The Plastic World Has Won: 90% of People Can't Tell AI Content from Real

Runway, ahead of the release of their new AI model Gen-4.5, published a very interesting study called The Turing Reel. Over a thousand participants were shown 20 five-second videos: ten real and ten generated. They had to determine which was real and which was AI.

The results make you think. 90% of users could not reliably distinguish generated videos from real ones. Only 99 out of 1043 people (about 9.5%) achieved statistically significant results, correctly identifying at least 15 out of 20 videos. The average accuracy was 57.1%. That's slightly better than flipping a coin or playing Russian roulette. You can also take the test at the link above and check yourself. By the way, my result was mediocre: I only guessed 13 out of 20.

Accuracy varied noticeably depending on the video category. Videos with leather bags people were easier to recognize (accuracy 58-65%). But videos with animals, nature, or architecture were identified much worse — with an accuracy of 45%, meaning generated videos were more often mistaken for real ones.

The plastic world has won: in 2025, we have passed the point of no return. Neither video nor especially images can serve as proof of anything on their own anymore. By default, content should not be trusted until proven otherwise. Presumption of guilt.

This new vulnerability of our world is already being actively exploited: here's a post about a case from my practice where a neural network was used to deceive restaurants with delivery.

"I don't want to live in a world where I first have to make sure the cat isn't generated before I can be charmed by it"


Speaking of cats. For those too lazy to follow the link and take the long test there, here's a test from me 👇