Microsoft’s Study Reveals Humans Are Struggling to Spot AI Images
Readers help support Windows Report. We may get a commission if you buy through our links.
Read our disclosure page to find out how can you help Windows Report sustain the editorial team. Read more
AI-generated images are no longer so easy to recognize. That’s the conclusion of a new Microsoft report, which revealed that the majority of individuals barely noticed the difference between real and AI-generated images, particularly in landscapes and urban areas.
Microsoft’s “AI for Good” team conducted a global quiz-type experiment involving more than 12,500 people. Each was asked to decide if the images were actual or not. In a staggering 287,000 evaluations of images, participants were only correct 62% of the time. Just a bit better than flipping a coin.
Humans were most accurate at detecting fake human portraits. However, in detecting AI-generated images of nature or cities, they only got to about 59% accuracy. Such scenarios don’t tend to feature glaring tells such as creased faces or outlandish textures, so it would be more difficult to mark them as suspicious.
The researchers ensured that the team used a wide variety of images that individuals may, in fact, encounter online. Nothing was selected to be tricky or manipulative. Nevertheless, the findings indicated that even everyday, simple AI images can evade human judgment.
To assist in addressing the dangers of disinformation, Microsoft is now advocating for greater transparency standards. That means they can be seen in the form of visible watermarks and more robust detection mechanisms. But watermarks might not be sufficient. The report says that one can crop or edit them out through simple means.
Notably, Microsoft’s internal AI detector was 95% accurate on the same photographs. That implies that we may have to get machine assistance in order to maintain pace with the velocity and realistic nature of modern generative models. All in all, this research is a reminder that visual deception is becoming more advanced, and we might not always know it’s coming.
