In the age of advanced technology, distinguishing between real and fake content has become increasingly challenging. The rise of AI “deepfake” technology has made it difficult to discern the authenticity of photos and videos circulating online. However, there are simple methods to identify the flaws in deepfake content.
According to American tech commentator William Antonelli, even the most sophisticated AI models struggle to maintain details without errors in long-duration footage. When determining if a video is AI-generated, there are common features to look out for. These include blurry or distorted outlines of people or objects, as well as unnatural changes or misplacements of background details after object movements, especially in complex scenes or large crowds.
Antonelli illustrated with an example of an AI-generated video showing rabbits jumping on a trampoline captured by a surveillance camera. While the video initially appears realistic, upon closer inspection, discrepancies such as disappearing rabbits during movements and color blending between them are noticeable. These anomalies indicate the video’s artificial origin.
The quality of AI-generated videos is often lower, with creators using low resolution to hide physical inaccuracies. Many of these videos may falsely claim to be surveillance footage to mask the blurry visuals. Even videos allegedly filmed with real cameras can reveal flaws. For instance, a popular video featuring a chance encounter on a subway, filmed with a modern smartphone, would not typically exhibit such excessive blurriness in normal lighting conditions.
When dealing with text in visuals, AI often makes errors like letter blending, inconsistent fonts, or meaningless typos. Pay attention to details like electronic clocks or ticking clocks in the frame to assess the sequence of digits and the appropriate passage of time for authenticity.
Although AI can generate texts, images, and audio, issues arise when handling multiple content types simultaneously. Therefore, audio in AI-generated videos often serves as a crucial clue in identifying their authenticity. Unnatural features in audio include strange echoes, fast speech pace, inconsistent accents, and vague sentences, sometimes bearing the earmarks of AI-generated text.
In the realm of celebrity videos, common sense is key when assessing authenticity. Recently, an AI-generated video portrayed Brad Pitt and Tom Cruise engaging in a fight, with convincing visuals and sound. However, a segment in the video depicting Pitt’s AI counterpart defending sex offender Jeffrey Epstein clearly contradicts the actors’ public personas, rendering the video’s narrative implausible.
If you encounter a suspicious video, scrutinize the account posting it for credibility. Newly established accounts rapidly sharing content seeking virality, posting repetitive themes, or exclusively focusing on political issues may indicate automated operation by bots. Accounts repeatedly posting similar content with minor alterations, such as changing camera angles or animal colors, are likely to feature AI-generated content.
Many AI models train on a vast database of influencer photos, resulting in portrait images that appear “unnaturally perfect.” The subjects often exhibit overly smooth skin, virtually invisible pores, impeccable makeup, and precise hair styling akin to professional studio shots.
One prevalent flaw in AI creations is the excessive use of background blur effects, combined with overly polished lighting on human subjects, which can evoke a synthetic feel. Notably, if the background is unnaturally blurred while the subject remains overly sharp, it indicates a telltale sign of artificial manipulation.
AI-generated content often exposes flaws in background details, such as blurry facial features, overlapping elements, or abnormal limb structures. In scenes with multiple individuals, the crowd may lack distinct details and appear fuzzy. Furthermore, architectural distortions, irregular window arrangements, and distorted perspective relations are common spatial issues found in AI-generated environments.
This article references reporting from “Reader’s Digest.”