synthetic content transparency is a useful technical reference on watermarking, provenance and detection.
Try Checking an AI-Generated Image Yourself
If you want to see how provenance checks work, try them with an image you generated yourself using a supported tool.
Download the original file, keep it unchanged and open the provider’s official verification tool. Check which signals are detected and what the result actually says.
For a more useful comparison, keep a copy of the original verification result alongside the image. This gives you a record of what the checker reported for the original file.
If you’re documenting the test, screenshots of the file and the verification result are far more useful than claiming an image looks AI-generated just because its lighting or details seem unusual.
And if no signal appears, don’t immediately label the image as human-made. All you’ve established is that the checker didn’t recognize a supported watermark or provenance record.
