Barely better than "chance": Microsoft reveals just how bad we are at detecting AI images

A new research paper from Microsoft AI for Good outlining an experiment consisting of over 12,500 global participants and 287,000 image evaluations revealed an overall success rate of just 62% in telling apart AI-generated images from real ones. This indicates that humans only have a modest ability of seeing through these fake images, with success just above chance.

It was found that participants could detect when human portraits were fake with the most ease but struggled significantly when it came to natural and urban landscapes with success rates dropping to 59-61%. The low scores highlight the challenges humans face when trying to distinguish AI images, especially those without obvious artifacts or stylistic cues.

In the study, participants had to play a Real or Not Quiz game where they were shown AI images they’d likely run across online in reality; those running the study avoided cherry picking images or only choosing highly deceptive images. It was also noted that AI is continually improving so future models may produce even more convincing images.

Based on the results of the study, Microsoft is calling for transparency tools such as watermarks and robust AI detection tools to remove the risks of misinformation arising from AI-generated content. To help educate people about these dangers, the Redmond giant previously launched a campaign addressing AI-generated misinformation.

The researchers also had access to their own AI detection tool and it was able to get a success rate above 95% on both real and AI-generated images across categories - this suggests that machine assistance is quite a lot more reliable that human judgment, but even it isn’t perfect.

It’s also important to point out that even if you have a visible watermark on the image in the corner, malicious actors looking to dupe people with fake images can easily crop this out or make it harder to see using rudimentary tools.

The researchers noted that humans may have found it easier to detect AI images of faces because of our innate ability to identify faces well and we can most likely spot abnormalities in the AI portraits. Interestingly the research found that older generative adversarial networks (GANs) and inpainting techniques were quite good at fooling users as they produce images that look like amateur photography rather than with a studio-like aesthetic used by popular models like Midjourney and DALL-E 3.

Inpainting is an interesting technique that replaces a small element of a real picture with something AI-generated. Microsoft noted that this makes forgery extremely difficult to identify and poses a significant risk for disinformation campaigns.

This study highlights just how susceptible humans are to getting tricked by artificial intelligence and reminds of the need for tech firms to develop technologies to try to prevent the malicious spreading of such images.

Source: ArXiv | Image via Depositphotos.com