We live in an era where artificial intelligence can generate videos so visually convincing that they blend seamlessly with footage shot by actual cameras. Text-to-video models like Sora and Veo create coherent motion, realistic lighting, and consistent object behavior across minute-long sequences. The natural assumption has been that this is a solvable detection problem: build better classifiers, train on more examples, identify the telltale artifacts of generation. But what if we’ve been testing detection on the easy cases all along?New Video Reality Test research challenges this assumption by asking a more uncomfortable question: when video and audio are tightly synchronized, as they are in real-world content, can our best AI detection systems actually tell the difference between real and fake? The answer is no, and the reason why reveals something fundamental about how AI perceives authenticity.
Read more