Level 1 — Absolute Beginner
Fake videos are a big problem. They look real, but a computer made them.
Scientists at UCLA built a new tool to find fake videos. The tool uses light.
It can check 15 videos at the same time. It is right almost 98 times out of 100.
The tool also uses very little energy.
- fake
- Not real.
- video
- Moving pictures, often with sound.
- scientist
- A person who studies the world with tests.
- tool
- Something that helps you do a job.
- light
- What lets us see.
- energy
- Power that makes things work.
- computer
- A machine that works with information.
- correct
- Right.
Level 2 — Elementary
Fake videos made with artificial intelligence are getting better every year. Many people cannot tell what is real.
Researchers at UCLA built a new system to find these deepfakes. Part of the work is done by light, not by a normal computer chip.
Most detectors check one video after another. The UCLA system can check 15 or more videos at the same time in one pass of light.
The system is correct almost 98 percent of the time. It also uses very little energy, so it could help websites check huge numbers of videos.
- artificial intelligence
- Computer systems that can learn and make choices.
- deepfake
- A fake video made with AI that looks real.
- detector
- A tool that finds something.
- researcher
- A person who studies a subject carefully.
- system
- A set of parts that work together.
- pass
- One trip of light through a device.
- accurate
- Correct and without mistakes.
- huge
- Very big.
Level 3 — Intermediate
As AI generated video grows more convincing, one of the hardest questions online is simple: is this clip real? Researchers at UCLA say they have a faster and cheaper way to answer it, and it relies on light.
The team, led by Aydogan Ozcan with Parnian Ghapandar Kashani and Shiqi Chen, describes an optical neural system in the journal eLight. A small digital front end prepares each video, and a programmable device called a spatial light modulator then lets light do the heavy math in a single pass.
Because the optics work in parallel, the system can examine 15 or more video streams at once instead of one after another. In tests it detected deepfakes with nearly 98 percent accuracy.
Energy use is tiny. The researchers estimate the optical decoder needs roughly 0.18 to 0.66 millijoules per video. They also say the design resists attacks, which matters because people who make fakes try to fool detectors.
- convincing
- Able to make you believe something.
- optical
- Related to light.
- neural system
- A computer design inspired by the brain.
- front end
- The first stage of a system that prepares the input.
- modulator
- A device that changes a beam of light.
- in parallel
- At the same time rather than one after another.
- decoder
- The part that reads a signal and gives an answer.
- resist
- To stand up against.
Level 4 — Advanced
The arms race between synthetic media and the tools built to expose it has a new entrant, and it runs on photons. Engineers at UCLA have demonstrated an optical neural architecture that screens video for deepfakes with nearly 98 percent accuracy while examining 15 or more streams simultaneously, according to a study in the journal eLight.
The design, by Parnian Ghapandar Kashani, Shiqi Chen and Aydogan Ozcan, is a hybrid. A lightweight digital front end distills each clip into a compact representation, and a spatially multiplexed optical decoder, driven by a programmable spatial light modulator, performs the classification as light propagates, in effect letting physics carry out much of the computation in a single pass.
The payoff is throughput and thrift. Conventional detectors grind through footage sequentially, burning electricity at every step, whereas the optical stage handles many videos in parallel and, by the team's estimate, consumes only about 0.18 to 0.66 millijoules per video at the decoder. The authors also report resilience to adversarial attacks, a vital property when forgers deliberately craft clips to slip past automated screens.
Whether such hardware can leave the laboratory is the open question, but the appeal is obvious for platforms that must triage a deluge of uploads. As generative tools improve, cheaper and faster verification may prove as consequential as the fakes themselves.
- arms race
- A contest in which each side keeps improving to beat the other.
- entrant
- A new participant in a field.
- architecture
- The overall design of a system.
- representation
- A simplified form of data that keeps the key features.
- multiplexed
- Combining several signals in one channel.
- throughput
- The amount processed in a given time.
- adversarial
- Designed to deceive or defeat a system.
- deluge
- An overwhelming flood.
%2C_(analogous_to_X-ray_crystallography).JPG%3Fwidth%3D800&w=3840&q=75)
Talk about this article with KatePractice this story's vocabulary in a real spoken conversation with your AI tutor