Does Removing Metadata Remove SynthID? No. Here's Why

No. SynthID is not metadata. It is a pattern embedded in the pixel values of an image, so any tool that only removes metadata, including BlankAI, leaves it exactly where it was. Here is how it works and what that means in practice.

What SynthID is

SynthID is an invisible watermark developed by Google DeepMind. A trained encoder adjusts the image in ways people cannot see, and a trained detector recognizes the pattern later. Google has published details of the image version and says it is built to survive common changes such as compression, resizing, cropping and screenshots. Google reports that more than 100 billion images and videos carry it.

Who uses it in 2026

• Google: all media generated by its models, including Gemini and Nano Banana images.
• OpenAI: ChatGPT, Codex and API images since 19 May 2026, alongside C2PA.
• Apple: announced for Apple Intelligence images and most edits in a software update later this year.
• NVIDIA, ElevenLabs and Kakao have also adopted SynthID for their generated media.

Meta does not use SynthID. Its Muse Image output carries Meta's own Content Seal, which Meta's detector reads and Google's does not.

How to check an image

Upload the image to the Gemini app and ask whether it is AI-generated. Google announced SynthID checks for Search and Chrome in May 2026, and offers an AI content detection API on Google Cloud. OpenAI provides its own verification for ChatGPT images. Because OpenAI uses SynthID, Google's checks recognize both companies' images.

Why metadata removal cannot touch it

A metadata remover works on the file's containers: EXIF, XMP, IPTC, C2PA and text chunks. It either deletes them or, like BlankAI, writes a new file from the pixels. In both cases the pixels, and the watermark inside them, carry over. BlankAI does not alter pixels to interfere with watermarks.

What about SynthID removal tools?

Academic work shows that watermarks can be weakened. The UnMarker attack presented at IEEE S&P 2025 reduced SynthID detection substantially, and regeneration approaches redraw an image with a diffusion model. These methods change the image, sometimes visibly in faces and text, and none guarantees removal. Browser tools that claim to erase SynthID with imperceptible tweaks are not credible against a watermark trained to survive JPEG compression.

There are also rules. The EU Code of Practice asks providers to prohibit removing markings in their terms and not to promote circumvention tools, China prohibits removing AI labels, and several generators' terms forbid removing provenance signals.

What this means for you

If an image came from Google or OpenAI, assume it can be identified as AI-generated no matter what you do to its metadata, and disclose it where platforms ask. Use metadata removal for what it does well: keeping your prompts, account details, location and editing history private.