I’ll always remember the time my friend, who’s in her early 20s, showed me a video of what she believed was Vladimir Putin working out at the gym.
Now, most of us wouldn’t fall for a Putin workout video. But with generative artificial intelligence (AI) becoming mainstream and advanced, it’s going to be more difficult to tell what is and isn’t AI.
Discerning this is important. Especially in a day and age where “we get so much of our information from social media, it’s basically our primary way of understanding what is happening around us or on the far side of the world,” said Siwei Lyu, director of the University at Buffalo’s Media Forensics Lab.
“The world is so connected, yet we are not able to physically present at every important event. We rely so much on audio visual information from social media to tell us what happened. So anybody, if they have the ability to manipulate the images or videos, can influence our decision-making process,” Lyu said.
AI-generated images and videos can be used to influence the outcomes of political campaigns or our day-to-day decision-making about purchasing products, for example, said Lyu. Recently during the Los Angeles wildfires, there were a slew of AI-generated images circulating online that caused misconceptions about what was truly happening on the ground.
What we see online can influence what we believe and how we formulate our opinions, noted Junfeng Yang, a professor of computer science and co-director of the Software Systems Lab at Columbia University.
Lyu and Yang both agreed that we need to regard any photo or video online with the mindset that it could be AI-generated. This skepticism will give us pause before jumping into any false conclusions.
How To Tell If A Photo Or Video Is AI-Generated
A caveat: Generative AI is evolving and improving as we speak. Lyu and Yang cautioned that the telltale signs they are using now to detect AI-generated content are likely to be fixed in the near future. The best way to detect AI is to approach everything you see with some skepticism, carefully inspect what you’re seeing, and investigate its source, both experts said.
Here are what AI experts look for to tell if an image or video is generated by AI, for now. But don’t rely too much on these signs, according to Lyu. A lack of them does not mean the image or video isn’t AI.
Inspect the human hands
Hands, arms and fingers do not look perfect yet in AI-generated content, according to Lyu and Yang.
“For instance, when two people are hugging each other, their arms and hands may be in a very unusual configuration, they may be twisted together, or their arms or fingers are extremely long and so on,” Lyu said.
Pay attention to whether the image respects real world laws like physics
At first glance, an AI-generated image captures the look and feel of a real world scene. But current generative AI models do not have the capability to replicate human anatomy, geometry and the laws of physics in their outputs, Lyu said. Current generative AI models are still unable to truly replicate how objects relate to each other in the real world.
Observe the physical movements of people in the video, Yang advised. Do they look natural, like how human beings in real life would move? Have a look at the lighting that falls on different objects in the image or video. For example, if the light source is shining from the top left of the picture, are the objects correctly illuminated on that side, or are some being illuminated from the right side?
While you’re scanning for lighting inconsistencies, scan also for shadows that do not make logical sense. If the light source is shining from the top left of the picture, are the objects’ shadows rightly cast on the opposite side (i.e., bottom right)?
Peer into the eyes
“AI-generated human faces look highly realistic, but one way to tell they’re AI-generated is to actually look into their eyes,” Lyu explained. “The human pupil is a circular shape, almost like a perfect circle, but AI-generated faces tend to have irregularly shaped pupils.”
Examine the mouth and teeth
You’ve probably seen lip-synced AI-generated videos, where they match a person’s mouth to an audio clip they didn’t originally say. If you pause the video and drag the progress bar along the video timeline slowly to inspect the video one frame at a time, you may see “inconsistencies in the number of teeth, or the shape of particular frontal teeth will change from frame to frame,” Lyu said.
The movement of the person’s lips may also appear unnatural, according to Yang.
“In a real person’s mouth, you’ll also be able to see the tongue and other structures, but for an AI-generated video, when the mouth is open, it’s a dark, empty, vacant space there,” Lyu noted.
Listen for breaths and natural pauses in speech
“In current AI-generated voices, there is a lack of what we call paralingual characteristics,” Lyu said. “These are features of human voices that do not carry meaning, like pauses or sound of breathing.”
As humans, we must first take air into our lungs, before we can expel the air and move our vocal cords to produce sound, Lyu explained. When we hear humans speak, you can hear their intake of breaths in and sometimes the exhales out. In AI-generated voices currently, “they do not have that breathing sound, so you hear a very clean human voice, with almost no background noise,” Lyu added.
Be skeptical of things that look too perfect and lack detail
“If you see an image or a video, that’s too perfect or too surreal, that may be an indication of some AI generation going on,” said Yang.
According to Yang, generative AI models today still face challenges with replicating all the imperfect details and realistic texture of the objects and background in the image or video.
For example, a human face or animal’s fur may seem too smooth and almost airbrushed.
Scan for repetitiveness and duplication in the background
“Sometimes in video or image, you see repetitiveness, like this background person right, gets copied like 20 times, and they all look similar,” Yang said. It may be a background person that is duplicated multiple times to create a crowd, or even a beach without much texture variation among the sand.
Observe the blinking of eyes if it’s a video of a person
Carefully watch how a person blinks in the video, says Yang. A real life person does not blink at a set rate. If the blinking pattern is repetitious, such as a blink every 10 seconds, it may be AI-generated.
Verify the source of the image or video
As AI-generated photos and videos become more and more realistic, and as the above “quirks” get fixed over time, Yang emphasised the importance of learning how to verify the source of the content.
Trace the original source of the photo or video. Is it a credible one? If the video came from a social media platform like TikTok, read the caption, hashtags and the account profile description for any mention of AI. Finally, check if the message being conveyed in the photo or video aligns with reports from reputable news outlets.