A truly cursed game of spotting differences.

Individuals can differentiate between actual and AI-generated images just 62% of the time, according to a study conducted by Microsoft’s AI for Good lab. This statistic comes from an extensive evaluation involving 287,000 images assessed by 12,500 participants. Each participant reviewed around 22 images, drawn from a set that contained both 350 copyright-free photos and 700 AI-generated images utilizing tools like DallE-3 and Stable Diffusion.

If you’re curious about how well you can identify the real from the artificial, you can try your hand at this quiz.

The study also evaluated Microsoft’s in-progress AI detector, which achieved over 95% accuracy. Despite the findings, the study calls for increased transparency concerning generative AI, emphasizing the need for content credentials and watermarking to better inform the public on media authenticity. Nonetheless, many argue that watermarks can easily be circumvented.

In developing AI detection tools, it remains vital to address not just detection itself but also the environmental impact of AI technology, which has raised serious concerns over energy consumption and greenhouse gas emissions.