Light Analysis Flags Deepfake Videos at 99.86%

The UCLA method remains research-stage, but it targets a detection gap where human viewers catch deepfakes only half the time.

Updated on Nov. 3, 2026 in Artificial Intelligence

Bold flat-color editorial illustration of a camera lens with a cyan beam of light raking across a cream surface, evoking light analysis. — UCLA researchers have developed a light-based method that flags deepfake videos with 99.86% accuracy, though the technique remains at the research stage. AI Illustration. Upload story photo >

Live Poll

Do you trust that videos you see online are real?

UCLA researchers have developed a technique that detects deepfake videos by analyzing the light present in a video. The method identifies manipulated videos with 99.86% accuracy and real videos with 95.72% accuracy, and it processes 15 videos at a time. The development is reported as research-stage; no shipped product is described in the source.

Why it matters

Human viewers identify deepfake videos only about 50% of the time, according to Nature Communications research, making automated detection a growing need. A detection approach built on light analysis points to a different signal than the visual artifacts most current tools examine.

The UCLA system reports 99.86% accuracy on manipulated videos and 95.72% on real ones, and processes 15 videos at a time. Human observers in Nature Communications research identified deepfakes only half the time. The source does not specify the benchmark dataset behind these figures.

The players

University of California, Los Angeles (UCLA)

A public research university whose researchers developed the light-based deepfake detection technique reported in this story.

The details

The technique works by analyzing the light present in a video to determine whether the footage is authentic or a deepfake. A deepfake is synthetic or altered media generated by machine learning models to show someone doing or saying something they did not. Because the method examines lighting characteristics rather than the facial content a generator tries to fake, it targets a different class of signal than human viewers rely on. The system handles 15 videos at a time, according to the source.

Article on the UCLA deepfake light-detection technique published November 3, 2026.

The Tech Race

The UCLA result sits in the research effort to build automated deepfake detectors that exceed human judgment, where human accuracy of about 50% sets the baseline. The source does not name competing detection programs or prior milestones for comparison.

The source does not indicate when or whether this detection technique will be available to the public, to platforms, or to newsrooms. It remains a research result, so no current tool is available for users to apply to their own videos.

The takeaway

Watch for peer-reviewed publication and independent testing of the 99.86% and 95.72% figures on datasets beyond the original study. Readers should treat the result as promising research until a deployable tool is documented.