A light-powered AI checks 15 videos at once for deepfakes, with about 98% accuracy
UCLA engineers built a detector that sends digitally encoded light through free space to an optical decoder, plus a small computer program, to judge whether many videos are real or AI-faked at the same time.
What the scientists did
Deepfakes are videos where AI has swapped or invented someone’s face. Checking huge numbers of videos with ordinary computer chips takes a lot of energy, so the team tried doing part of the work with light instead of electronics.
First, a small digital program pulls out clues from each video, such as patterns in space, color and time. Those clues are encoded as light patterns on a screen-like device. In the experimental device, that light then travels through free space (open air, with no extra layers) to light detectors, and the way it spreads out does the rest of the “thinking.” The detectors give each video a real-or-fake score. One pass of light handles 15 videos side by side. The team separately tested, mostly in simulation, a version that adds specially designed see-through (diffractive) layers in the light path.
Words to know
Deepfake
A video, image or audio clip that AI has created or changed to show something that never happened.
- A face swapped onto another person’s body
- A clip made entirely by a text-to-video AI
Optical computing
Doing calculations with light as it passes through specially shaped materials, instead of with electricity in a chip.
Sensitivity and specificity
Two ways to grade a detector. Sensitivity is how many fakes it catches. Specificity is how many real videos it correctly leaves alone.
- 99.86% sensitivity: almost every fake caught
- 95.72% specificity: about 4 in 100 real videos wrongly flagged
What they found
In real lab experiments on a standard deepfake test set (Celeb-DF), the free-space device was 97.79% accurate while checking 15 videos at once. It caught 99.86% of fakes and correctly passed 95.72% of real videos. The model was trained in a computer and then moved to the physical light setup without any retuning.
The optical decoding step was estimated (not measured) to use about 1.38 to 4.11 millijoules of energy per video, compared with about 30.4 millijoules for the decoder of an all-digital system of similar accuracy. Both designs also need the same digital encoding step, so the estimated whole-system totals are about 180 to 183 millijoules versus about 209 millijoules per video: a real saving, but much smaller than the decoder-only comparison suggests. The team also tested other kinds of fakes, including videos made by a text-to-video AI, and squeezed in 18 videos at once, with somewhat lower accuracy.
Why it matters to you
You probably scroll past dozens of videos a day, and some are fake. Fast, low-energy detectors like this could someday help social apps and news sites screen huge numbers of videos before they spread, so you’re less likely to be fooled by a fake clip of a celebrity, politician or classmate.
Where this could lead
Faster fake-video screening
Platforms could check many video streams at once using less electricity.
Few years, if it can be built at scaleLow-power AI hardware
The same light-based approach could speed up other AI tasks that sort images or video.
Longer termThese are possibilities the research points toward, not promises. Most early findings take years of testing before they reach everyday life, and some never do.
Why scientists care
AI is using more and more electricity. Scientists are exploring whether light can do some of the heavy math faster and with less energy, and deepfake detection is a useful real-world test.
Keep in mind
This was a lab setup tested on research datasets, not on the messy mix of videos found online. Videos near the edges of the setup scored worse, and packing in more videos lowered accuracy. The energy figures are the authors’ estimates, not measurements. New deepfake tools appear constantly, so any detector has to keep being updated, and none is perfect.
How far along is the science?
- Peer-reviewedReviewed by independent experts for a journal before publication. Peer review is a quality check, not a grade of evidence strength: confidence still depends on study design, limitations, replication and the wider evidence.
- PreprintA manuscript shared publicly, often before or alongside journal review. Check its separate journal and review status (each story says what we know). Results may change.
- Conference talkPresented to other scientists at a meeting. Usually short and early; a full paper may come later.
- Agency reportPublished by a science agency such as NASA, NOAA or NIH. Reviewed internally, but not by an outside journal.
- Press releaseAn announcement from a university or company. Useful, but always check the research it describes.
How soon could it reach you?
- Could reach you soonCould show up in products, services or advice soon. Our rough editorial guess is about 1 to 3 years, not a forecast from the research.
- A few years outPromising, but needs more testing or engineering first. Our rough editorial guess is several years, with a lot of uncertainty.
- Long-term scienceBasic research that builds knowledge. Real-world uses, if any, are far off.
- Small stepA small update to earlier work. Useful to researchers, but not a change you would notice yet.
Every story links to its original source: the paper itself, or the conference program when no paper is out yet. Explanations are written with AI, usually from the paper’s abstract (each story says what it was written from), then checked against the source. We only use journals, recognized preprint servers, science agencies, scientific conferences and established science news outlets.