Skip to main content
A Student-Led Research Initiative

Know what your health tools know.
And what they don't.

AI health tools don't work equally well for everyone. This project gives you the research, the data, and the framework to understand why — and to make better decisions about your own health.

Does the AI answering your health question work the same for everyone?

Peer-reviewed studies in Science Advances, The Lancet Digital Health, and npj Digital Medicine show that many AI health tools produce different results depending on a patient's race, skin tone, or background. These aren't edge cases — they're measured, documented, and published in the most respected journals in medicine.

The reasons trace back decades. Medical algorithms were built on data that reflected historical biases in clinical practice — biases that have since been identified and, in many cases, corrected at the institutional level. But the training data that AI models learned from still carries those patterns.

Understanding this isn't about fear. It's about being an informed consumer of the tools your generation uses more than any before it.

The data is clear. Here's what the science says.

These findings come from peer-reviewed research. Knowing them puts you ahead of most adults.

Why It's Called

Diagnose the Bias

Doctors diagnose problems so they can treat them. The same principle applies here — you can't fix what you can't see. This project teaches you to identify where bias lives in health AI so you can make better decisions about the tools you use.

97%

Stereotyping rate in one AI model's generated cases

Asked to generate a sarcoidosis case, one major AI model described a Black patient 97% of the time (966 of 1,000) — wildly exaggerating a real but much smaller demographic pattern.

Zack et al., Lancet Digital Health, 2024
100%

of the four major AI models tested repeated race-based medical myths

When tested with race-sensitive medical questions, all four major AI models tested — GPT-3.5, GPT-4, Bard, and Claude — produced at least some debunked race-based claims.

Omiye et al., npj Digital Medicine, 2023

Explore all 47 documented cases →

This isn't about avoiding technology. It's about using it better.

AI health tools are useful. They're also imperfect — and the imperfections aren't random. When you understand where bias lives in these systems, you can make smarter choices about what to trust, what to question, and when to seek a second opinion.

Diagnose the Bias is a free, freely shareable curriculum designed for high school students. It's built on published research, grounded in real data, and focused on one goal: giving you the knowledge to be a more informed, more autonomous health consumer.

📊

See the data

Real findings from peer-reviewed journals, translated for a student audience.

🔍

Understand the why

How historical patterns in medical practice became training data for modern AI.

💪

Own your decisions

A practical framework for evaluating health information — on your terms.

Did You Know?
Loading...