AI Can Draw Unusual Experiences, But Misses How They Felt

Some experiences are hard to put into words — the shapes someone sees during a migraine, or the swirling imagery some people get from flickering light. A new paper called the Perceptual Reality Transformer (PRT) asks a simple question: when AI draws a picture from someone's description, what must that picture preserve to stay true to them?

The answer, it turns out, is more than just the objects in the scene. The researcher studied 3,145 public reports of Ganzflicker experiences — a flickering-light effect that makes some people see vivid patterns. AI handled the basic content well, correctly identifying complex imagery about 89% of the time. But it was much weaker at capturing qualities like vividness, how long the image lingered, and how unpleasant it felt. Those are often the parts people most want to communicate.

The proposed fix isn't a better AI model — it's a different format. The PRT workflow keeps the person's original words right next to the generated image, so readers can see what was said versus what was drawn. It also flags what's genuinely missing from the description versus what the AI simply failed to represent. The paper also tested 72 matched illustrations from 24 reports and found that automated scoring tools couldn't prove this structured approach was better, largely because the scores were too close to the ceiling to be meaningful.

That last point is the paper's quiet warning. AI illustrations of private, subjective experiences can look convincing and still be wrong. The atlas format is designed to make those gaps visible rather than paper over them — a deliberate refusal to treat a generated image as a measurement of what someone felt.