Blinded Scientific Drawing: Can Expert Interaction Beat AI Prompting

Описание: Can human experts collaboratively convey subtle phenotypic information more effectively than text-based AI prompting?

In this blinded scientific drawing experiment, a professional scientific illustrator starts from a neutral, average face and iteratively refines the sketch based solely on feature-level instructions provided by expert dysmorphologists. The diagnosis is never named, and only facial characteristics (HPO-style descriptors) are allowed.

The final expert-guided drawing is evaluated using GestaltMatcher, a deep-learning–based facial phenotyping tool. In parallel, generative AI images are created using exactly the same verbal instructions given to the illustrator. Both approaches are assessed under identical, fully blinded conditions.

This session builds on prior observations showing that collective expert feedback consistently improves phenotypic drawings—and explores whether human–human interaction may currently outperform human–machine prompting when translating complex clinical gestalt into visual representations.

🎄 Christmas Special Edition

Key topics:

Blinded phenotyping experiments

Human-in-the-loop vs. AI-only approaches

Dysmorphology & facial gestalt

HPO-driven feature reinforcement

Generative AI limitations in rare disease phenotyping

After the experiment, the secret syndrome is revealed and discussed using a positive control derived from monozygotic twins discordant for the condition.