What Is SynthID? The Invisible AI Watermark Now Inside ChatGPT, Gemini and More

AI watermarking stopped being a niche research topic in 2026 and became the default fabric of how AI-generated content is identified. If you make images with ChatGPT or Gemini today, they leave with an invisible signature baked in — whether or not you asked for it. Here is what SynthID actually is, who is now using it, and what it changes for anyone who creates with AI.

The news: SynthID went industry-wide in 2026

For its first couple of years, SynthID was a Google feature for Google's own models. That changed at Google I/O on May 19, 2026, when Google confirmed OpenAI, NVIDIA, Kakao and ElevenLabs as SynthID adopters — and OpenAI announced that every image from ChatGPT, DALL·E and the OpenAI API now ships with a SynthID watermark plus C2PA metadata.

The scale is already enormous: Google says SynthID has tagged more than 100 billion images and videos, plus the equivalent of 60,000 years of audio. What began as a single-vendor feature is now a shared, cross-industry default — a rare moment of alignment between companies that normally compete head-to-head.

What SynthID actually is

SynthID is an invisible watermark from Google DeepMind. Rather than stamping a visible logo, it embeds a statistical signal directly into the content at the moment of generation — into the pixels of an image, the samples of an audio clip, or the token choices of generated text. To a human it is imperceptible. To a detector that knows what to look for, it is a clear "this was made by AI" flag.

The property that makes it matter is robustness. SynthID is designed to survive the ordinary things that happen to a file: cropping, resizing, compression, screenshots, filters and format conversion. That is what separates it from metadata — which any "save as" or screenshot quietly destroys.

It spans four modalities: images (Imagen, Nano Banana), video (Veo), audio (Lyria), and text (SynthID-Text in Gemini).

Who embeds it now

The list of generators that bake SynthID into their output:

• Google — Gemini, Imagen, Nano Banana, Veo, Lyria (the origin)
• OpenAI — ChatGPT, DALL·E and the OpenAI API, for images (since May 2026)
• NVIDIA — its Cosmos foundation models (since 2025)
• Kakao — image generation (2026)
• ElevenLabs — AI-generated audio (2026)

A second tier is wiring up detection rather than embedding. Google is previewing a Content Detection API with partners including Canva, Snap, Shutterstock and Fox Sports — the plumbing platforms need to flag AI content at upload. SynthID detection is also rolling into Google Search, Chrome and Circle to Search.

SynthID vs C2PA: two layers, different jobs

Most of these rollouts pair SynthID with C2PA "Content Credentials." The two solve the same problem from opposite ends:

• C2PA is signed metadata attached to the file: rich provenance — what made it, when, and its edit history. But it travels outside the content, so a screenshot or a metadata strip removes it.
• SynthID is inside the content: less detail, but it survives when the metadata does not.

Together they form a deliberate dual layer — C2PA carries the full story, SynthID carries a fallback signal when the story gets stripped. That redundancy is the entire point, and it is why "just clear the EXIF" no longer means an AI image is clean.

What this means if you create with AI

For creators, media buyers and agencies, the practical shift is simple: AI-generated images now carry provenance signals by default, and platforms are actively building the detection to read them. So:

• An image from ChatGPT or Gemini is no longer anonymous — it can be identified as AI-made even after you edit it.
• Stripping metadata is not enough. The in-content signal persists through the edits that wipe C2PA.
• Expect "AI-generated" labels to spread across surfaces as the detection API rolls out to platforms.

None of this is inherently good or bad — it is provenance infrastructure, the same way HTTPS is transport infrastructure. But if any part of your workflow assumed AI images were untraceable, that assumption expired in 2026.

How to check an image

The honest first move is to detect, not assume. OpenAI previewed a verification tool that inspects an upload for both C2PA and SynthID and reports whether either signal is present; Google's SynthID Detector does the same across its ecosystem. If you want to know whether an asset carries a signal — your own or someone else's — run it through one of these before you build on it.

The reality of removal

Because SynthID lives in the pixels and is built to survive edits, it does not come off with a crop or a re-save. Removing it means regenerating the image so the statistical signal no longer holds — and even then it is best-effort. Watermarking is an active arms race: detectors improve, new watermark versions ship, and no method is guaranteed against every image or every future iteration.

Our Watermark Cleaner does exactly this — it diffusion-regenerates an AI image on a GPU to strip the SynthID signal while keeping the picture intact. It is deliberately framed as best-effort: always verify the result yourself rather than assume it is gone.

The takeaway

SynthID stopped being a Google curiosity and became the default fabric of AI content provenance in 2026. If you generate with ChatGPT, Gemini, or anything downstream of them, your images leave with an invisible signature — and the tooling to read it is spreading fast. Whether you want to label your AI content honestly or simply understand what platforms can now see, step one is the same: know the signal is there, and check for it.