A Light-Powered AI That Spots Deepfakes With Nearly 98% Accuracy
In this post
Deepfake videos are getting better, cheaper and more common, and checking them one by one on power-hungry computers does not scale. A team at UCLA led by Professor Aydogan Ozcan has proposed a different kind of deepfake detection: let light do part of the thinking. Their prototype screens 15 videos at once and flags fakes with 97.79% accuracy.
The study appeared in the journal eLight on 22 September 2026 and drew wide coverage in October.
How a detector can "think" with light
The system is a hybrid of ordinary computing and optics:
- Digital encoder. A conventional program samples frames from each video and extracts key features.
- Light pattern. Those features are converted into a pattern displayed on a programmable device called a spatial light modulator. Several videos are packed into one combined pattern.
- Optical decoder. Laser light passes through thin, passive diffractive layers. As the light bends and interferes, it performs part of the classification math for free, at the speed of light.
- Sensors. Light detectors measure the result and produce a real-or-fake score for each video.
Because step 3 needs no electricity once the layers are made, the approach could cut the energy and time needed for first-pass screening.
The results
| Test | Result |
|---|---|
| Videos per optical pass | 15 (up to 18 shown) |
| Accuracy, 15 videos | 97.79% |
| Fakes correctly caught (sensitivity) | 99.86% |
| Real videos correctly cleared (specificity) | 95.72% |
| Accuracy, 18 videos at once | 96.13% |
| Unseen Google VEO-3 videos, light fine-tuning | 94.80% |
On a set of harder manipulations, adding two optimized diffractive layers raised accuracy by about 6.8 percentage points, from 89.0% to 95.8%, with little extra energy or processing time.
Why it's harder to fool
Many AI detectors can be tricked by adversarial attacks, small hidden changes designed to confuse the model. Here, key parameters are physically built into the diffractive layers. The researchers report that this gives resistance to black-box attacks and raises the bar for attackers who try to study the system directly.
The honest limits
- It's a lab prototype. Deployment would need dedicated optical hardware.
- Accuracy depends on the dataset. The 97.79% figure comes from the Celeb-DF benchmark under specific conditions.
- It's a first filter. The team positions it as a highly sensitive first stage that passes suspicious videos to more detailed checks, not a final judge.
Detection also answers a different question than provenance. Systems like C2PA Content Credentials and Google's SynthID try to label where content came from at the source. A realistic future likely uses both.
How to protect yourself today
Until tools like this reach your phone, a few habits help: check whether trusted outlets report the same clip, look for the original upload, be wary of videos that trigger strong emotion, and look for content credentials or AI labels where platforms provide them, such as YouTube Music's new labels for AI-made songs.
Sources
Frequently asked questions
How does the UCLA deepfake detector work?
It converts video features into a light pattern, then passes laser light through diffractive layers that perform part of the classification optically.
How accurate is it?
97.79% on the Celeb-DF dataset while screening 15 videos at once, with 99.86% of fakes caught.
Can it detect videos from new AI models?
With minimal fine-tuning it reached 94.80% accuracy on unseen VEO-3 videos.
Is it available to the public?
No. It is a research prototype that would need special hardware.
Why use light instead of a GPU?
Light can process many signals in parallel with very little energy, making large-scale screening cheaper.