New AI Creates Realistic Eye-Tracking Data to Protect Privacy and Cut Costs

Eye-tracking technology watches where people look, and it's used for everything from designing websites to studying how we read or drive. But collecting this data is a pain: you need specialized hardware, a controlled lab, and volunteers who stare at screens while their every glance is recorded. Even then, the resulting data is sensitive personal information, so sharing it with other researchers is risky. Because of these hurdles, progress in fields that rely on eye movement data can be slow.

Now, a team of researchers has found a clever workaround. They trained a type of AI called a diffusion model — think of it as a pattern learner that can generate brand-new examples that look like the real thing — to produce synthetic eye-tracking data. They gave the AI real eye-gaze recordings from 28 people, and it learned the typical patterns of how eyes jump, pause, and wander. Once trained, the AI could generate unlimited amounts of fake eye-tracking data that statistically mimics real human gaze.

The results are impressive. In tests, the synthetic data was almost interchangeable with real data for certain measures, like how long people fixate on a spot. When another AI was trained only on synthetic data and then tested on real eye tracking, it performed about 83% as well as an AI trained on real data. That tells researchers the fake data captures genuine underlying patterns, while keeping the identities and privacy of the original participants completely out of the picture.

So why should you care? For one, synthetic data could let researchers develop and test new products—like smarter car interfaces or more accessible software—without needing to recruit hundreds of volunteers. It also helps with privacy: researchers could share a 'dataset' of eye movements without exposing any real person's data. While current synthetic data still struggles with long-range behaviors like counting saccades (the quick jumps your eyes make), it's already good enough for many practical uses. The next step is making even bigger, richer datasets that any researcher can use freely.