The proliferation of AI-generated images has created demand for a new kind of tool: sites and apps that claim to determine whether an image was created or modified by AI. We tested 10 of the tools on the same image.
There are dozens of AI detection tools claiming to identify images generated or modified by AI, but their results are not always accurate.
We tested an image of the American sex offender Jeffrey Epstein that we knew was generated by AI. The image, showing a man with sunglasses, white hair and a beard, was posted in February 2026 with a claim that it showed Epstein still alive. But Epstein died in a US prison in 2019.
We tested the photo in 10 AI detection tools: WasitAI.com, Isgen.ai, AIorNot.com, Decopy.ai, Rephrasy.ai, Sightengine.com, Hivemoderation.com, Undetectable.ai, Mydetector.ai and ZeroGPT.com.
Some said it was real, one even ranking it "100% human".
Only three of the 10 tools – AIorNot, ZeroGPT and Undetectable.ai – correctly identified the Epstein image as being generated by AI.
The others said the image was real. That is what is known as a "false negative".
"A false negative is when the tool fails to detect that there is a problem,” says Tina Nikoukhah, vice-president of research at US-based GetReal Security. “Perhaps the tool is not up to date, or is not well-adapted for the task.”
Nikoulah continues: “Developing AI-detection software is complex. Most AI-detection tools are ‘classifiers’. That means they put images into two categories: ‘natural’ and ‘synthetic’. But sometimes the AI detectors have not been sufficiently trained on the images in question. They might be good at detecting images created by one AI generator, but not by another.”
AI detection tools can also be misled by images that are low-resolution or compressed. The low quality makes it harder to spot the AI.
'Created with Google AI'
Nikoulah says there are some safeguards in place: “Certain AI image generators embed an invisible watermark into the pixels of the images they create. It's like the watermark a photographer might put on a photo when he sells it. You can see his name written on the photo. It's the same idea here, except the watermark is invisible."
That's the case for images created by Google's AI tools. When you click the "About this image" tab in Google, you get a message saying it was “created with Google AI".
We contacted the detection sites that failed to identify the Epstein image as being AI. They said they would continue to improve their tools as AI technology evolves.
This article was published on the occasion of France's Media in Schools Week, March 23-27, 2026.
English (US) · 