New AI Invention Uses Light to Catch Deepfakes — 15 Videos at Once With 97.8% Accuracy

Introduction

Deepfakes are getting cheaper to make and harder to spot, while the AI systems that detect them burn huge amounts of computing power. Researchers at UCLA in Los Angeles, California, have come up with a different answer: an AI system that uses light to do the heavy lifting. Published in the journal eLight and led by Professor Aydogan Ozcan, the new optical deepfake detector checks 15 or more videos at once in a single pass of light, with an average accuracy of 97.79%.

This guide explains this new AI invention in plain English — how light-powered AI works, why it matters for deepfake fraud, and what it means for businesses in California — from Los Angeles media studios to San Francisco fintechs — that rely on video, identity checks and online trust.

The Problem: Deepfake Detection Is Expensive at Scale

Platforms, banks and newsrooms now face a flood of AI-generated video. Conventional detectors run on GPUs and can need hundreds of billions of calculations to analyse a single clip. Checking every upload, video call or identity selfie that way is slow and costly — so in practice, most content is never checked at all.

At the same time, deepfakes are moving into fraud: fake CEOs on video calls, synthetic identities in onboarding and manipulated evidence in claims. Businesses need detection that is fast, cheap and always on.

Why It Matters

AI's energy use is now a business and policy issue. Data centres are straining power grids, and every extra GPU workload adds cost. Optical computing — using light instead of electricity to do maths — promises much faster, far more energy-efficient AI for certain tasks.

The UCLA work is one of the clearest real-world examples yet: a practical security task, solved with a hybrid of a small digital model and passive optics. If the approach scales, it could make screening every video stream affordable.

How Light-Powered Deepfake Detection Works

1. A small digital encoder. A lightweight AI model extracts compact spatial, spectral and timing features from each video and turns them into a pattern displayed on a programmable spatial light modulator.

2. A passive optical decoder. Light passes through specially designed diffractive layers. As it bends and interferes, it performs the final calculation — with no extra electrical power needed during inference.

3. Many videos in parallel. Paired optical detectors read out an authenticity score for each video. The team tested 15 videos per pass, and 18 at once with 96.13% accuracy.

4. First-stage screening. The researchers describe it as "a highly sensitive first stage of a larger detection system" — suspicious videos are passed to digital models for deeper analysis.

Benefits

• Massive parallelism — many videos analysed in one optical pass instead of one by one.
• Lower energy for the optical stage, since diffractive layers compute without extra power.
• Very few missed fakes, with sensitivity near 99.9% and a false-negative rate around 0.14%.
• Built-in robustness — key parameters live in physical hardware, which the team reports makes it resistant to adversarial attacks, noise, blur and compression.
• A blueprint for green AI that could apply to other high-volume screening tasks.

Real-World Examples

Content moderation: a social platform could screen every uploaded video optically and send only flagged clips to expensive GPU models and human reviewers.

Banking and fintech: a San Francisco fintech verifying customers by video selfie could add a high-speed deepfake screen at onboarding — the same threat we covered in our guide to AI payment fraud and deepfakes.

Newsrooms and media: Los Angeles studios and broadcasters could check incoming user-submitted footage in real time before it goes on air.

Why This Matters for California Businesses

California is both where much of today's generative AI is built and one of the most active states in regulating it. The California AI Transparency Act (SB 942) requires large generative AI providers to offer AI-detection tools and disclose AI-generated content, and the state has passed several laws targeting deceptive deepfakes. For California platforms, studios, banks and marketplaces, reliable deepfake screening is moving from a nice-to-have to something users and regulators expect.

Common Mistakes

• Trusting video as proof. A face and a voice on a call are no longer enough to approve a payment or verify identity.
• Relying on one detector. Every detector can be fooled by new generators; layered checks are safer.
• Ignoring false positives. Blocking real customers damages trust — always have a human review path.
• Waiting for perfect technology. Process controls like call-backs and dual approval work today.

Best Practices for California Businesses

• Layer your defences: fast screening first, deeper AI analysis second, human review last.
• Add liveness and identity checks to onboarding and high-value transactions.
• Verify out of band: confirm unusual requests through a known phone number or second channel.
• Log and audit every verification decision for compliance and dispute handling.
• Watch energy-efficient AI — optical and edge computing will cut the cost of always-on AI.

Most businesses won't build optical chips, but they can build smarter verification into their software today. As an AI development company in California, we integrate deepfake detection, liveness checks and risk scoring into onboarding, payment and content workflows, and our cyber security and custom software development company in California teams build the audit trails and approval flows around them. Find your best starting point with our free AI Readiness Assessment.

Conclusion

UCLA's light-powered deepfake detector is a glimpse of where AI hardware is heading: hybrid systems where a small digital model does the setup and light does the heavy computation, screening many video streams at once with very high sensitivity. It is still research, but it shows deepfake detection can become fast and affordable enough to run everywhere. Until then, businesses should combine AI detection with strong verification processes and human oversight.