Identifying 9 techniques for recognizing computer-generated viral video content
In the era of sophisticated AI video technology, it can be challenging to distinguish synthetic content from real footage on social media platforms. To help identify AI-generated videos, look for subtle visual and temporal inconsistencies beyond just facial cues.
Key signs to watch out for include:
- Visual glitches during motion transitions, such as freezing, melting, morphing, or repetitive backgrounds. Pausing the video at random points can reveal these inconsistencies.
- Anatomical oddities like unnatural extremities, wrong or floaty fingers, asymmetrical or glitchy blinking, weird or frozen smiles, and glassy or flickering eyes.
- Audio mismatches, such as lip-sync issues or robotic speech tones.
- Inconsistent scene details, such as objects vanishing, changing features, or unnatural object placement within backgrounds.
- Generic or contextually odd scenarios where no credible source is available.
- Sometimes, specific AI generation methods leave telltale artifacts like warped edges, ghosting, or unrealistic physics.
Recent advances have led to the development of universal detection systems like Google's UNITE. This tool uses transformer-based deep learning models to analyze whole scenes for spatial and temporal inconsistencies, not just faces. UNITE examines multiple regions per frame, detecting both manipulations of faces and fully synthetic content, even without people in the video.
To spot AI-generated viral videos on social platforms:
- Look beyond faces—examine backgrounds, object interactions, and motion flow for unnatural distortions or repeated patterns.
- Notice subtle glitches that occur when motions transition or objects appear/disappear.
- Check for inconsistencies in audio-visual synchronization.
- Be wary of videos with no identifiable credible source or overly generic content.
- Familiarize yourself with common AI-video generation weaknesses (e.g., unnatural head turns or ghosting) for specific generation types.
In addition, keep an eye out for:
- Legible writing issues, such as warped letters, random symbols, or gibberish text on clothing, signage, or packaging in the frame.
- Videos that are 10 seconds long or consist of very short clips stitched together, as these may provide AI less opportunity to reveal its flaws.
- Synthetic clips often have strangely clean audio, mismatched ambient noise, or none at all.
While AI has made significant strides, it's important to remember that it still struggles with legible writing and may not always perfectly mimic real-world motion, which often obeys physical laws. As such, pairing human scrutiny with emergent AI detection tools can help identify AI-generated video content spreading on social media.
It's also worth noting that AI research scientist Negar Kamali from Northwestern University's Kellogg School of Management stated that it's not always possible to definitively identify AI-generated videos. Therefore, maintaining a critical and discerning mindset when consuming social media content is essential.
References:
- Google's UNITE
- OpenAI's Sora
- Google Veo 3
- SynthID by Google DeepMind
- Aruna Sankaranarayanan, research assistant at MIT's Computer Science and Artificial Intelligence Laboratory
- Hany Farid, UC Berkeley professor of computer science and digital forensics expert
- Negar Kamali, AI research scientist at Northwestern University's Kellogg School of Management
- The development of advanced AI video technology has led to the creation of tools like Google's UNITE, which employs transformer-based deep learning models to examine whole scenes for spatial and temporal inconsistencies.
- To identify AI-generated viral videos on social media, one should not limit their analysis to facial cues, but also examine backgrounds, object interactions, and motion flow for unnatural distortions or repeated patterns.
- Synthetic clips may have legible writing issues, such as warped letters, random symbols, or gibberish text on clothing, signage, or packaging, which can serve as telltale signs of AI-generated content.
- Subtle glitches that occur when motions transition or objects appear/disappear can be key indicators of AI-generated videos, and these inconsistencies can often be revealed when pausing the video at random points.
- Tech giants like Google, OpenAI, and Google DeepMind, along with researchers such as Hany Farid and Negar Kamali, are continuously working to advance AI detection systems and reveal the flaws of AI-generated video content.
