AI Labels: Loss of Evidence After Screenshots

A recent test by AI Insider has shown that the AI label for four different AI services, namely ChatGPT, Gemini, Grok, and Meta AI, completely disappears after taking a screenshot. This raises questions about the effectiveness of the current EU regulations, which require AI services to make images recognizable as AI since August 2026. The test took place on October 19, 2026, where four AI models generated the same image of a woman on a bicycle in the Amsterdam canal.

The test revealed that while ChatGPT and Gemini initially displayed a reliable label as derived from their respective publishers, both labels were lost after a screenshot. With Grok, the label also disappeared, but this was already not trusted due to the method of signing. Meta AI, on the other hand, was able to leave an invisible watermark that survived the screenshot. This points to a significant problem for the identification of AI-generated images, especially when shared on social media.

The test utilized a public checker of Content Credentials, which could read the labels on the images. After editing or resizing the images with the features of a screenshot, none of the AI services could identify the images anymore. OpenAI itself could not determine its own invisible watermark, while Gemini and Meta left some traces behind. This makes it clear that although the implementation of AI labels is a step forward, their effectiveness in practice is far from satisfactory.

European legislation requires machine-readable labels for AI services, but the current test demonstrates that a screenshot can effectively erase these labels. This creates uncertainty about the origin and authenticity of images and indicates that consumers and organizations must remain vigilant when sharing AI-generated content. The results of this test may prompt further investigation, especially as the transitional period for the regulations expires on December 2, 2026.

Read the full article from AI Insider.