Sorry to shatter your expectations, but if you’re looking for a simple checklist to identify AI-created videos, you’re not going to find one. The era of peculiar AI-generated phenomena like [Will Smith oddly eating spaghetti](https://www.reddit.com/r/StableDiffusion/comments/1244h2c/will_smith_eating_spaghetti/) is over. Nowadays, AI tools can produce persuasive, photorealistic videos in just a matter of clicks.
Currently, AI-generated videos are still relatively primitive compared to their AI-generated text, images, and audio counterparts. Producing realistic videos necessitates a significant amount of high-quality data, which is labor-intensive to compile, according to Siwei Lyu, a computer science and engineering professor at the University at Buffalo SUNY. Nevertheless, there are no fundamental obstacles to enhancing the quality of AI-generated videos, suggesting that we can anticipate significant improvements soon. As these videos advance, the conspicuous flaws—such as deformed faces or morphing objects—are likely to vanish. The secret to recognizing AI-generated videos, or any content created by AI, lies in AI literacy. “Realizing that [AI technologies] are progressing and keeping the mindset that ‘what I’m witnessing could be AI-generated’ is more crucial than zeroing in on specific indicators,” states Lyu, who also oversees UB’s Media Forensic Lab.
### Navigating the AI-Loaded Web
In an increasingly AI-saturated environment, your best strategy is to enhance your online discernment and judgment. Spotting when something seems off is vital to steering clear of deepfakes, misinformation, or mere low-quality AI content. While mastering this skill can be challenging—particularly in an online realm that relentlessly competes for your focus—it is certainly feasible to hone your instincts.
“By examining [AI-generated images], we believe individuals can boost their AI literacy,” explains Negar Kamali, an AI research scientist at Northwestern University’s Kellogg School of Management. Kamali, who has co-written a [guide](https://arxiv.org/pdf/2406.08651) on identifying AI-generated images, adds, “Even without spotting any artifacts [that suggest AI-generation], my mind instantly thinks, ‘Oh, something isn’t right.’ Even if I can’t identify the flaw, I can’t assert that it’s authentic—and that’s our goal.”
### Classes of AI-Generated Videos: Imposter vs. Text-to-Image
Before we explore how to detect AI-generated videos, it’s essential to differentiate between the two primary categories: imposter videos and text-to-image diffusion model videos.
#### Imposter Videos
Imposter videos utilize AI-editing techniques such as face-swapping—where one individual’s face is substituted with another’s (often a celebrity or political figure)—and lip-syncing, where the person’s lips are altered to synchronize with different audio. These videos tend to be more convincing since they rely on existing footage rather than creating something from scratch. Remember those viral Tom Cruise deepfake videos? They were remarkably realistic because the creator, Chris Ume, collaborated with a professional Tom Cruise impersonator and conducted extensive editing, as outlined in [this interview](https://www.theverge.com/2021/3/5/22314980/tom-cruise-deepfake-tiktok-videos-ai-impersonator-chris-ume-miles-fisher) with *The Verge*. Today, numerous applications can replicate similar effects, even generating audio from short snippets found online.
If you suspect that a video could be an AI deepfake, begin by analyzing the format. AI deepfakes are frequently presented in a “talking-head” style, where only the speaker’s head and shoulders are visible, with their arms concealed. To identify face swaps, watch for artifacts around the face, especially during head movements at angles. Digital forensics expert Hany Farid, a professor at UC Berkeley, remarks that “you usually notice artifacts when the head shifts obliquely to the camera.” Additionally, take note of unnatural body movements—if the person’s arms seem rigid or their body hardly moves, it may indicate a fake.
For lip-syncing deepfakes, pay attention to the subject’s mouth, and particularly their teeth. In some instances, the tooth count may fluctuate, or the teeth might appear oddly shaped. Lyu further highlights that the lower half of the face may “wobble” or seem rubbery due to minor misalignments that occur during editing.
#### Text-to-Image Videos
The newer, more glamorous type of AI-generated video derives from text-to-image diffusion models, which produce videos based on text or image prompts. OpenAI’s recent showcase of its video generator, [Sora](https://mashable.com/article/openai-sora-dropped-trippy-music-video-fan-ai-hype-flames), amazed viewers with its phot