The AI Was Consistently Wrong

In the HealthRadar Yardstick study, Google, OpenAI, and Anthropic repeatedly failed the same difficult synthetic laboratory case. Edward A. Rose examines what that result may—and may not—mean.

Perspective

Most conversations about generative AI in healthcare eventually arrive at the same concern: inconsistency.

Ask the same question twice and you may get two different answers.

But our recent experiment produced a result that bothers me for almost the opposite reason.

The AI was remarkably consistent.

We gave the same difficult synthetic laboratory case to three leading AI model families—Google Gemini, OpenAI GPT, and Anthropic Claude. Each model received the case three separate times, in independent stateless requests.

Nine interpretations in all.

Under our prespecified physician standard, every one of them failed the same safety requirement.

Nine attempts. Nine potentially unsafe classifications.

That doesn’t tell us how often an AI system would make a dangerous mistake in ordinary clinical use. Our study was far too small for that conclusion.

But it raises a question I think deserves considerably more attention:

What happens when AI isn’t randomly wrong—but reproducibly wrong about the same clinically important thing?

Read the research

The HealthRadar Yardstick original cohort is a small synthetic benchmark, not a clinical validation study. We have made the graphical abstract, complete manuscript/preprint, figures, study limitations, and public HealthRadar demonstration available online.

Read the HealthRadar Yardstick study

We’re also interested in hearing from researchers, clinicians, health systems, and others working on clinical AI evaluation, laboratory-result communication, and governed approaches to AI-assisted interpretation.

Most of all, we’d welcome independent scrutiny.

If this finding is real, it should be reproducible.

An earlier version of this essay was published on Edward Rose's Substack.