OpenAI can accurately detect images from its DALL-E 3 generator 98% of the time but struggles with other AI generators, especially if images are extensively modified.
Minor modifications like cropping, compression, and saturation changes have a success rate of 95-97%, while adjusting hue drops it to 82%.
The toolset struggles to classify images with more extensive modifications, with OpenAI not even publishing success rates for these cases.
AI-generated images, frequently used in online misinformation during election years, pose a challenge as these images are often heavily modified post-generation.
[
add
]
[
|
|
...
]