When most people want to watch a sports highlight video, breaking news clip, death-defying stunt, cat meme or video du jour, we typically use the internet. In fact, videos make up more than 80% of all internet traffic as people choose it over other types of content.
With the advent of generative AI, computer produced video is expected to drive online video consumption even higher in coming years. In 2023, the AI-generated video market was worth approximately $555 million and is projected to nearly quadruple to $1.96 billion within the next five years, according to Grand View Research.
Even before that type of explosive growth, AI blended videos have caused problems. Deepfake videos are one of the more high-profile challenges produced using generative AI technology. While some of the deepfakes have been nothing more than playful parodies, there have been much more sinister video creations from static images and text using AI resulting in fake footage.
Some of those AI generated videos have been determined by authorities to violate a host of privacy laws, as well as sexual and child abuse statutes and even national security regarding election tampering. This issue has become so important that California recently enacted the nation’s first anti-deepfake legislation, but the constitutionality of those laws is likely to be challenged.
It seems that regulation to shutter AI-generated videos is fuzzy at best for now, and we need to learn what to look for ourselves.
As an extension of my article from Friday regarding detecting AI-generated written content, here are five ways to help tell if the next video snippet you watch was AI made.
Facial and Lip Syncing
At this stage of AI-generated videos, especially when it comes to deepfakes, there can be misalignments between lip movements and the spoken words. It’s akin to watching foreign films with an overdubbed audiotrack in English.
Except in this instance, the actor on the screen and the audio performer are speaking the same language. This mismatch can be a signal that it’s an AI copy.
Blinking and Facial Movements
An interesting aspect of depictions of people in generative AI videos is that they can have a tendency of not blinking naturally. When humans blink in the real world, we do it at irregular intervals, which are also tied to verbal cues in conversations. Additionally their facial movements can be overly smooth or limited—either one is noticeable if you’re looking for it.
AI programming hasn’t mastered these nuances yet—but it will eventually and that’s why AI video detection tools are necessary.
AI-Video Detection Tools
The first tool for consideration is the Deepware scanner, which is an easy-to-use detector that specializes in deepfake videos by analyzing frame-by-frame inconsistencies and micro irregularities in facial expressions.
Another useful resource comes from WeVerify, which offers its nifty InVid Chrome plugin so you can have an AI detector tool box always open on your browser. The plugin won first place in the 2021 US Paris Tech challenge and has been recommended by the non-profit Poynter Institute for journalism ethics.
The reason for sharing more than one AI detection tool is that it’s a best practice to use both tools. You should get in the habit of using the second tool to verify the results of the first.
Background Noise And Distortions
As we saw with AI-generated images, the current batch of generative AI platforms have some difficulty accurately rendering background details and depth of scale.
AI videos often have slight distortions that can appear as pattern noise, flickering or unusual shifts in the background scenery. These constructs are typically due to current limitations with AI’s rendering processes. Expect those to improve over time.
Emotional Incongruence
AI has an uncanny ability to subtly inject stilted dialogue, awkward expressions and unnatural microgestures into its videos. These small missteps cumulatively create a lack of congruence between the emotional tone of the spoken words and the displayed reactions.
Again, the disconnect is subtle but often just enough to let you know that something about the video is not quite right and perhaps not entirely authentic.
The Video is Mightier Than The Pen
The ability to separate true videos from fake videos is critical when you consider the power of video messaging. The brain processes video images 60,000 times faster than text and as a result 95% of a video message tends to be retained compared to just 10% of written text.
If the proliferation of online video — and generative AI video in particular — continues as projected, one day we may be forced into a state where “seeing is believing,” even if what we see is completely AI fabricated.
