ChatGPT is bad at creating diagrams of body organs (masto.ai)

Researchers tested ChatGPT’s ability to generate anatomical diagrams of pregnancy and found consistently wrong, misleading outputs. Simple prompts like “Show me a diagram of the organs during pregnancy” produced side-view images with misplaced labels (vagina linked to the rectum), organs drawn from the wrong perspective, the liver inside the uterus, and a fetus shown in breech position. Even increasingly specific prompts—requesting a sagittal view, placenta, umbilical cord, rectum and bladder, or literature-backed sources—still yielded errors: missing or mislabeled bladder and placenta, umbilical cord wrapped unnecessarily around the fetus, and label placements that reflect where text should go rather than actual anatomy. This matters because generative AI assembles plausible-looking visuals from statistical patterns, not grounded anatomical knowledge, so near-correct diagrams can be dangerously convincing to lay users or educators without domain expertise. Technical implications include multimodal misalignment (labels vs. image content), dataset gaps or noisy training images, and models reproducing rare complications as “typical.” Practical remedies are clear: do not rely on general-purpose generative models for medical visuals; use curated medical datasets, specialist-trained models, and expert-in-the-loop validation; and add UX guardrails and provenance metadata so users know when an image is unverified.