Let's be clear. This isn't evidence. AI videos are not more likely to look bad. The best AI tools can deliver beautiful, polished clips. And low-quality clips aren't necessarily made by AI, either. "If you see something that's really low quality that doesn't mean it's fake. It doesn't mean anything nefarious," says Matthew Stamm, a professor and head of the Multimedia and Information Security Lab at Drexel University, US.

Instead, the point is that blurry, pixelated AI videos are the ones that are more likely to trick you, at least for right now. It's a sign you may want to take a closer look at what you're watching.

AI still adds distortions to videos, but they're getting harder to spot. When a clip is low quality, you're more likely to miss red flags (Credit: Serenity Strull/ Getty Images)

"The leading text-to-video generators like [Google's] Veo and OpenAI's Sora still produce small inconsistencies," Farid says. "But it's not six fingers or garbled text. It's more subtle than that."

Even today's most advanced models often introduce problems such as uncannily smooth skin textures, weird or shifting patterns in hair and clothing, or small background objects that move in impossible or unrealistic ways. It's all easy to miss, but the clearer the picture is, the more likely you are to see those tell-tale AI errors. 

That's what makes lower-quality videos so seductive. When you ask the AI for something that looks like it was shot on an old phone or security camera, for example, it can hide the artefacts that might otherwise tip people off.

Over the past few months, a few high-profile AI videos fooled huge numbers of people. They all had something in common. A fake but delightful video of wild bunnies jumping on a trampoline got over 240 million views on TikTok. Millions of other online romantics hit the like button on a clip of two people falling in love on the New York subway, only to face the same disappointment when it turned out to be a fake. I personally fell for a viral video of an American priest at a conservative church giving a surprisingly leftist sermon. "Billionaires are the only minority we should be scared of," he bellows in a southern accent. "They have the power to destroy this country!" I was stunned. Have our political boundaries really grown that blurry? Nope. Just more AI.

The other two factors, resolution and quality, are related but different. Resolution refers to the number or size of pixels in an image, while compression is a process that reduces the size of a video file by throwing away detail, often leaving behind blocky patterns and blurred edges.

In fact, Farid says low-quality fakes are so compelling that the bad guys downgrade their work on purpose. "If I'm trying to fool people, what do I do? I generate my fake video, then I reduce the resolution so you can still see it, but you can make out all the little details. And then I add compression that further obfuscates any possible artefacts," Farid says. "It's a common technique."

Low resolution images have fewer pixels, while compression adds other errors. Both hide artefacts that can make AI's work more obvious (Credit: Serenity Strull/ Getty Images)

The trouble is that, as you read this, the tech giants are spending billions of dollars to make AI even more realistic. "I have some bad news to deliver. If those visual tells are here now, they won't be very soon," Stamm says. "I would anticipate that these visual cues are going to be gone from video within two years, at least the obvious ones, because they've pretty much evaporated from AI-generated images already. You just can't trust your eyes."

That doesn't mean the truth is a lost cause. When researchers like Farid and Stamm are verifying a piece of content, they have more advanced techniques at their disposal. "When you generate or modify a video, it leaves behind little statistical traces that our eyes can't see, like fingerprints at a crime scene," Stamm says. "We're seeing the emergence of techniques that can help look for and expose these fingerprints." Sometimes the distribution of pixels in a fake video might be different than a real one, for example, but factors like these aren't foolproof.  

Technology companies are also working on new standards to verify digital information. Essentially, cameras could embed information in the file the moment they create an image to help prove that it's real. By the same token, AI tools could automatically add similar details to their videos and images to prove they're fake. Stamm and others say these efforts could help.

The real solution, according to digital literacy expert Mike Caulfield, is for us all to start thinking differently about what we see online. Looking for the clues AI leaves behind isn't "durable" advice because those clues keep changing, he says. Instead, Caulfield says we have to abandon the idea that videos or images mean anything whatsoever out of context. 

"My perspective is that largely video is going to become somewhat like text, long term, where provenance [the origin of the video], not surface features, will be most key, and we might as well prepare for that," Caulfield says.