Computer scientists at the University at Buffalo have developed an AI tool that can detect a deepfake photo by analyzing the light reflections in the eyes
THE METHOD HAS PROVEN ACCURATE ON PORTRAIT-STYLE DEEPFAKES — Deepfakes are being used for a range of nefarious purposes …
Context & Ripple Effects
The University at Buffalo's eye-reflection detector lands mid-way through a detection arms race that began with the 2019 Deepfake Detection Challenge backed by Facebook and Microsoft. Academic teams have since chased different forensic signals: Facebook partnered with Michigan State on reverse-engineering which ML model made a fake (model attribution), and Intel later shipped FakeCatcher reading blood-flow signals from video pixels.
What distinguishes this tool is its target — still portrait photos, where the corneal highlight must match between both eyes if the image is real — and its timing: it arrives just months before researchers exposed an AI service making nonconsensual face-swapped imagery trivially easy to generate.
First-order effects
- Investigators, platforms, and fact-checkers gain a low-cost forensic check for portrait-style fake photos — a modality most prior detectors like Intel's video-focused FakeCatcher did not address.
Second-order effects
- A growing market of commercial deepfake-detection startups will absorb techniques like this, even as their accuracy claims outrun independent testing — pushing buyers to demand verified benchmarks before licensing.
Third-order effects
- Because any single artifact-based signal (eye highlights included) is something generative models can eventually learn to fake, the pattern points toward a layered trust stack where detection is one layer and capture-time provenance becomes the durable anchor.
The trend: Synthetic-media defense is splitting into specialized forensic detectors per artifact type while the industry hedges toward provenance standards that don't depend on spotting artifacts at all.