AI chatbots reading X-rays can be dangerously confident even when they're wrong
Quick Answer
AI chatbots struggle with X-ray interpretation, scoring significantly lower than human radiologists, with the best model, Google's Gemini 3 Pro, achieving only 758 points out of 2000.
Quick Take
The study emphasizes the dangers of overconfidence in AI, as many models frequently misdiagnose while exhibiting high confidence, undermining their reliability in medical contexts.
Key Points
- Human radiologists scored 988.7 points, outperforming all AI models tested.
- The scoring system penalizes overconfidence, rewarding honesty in diagnoses.
- Meta's Muse Spark 1.1 excelled at recognizing when to defer to human experts.
- Patients increasingly trust AI chatbots for medical image interpretations.
- AI models often misdiagnose with high confidence, raising safety concerns.
DeepSignal Analysis
What happened
A recent study evaluated 16 AI models against human radiologists in interpreting X-rays, revealing that human experts scored significantly higher. The best AI model, Google's Gemini 3 Pro, achieved only 758 points out of 2000, while human radiologists scored 988.7. The study highlights the issue of AI overconfidence, as many models misdiagnose while exhibiting high confidence, raising concerns about their reliability in medical settings.
Key evidence
- The study involved 200 cases and found that human radiologists outperformed all AI models, scoring 988.7 out of 2000 points compared to the best AI model's score of 758.
- Meta's Muse Spark 1.1 was noted for its ability to recognize when to defer cases to human radiologists, showing a significant reduction in its hallucination rate.
- The research team criticized claims that AI systems diagnose better than 99 percent of doctors, stating that such assertions are often based on anecdotes or simulations rather than empirical evidence.
Why it matters
The findings underscore the limitations of current AI models in medical diagnostics, particularly their tendency to misdiagnose with high confidence. This overconfidence can lead to dangerous outcomes in clinical settings, where accurate diagnosis is critical. The study also raises ethical concerns regarding the reliance on AI in healthcare, as patients increasingly trust chatbots for medical advice despite their unreliability.
Source Excerpt
From the original publisher, up to about 700 charactersThe RadLE 2. 0 benchmark tests whether AI models in radiology can tell when they should leave a diagnosis to a human. Many models deliver wrong findings with full confidence, and human radiologists are still well ahead. Before AI can diagnose on its own, it needs to learn when it's better to say nothing.
Want this in your inbox every morning?
Daily brief at your local 8am — bilingual EN/中文, free.
More from The Decoder
See more →
An AI model programmed nonstop for 19 days on a single MirrorCode task that cost $2,600 to run
Epoch AI's MirrorCode benchmark reveals Claude Opus 4.7 as the leader with a 56% solve rate, reconstructing a 16,000-line toolkit in 14 hours. Despite this, all models tested struggle with the most complex tasks, highlighting limitations in current AI capabilities. The single task consumed $2,600 over 19 days, raising questions about cost-effectiveness in AI development.
