AI chatbots fail at differential diagnosis, according to a new study from Mass General Brigham that tested 21 popular models on clinical cases. The results show that even the most advanced AI systems struggle with the kind of clinical reasoning that doctors rely on every day.
The study, conducted from January to December 2025, evaluated chatbots including Claude 4.5 Opus, DeepSeek, Gemini 3.0 Pro, GPT-5, and Grok 4. Researchers gave the models dozens of clinical cases with basic information—the same information a human doctor might have at the start of a diagnostic workup—and asked them to identify the most likely diagnosis.
What the study found: AI chatbots fail at differential diagnosis
The researchers from Mass General Brigham, a nonprofit healthcare network that includes two of the most prestigious medical teaching institutions in the U.S., wanted to see if large language models could serve as clinical allies. The answer, according to the results, is no.
While models optimized for reasoning—such as those with chain-of-thought capabilities—scored much higher than simpler models, even the best performers made significant errors. The study highlights that AI chatbots often hallucinate or provide incorrect diagnoses when faced with complex, real-world clinical scenarios.
This isn't the first time AI has stumbled in medicine. Previous research has shown that AI can be biased, inconsistent, and prone to overconfidence. But this study is particularly notable because it tested the most popular and advanced models available today.
Why differential diagnosis is so hard for AI
Differential diagnosis is the process of distinguishing between multiple conditions that could explain a patient's symptoms. It requires not just medical knowledge but also clinical judgment, pattern recognition, and the ability to weigh probabilities.
Human doctors use years of training and experience to narrow down possibilities. They consider factors like patient history, physical exam findings, and test results. AI models, on the other hand, rely on patterns in text data. They don't have hands-on experience or the ability to truly understand the context of a patient's life.
Moreover, the study points out that AI models can be overconfident, presenting incorrect diagnoses with high certainty. This is dangerous in a clinical setting, where a wrong diagnosis can lead to harmful treatments or missed opportunities for proper care.
- Reasoning-optimized models performed better but still made critical errors.
- Basic clinical information was provided, mirroring real-world diagnostic starting points.
- Hallucinations and incorrect diagnoses were common across all models.
- Clinical reasoning remains a challenge for AI, despite advances in language generation.
- Patient safety is at risk if AI is used without human oversight.
How the 21 chatbots performed
The study tested a wide range of models, from general-purpose chatbots to those specifically designed for medical applications. The table below summarizes the key differences in performance based on the study's findings.
| Model Type | Example Models | Performance | Key Limitations |
|---|---|---|---|
| Reasoning-optimized | Claude 4.5 Opus, Gemini 3.0 Pro, GPT-5 | Higher scores than simpler models | Still made significant errors; overconfident |
| General-purpose | DeepSeek, Grok 4 | Moderate scores | Prone to hallucinations; inconsistent |
| Simpler models | Various older chatbots | Lower scores | Frequent incorrect diagnoses |
It's important to note that even the best-performing models were not reliable enough for clinical use. The researchers emphasize that AI should not replace human judgment in diagnosis.
What this means for the future of AI in medicine
Despite the disappointing results, AI still has potential in healthcare. It can assist with administrative tasks, image analysis, and even suggesting possible diagnoses as a second opinion. But the study makes clear that AI is not ready to take over the diagnostic process. The background to this story is set out in AI in Rural Healthcare: How Ant Afu Helps One Doctor Treat 2,000 Patients. A related thread runs through AI Battle in US Healthcare: Algorithms Fight Over Billing and Denials. Earlier coverage in Health: latest articles shows how the argument developed.
For patients, this means that you should be cautious about using AI chatbots for medical advice. Always consult a qualified healthcare professional for diagnosis and treatment. For doctors, the study serves as a reminder that AI tools should be used with caution and always under human supervision. A closer look at the same subject appears in Blood Test to Predict Alzheimer's 10 Years Early: Promise and Risk. Anyone following the wider picture can start with Why Doctors Still Work 24-Hour Shifts: The Cocaine Origin Story. The same theme surfaces in Carnivore Diet Movement Gains Ground in US Healthcare, which takes a different angle.
The researchers call for more rigorous testing and development of AI models that can truly understand clinical reasoning. Until then, the differential diagnosis remains a human domain. Context on how this played out before sits in NHS Staff Snooping on Patient Records: 3,000 Incidents Spark Outrage. A companion piece, AI Chatbots Fail Differential Diagnosis in New Study, approaches the same subject from another angle.
FAQ
What was the main finding of the Mass General Brigham study?
The study found that 21 popular AI chatbots, including Claude 4.5 Opus, Gemini 3.0 Pro, and GPT-5, failed at differential diagnosis. They often provided incorrect or hallucinated diagnoses, even when optimized for reasoning.
Why is differential diagnosis difficult for AI?
Differential diagnosis requires clinical judgment, pattern recognition, and the ability to weigh probabilities based on patient context. AI models lack hands-on experience and can be overconfident in incorrect answers.
Can AI chatbots be used for medical advice?
Based on this study, AI chatbots are not reliable for medical diagnosis. They should not replace human doctors. Always consult a healthcare professional for medical concerns.
What should patients and doctors take away from this study?
Patients should avoid relying on AI for diagnosis. Doctors should use AI tools with caution and always verify AI-generated suggestions with their own clinical judgment.