What is SynthID?

Put simply, SynthID is a method for watermarking and detecting synthetic (e.g., AI-generated) media. It can be applied to images, videos, audio, and text. AI platforms can choose to use embed a SynthID watermark in content they generate and provide tools to detect the presence of that watermark.

SynthID watermarks are invisible to the human eye but detectable by a computer, even if the content is compressed or altered (such as screenshots, posting on social media, re-encoding, or even taking a photo of a print-out).

How does it work?

SynthID uses a form of steganography to embed a watermark in the content itself. In the case of images, a transformation is applied to the image before it's served to users.

Transformation applied to the image before it's served to users.

This is a heavily simplified representation of the transformation applied to an image. The actual transformation is more complex and imperceptible to the human eye. Details of the specific alterations are kept secret to prevent malicious actors from attempting to remove the watermark.

SynthID detectors know what this transformation looks like and can detect the presence of the watermark in content.

Each AI platform has its own unique watermarking transformation, so a SynthID detector for one platform will not be able to detect watermarks from another platform.

Unique SynthID transformations for each AI platform.

Is it accurate?

Yes. Unlike other unreliable "AI detection" methods, SynthID is not a statistical guess. The content itself is modified to include a watermark, and the detection process is deterministic. This means if you get a positive result from a SynthID detector, you can be very confident that the content was generated by the AI platform that embedded the watermark. False-positives are possible but very rare.

How can I check for a SynthID watermark?

Each AI platform has their own tools for detecting SynthID watermarks. There is no single tool for detecting all platforms' watermarks due to how the technology works.

Visit each of the tools below and upload the content you want to check:

How do I interpret the results?

If the tool returns a positive result, then you can be confident the content was AI-generated.

If the tool returns a negative result, then it's not possible to determine whether the content was AI-generated. It does not mean the content wasn't AI-generated. It's possible for watermarked images to produce a negative result if they have been significantly altered.

Regardless of the result, you should use it only as one signal for your interpretation. As always, you should seek corroborating information from trusted sources before coming to a conclusion.

Where can I learn more?

If you're interested in diving deeper into how SynthID works, refer to these resources:

I think you are wrong. How do I complain?

You can contact me at any of the methods listed on my website.