Krux
May 21, 2026
OpenAI Images Now Carry Watermarks That Survive Cropping
Published: May 21, 2026 at 12:24 AM
Updated: May 21, 2026 at 12:24 AM
100-word summary
OpenAI is embedding dual provenance signals into every generated image: C2PA metadata plus Google DeepMind's SynthID watermark. The watermark survives cropping, compression, and filters with a 98.7% detection rate. A new public verification tool lets anyone upload an image to check if it came from ChatGPT or OpenAI's API. The catch? It only spots OpenAI content for now, and determined actors can still strip the signals. The company plans to add cross-platform detection in coming months. For the first time, a major AI company is making "who made this" as standard as a file format.
What happened
OpenAI is embedding dual provenance signals into every generated image: C2PA metadata plus Google DeepMind's SynthID watermark. The watermark survives cropping, compression, and filters with a 98.7% detection rate. A new public verification tool lets anyone upload an image to check if it came from ChatGPT or OpenAI's API. The catch? It only spots OpenAI content for now, and determined actors can still strip the signals. The company plans to add cross-platform detection in coming months.
Why it matters
For the first time, a major AI company is making "who made this" as standard as a file format.