Fun with Data – Fool Me Once
AI image detection is a difficult task and one where there is a clear leader, and it is not human. Microsoft’s (MSFT) AI Lab for Good did a study as part of an AI image recognition development project that was both telling and not surprising. What sets this research apart is that it relies on a large sample of human evaluations through an online game. To date, the game has logged more than 82,500 distinct sessions. For the analysis below however the data comes from those that played between August 1st and August 8th 2024. This correlates to12,500 participants around the world, all of whom completed the full game, which represent 17,340 unique games, producing roughly 287,000 image evaluations.
The images displayed in the game come from the collection of 350 copy-right free ‘real’ images and the generation of around 700 diffusion-based images using DallE-3, Stable diffusion-3, Stable diffusion XL, Stable diffusion XL inpaintings[1], Amazon Titan v1 and Midjourney v6. Also added were GAN-based fake faces. Each user is shown one image at a time, for 15 rounds. The 15 images in a game are randomly selected from all images, both AI and real. The statistics are derived only from the results of completed games - in which all 15 images were evaluated by a user.
The game was designed to raise awareness about AI-generated images and how challenging it can be to spot them, and Microsoft could have cherry-picked AI images that can fool most people most of the time, but the goal was to provide a realistic panorama of the AI images that people are likely to be exposed to, an “average” output of generative AI.
[1] Inpainting are images that are added to both real and AI images as items that were not in the original image
The images displayed in the game come from the collection of 350 copy-right free ‘real’ images and the generation of around 700 diffusion-based images using DallE-3, Stable diffusion-3, Stable diffusion XL, Stable diffusion XL inpaintings[1], Amazon Titan v1 and Midjourney v6. Also added were GAN-based fake faces. Each user is shown one image at a time, for 15 rounds. The 15 images in a game are randomly selected from all images, both AI and real. The statistics are derived only from the results of completed games - in which all 15 images were evaluated by a user.
The game was designed to raise awareness about AI-generated images and how challenging it can be to spot them, and Microsoft could have cherry-picked AI images that can fool most people most of the time, but the goal was to provide a realistic panorama of the AI images that people are likely to be exposed to, an “average” output of generative AI.
[1] Inpainting are images that are added to both real and AI images as items that were not in the original image
Certain types of images are more easily detected by us humans, such as
:
· People 65%
· Vehicles 63%
· Objects 62%
· Urban 61%
· Nature 59%
But when it comes down to really being able to discern which are real and which are not, we humans are not so good, while AI detectors, such as the one Microsoft is developing as part of this study, are. Here’s the data:
:
· People 65%
· Vehicles 63%
· Objects 62%
· Urban 61%
· Nature 59%
But when it comes down to really being able to discern which are real and which are not, we humans are not so good, while AI detectors, such as the one Microsoft is developing as part of this study, are. Here’s the data:
We hate to admit it but we took the test and ended with a 53% score, quite a bit below average, although the system was kind enough to tell us that our score was still better than 18% of the participants. (It did little to make us feel much better). The scary part is that as Generative AI images improve, it will get harder for humans to tell them apart from real images and the necessity for watermarking or other embedded forms of identification becomes greater. The problem however is not with most rank and file content consumers, it is with those interested in disrupting news and media cycles or influencing consumer sentiment, those who are not going to watermark their devious works. It’s a hard problem to solve and one that will continue to get more difficult, but there is no free lunch. Everything in technology has a price.
Here are some of the images used in the survey. You can convince yourself that you are better than average by looking at them, but it means nothing unless you play the game:
https://arxiv.org/html/2507.18640v1
Here are some of the images used in the survey. You can convince yourself that you are better than average by looking at them, but it means nothing unless you play the game:
https://arxiv.org/html/2507.18640v1
RSS Feed