IE 11 Not Supported

For optimal browsing, we recommend Chrome, Firefox or Safari browsers.

How accurate is OpenAI at detecting AI-generated images?

Answer: More than 90 percent, but it depends on the image.

Illustration of two white robotic arms typing on a laptop. Light-blue background.
Shutterstock
As artificial intelligence models get better and better at creating their own images, so must our ability to detect such images. That’s why OpenAI, creator of popular AI chatbots including ChatGPT and DALL-E 3, has developed a program to detect when an image was generated by AI.

The system is 98 percent accurate at telling that an image was created by DALL-E 3. Most of the time, that is. If the image went through any minor edits such as cropping or changes in saturation, the accuracy rate declines slightly to 95-97 percent. However, if the hue is adjusted, there is a much more dramatic drop to 82 percent.

Any more extensive changes “can reduce performance,” according to OpenAI. The company did not specify by how much, but there’s a good chance it’s a significant amount. And the tool basically only works on images created by DALL-E 3 — OpenAI noted that only 5-10 percent of images created by other models were accurately identified as AI-generated.