What is SynthID, and how do you check for it?

SynthID is an invisible watermark for AI-generated content, made by Google DeepMind. When an AI model that uses it creates an image, it hides a pattern in the pixels that people cannot see but a detector can read. Google uses it in images from its own AI, and in 2026 OpenAI started adding it to images from ChatGPT. This guide explains what the watermark is, every official way to check for it, and what a result does and does not tell you.

What SynthID is

Google DeepMind describes SynthID as a watermark for images, audio, text and video (Google DeepMind). For images and video, Google says it is designed to stand up to changes such as cropping, adding filters, changing the frame rate and lossy compression. That is the key difference from labels stored in a file’s metadata, which a screenshot or a re-save throws away.

It is widely used. On 7 October 2026 Google said SynthID had marked more than 180 billion images and videos (Google).

It is not unbreakable. Google’s Gemini help says that after many alterations the watermark may not be detected (Gemini Apps Help), and OpenAI notes that screenshots can make its signals harder to detect (OpenAI Help Center).

Which AI tools add SynthID

Company	What carries SynthID	Since
Google	Images, video, audio and text from Google’s AI models, including the Gemini app	2023 onwards
OpenAI	Images generated through ChatGPT, Codex and the OpenAI API, then audio	19 May 2026 (audio 31 July 2026)
NVIDIA	Video from its Cosmos world foundation models	Partner, date not published
Kakao	AI content from Kakao services	Partner since May 2026

OpenAI left out models planned for early retirement, including gpt-image-1, gpt-image-1.5 and sora-2 (OpenAI Help Center). In its 7 October 2026 update, Google listed its partners as OpenAI, NVIDIA, Kakao and, soon, Apple (Google). Midjourney, Stable Diffusion and FLUX are not on that list, so a SynthID check says nothing about their images.

How to check an image for SynthID

Only the companies that add SynthID can read it. Pick the check that matches where the image probably came from.

Google’s SynthID Detector

The SynthID Detector checks images, video and audio for the watermark from Google and its partners. Google opened it to everyone on 7 October 2026, in English. It says plainly that it is not a general AI detector: it only looks for SynthID. Its privacy notice mentions daily usage quotas.

The Gemini app

Sign in to Gemini, upload the image (up to 100 MB) and ask, for example, “Was this image created or edited by Google AI?”. Gemini only recognises content made by Google AI tools, and allows about 10 image checks in any 24 hours (Gemini Apps Help). It also reads Content Credentials when the file has them.

Google Lens and Chrome

Since May 2026 you can ask “Is this made with AI?” in Google Lens, AI Mode, Circle to Search and Gemini in Chrome, which run the same SynthID check (Google).

OpenAI’s verify page and API

For images from ChatGPT, Codex or the OpenAI API, upload the file to OpenAI’s verify page, which does not ask you to sign in. Developers can run the same check with the Content Provenance API, which reads OpenAI’s SynthID watermark and its Content Credentials (OpenAI). OpenAI says the API check is currently free and subject to usage limits. The uncrop.si AI image checker runs it with your own OpenAI key, from an isolated panel that sends the file straight to OpenAI.

What a result means

A positive result from Google means all or part of the image was created or edited by Google’s AI models (Gemini Apps Help). From OpenAI, it means the image was probably made or exported with OpenAI’s tools.

A negative result proves much less. Google says an image without the watermark may still come from other AI systems, and OpenAI lists the reasons its own images can test negative: the metadata was stripped, the watermark was degraded, the image came from an older model, or it was made before these signals existed (OpenAI Help Center). Mistakes the other way are rare: the SynthID Detector says that in very rare cases it can trigger on content that was never watermarked.

SynthID and Content Credentials

The two work together. Many images carry both, and they answer different questions.

Question	SynthID	Content Credentials
What it is	An invisible watermark in the pixels	A signed record stored in the file
What it tells you	That a participating company’s AI made or edited the image	Who signed it, which tool made it, and what edits followed
Who can read it	Only Google’s and OpenAI’s detectors	Anyone, with free tools
Screenshots and re-saving	Built to survive many edits, though they can weaken it	Usually removed

ChatGPT images carry both. When we checked ChatGPT images made in October 2026, their Content Credentials were signed by OpenAI and recorded that an invisible watermark had been added. Read Content Credentials explained for the signed record, and how to tell if an image is AI-generated for the full routine.

Can SynthID be removed?

Google says the watermark is not infallible and that someone who tries many transformations may find one it does not survive. That is a statement about its limits, not a feature. Removing a provenance watermark so an AI image passes as a real photo is deceptive, and rules in China and the EU require AI-generated images to be labelled. uncrop.si does not offer removal. If one of your own photos is wrongly flagged as AI, the original file, and its Content Credentials if your camera or app added them, are the way to show where it came from.