As generative AI continues to improve at a breakneck pace, it’s becoming increasingly difficult to tell when something you see online is real or fabricated. And it’s not just static images that we have to worry about either. AI-generated videos are on the rise as well, thanks to video generators like OpenAI’s Sora and Google’s Veo. Now, discerning fact from AI-generated fiction is tougher than ever.
So how exactly do you spot videos made by AI? Currently, software-aided detection tools for AI-generated video are extremely scarce (unless you are only focused on deepfakes that have a face visible), so for now, you’ll have to trust yourself instead of a robot.
To help you sharpen your skills and learn what to watch out for, we’ve gathered some Sora video samples from OpenAI (including a few that accompanied the company’s research paper “Video Generation Models as World Simulators.) Using these samples as examples, we’ll show you some of the telltale signs that a video might be AI-generated. Here’s what you should look for.
Does the Video Follow the Laws of Physics?
Unless you’re watching Inception, most filmed things follow the laws of physics. This is not so for Sora. In its paper, OpenAI talks about Sora’s limitations as a simulator, saying “For example, it does not accurately model the physics of many basic interactions, like glass shattering.” Here is a video showing a similar scenario. Glass jumps into the air, with no apparent cause nearby, and the liquid passes through the solid glass, which itself dissolves as it hits the table.
When trying to determine whether a video has been created by AI, you should closely observe all of the phenomena in it, whether it’s the main action, or something taking place in the background.
Also, gauge your feelings while you’re watching. When things behave differently than they should, it’s often uncomfortable for us humans, a phenomenon known as the uncanny valley. So trust your gut. If something seems a bit off—even if you can’t quite put your finger on what exactly it is—take that as a sign that the video deserves some additional scrutiny.
Get Our Best Stories!
Get the Most Out of Your Technology
By clicking Sign Me Up, you confirm you are 16+ and agree to our Terms of Use and Privacy Policy.
Thanks for signing up!
Your subscription has been confirmed. Keep an eye on your inbox!
How Does the World React?
Bite into an apple in real life and a chunk disappears. That’s not necessarily how it works in Sora’s world. OpenAI notes that “interactions, like eating food, do not always yield correct changes in object state.” So while you should be monitoring actions to tell whether something is AI, it’s just as important to track the reactions as well. A lot happens in the video below, but where the man walks through the sand, there are existing footprints behind him, but no new prints are created.
Does Something Just Feel Off?
Pay careful attention to a video created by AI and you’ll likely notice inconsistencies. They might be glaring or just something that tugs at the corners of your thoughts. If it’s the latter, note where you start to sense that something is off and re-watch it closely. In the video below, a hand paints a cherry blossom tree. While the painting develops along with the brushstrokes, there are moments when the paint from the brush changes color despite the paintbrush never leaving the canvas.
Recommended by Our Editors
Can You Recreate the Videos Elsewhere?
In an interview with the Wall Street Journal’s Joanna Stern, now former OpenAI CTO Mira Murati repeatedly would not say what Sora was trained on beyond “publicly available and licensed data.” Putting aside the ethical and potential legal repercussions of this, it means that you can attempt to reverse engineer a video to see if it’s real or not.
Nick St. Pierre, a creative director and fan of Midjourney, followed a feeling he had about the sources of Sora’s work. He came up with prompts that suited Sora videos he’d seen and put them through Midjourney to generate AI still images. Sure enough, he came up with over a dozen examples where Sora’s videos are not much more than Midjourney images in motion. If you want to use more than your senses when assessing a video, come up with the prompt you believe it took to create them. Now, put it through a text-to-image generator to see how similar they are and a Google search to locate sources that might have been used for the video.
This Tweet is currently unavailable. It might be loading or has been removed.
Stay Vigilant!
Text-to-video AI is still very much in its infancy right now, and since the technology is constantly evolving, the best way to avoid being fooled is to evolve along with it. As AI progresses and generates increasingly realistic footage, we’ll need to adopt new tools and techniques in order to effectively spot it. So don’t get comfortable just yet! Keep an eye out for new detection tools and authenticity verification systems, and be sure to check back on this article periodically—we’ll update it as new tools emerge.