Detailed Analysis
A Reddit post circulating in mid-2026 highlights a potentially dangerous failure by Anthropic's Claude AI assistant, in which the system incorrectly characterized a blood pressure reading of 85/45 mmHg as "optimal." The post, accompanied by a screenshot of the exchange, quickly gained attention due to the severity of the error. A blood pressure reading of 85/45 mmHg falls well below the clinical threshold for hypotension, which is generally defined as anything under 90/60 mmHg. Such a reading in a real patient would typically indicate a medical emergency — potentially consistent with septic shock, severe dehydration, or cardiac failure — and would warrant immediate emergency intervention rather than reassurance.
The nature of the error points to a category of AI failure sometimes called a "confident hallucination," where a model not only produces incorrect information but does so with a tone of authority that discourages the user from seeking a second opinion. In a medical context, this is particularly hazardous. The irony captured in the post's title — that asking for medical advice yielded what amounts to a "funeral plan" — reflects a dark but pointed critique: that validating a critically low blood pressure as healthy could, in a worst-case scenario, discourage someone from seeking urgent care. The screenshot format of the post also suggests the failure was clear-cut and visually demonstrable, lending credibility to the complaint.
This incident connects to a long-standing and unresolved tension in large language model deployment around medical use cases. Anthropic and other AI developers have consistently warned that models like Claude are not substitutes for professional medical advice, and Claude's system-level guidelines are designed to encourage users to consult healthcare providers. However, the failure described here suggests that such guardrails did not prevent a materially false and dangerous medical assessment from being delivered with apparent confidence. Whether the error stemmed from a misreading of the numerical values, a confusion between blood pressure and some other metric, or a broader gap in the model's clinical reasoning is unclear from the available information.
Broader industry trends make this case worth examining carefully. As AI assistants become increasingly embedded in everyday consumer behavior, users frequently turn to them for quick medical triage — interpreting lab results, assessing symptoms, or evaluating vital signs — regardless of whether developers intend or endorse that use. The gap between intended use and actual use creates a zone of liability and risk that the AI industry has not fully resolved. Anthropic has invested significantly in Constitutional AI and safety-focused training methods, but cases like this illustrate that even well-resourced safety programs do not eliminate the risk of high-stakes factual errors in sensitive domains. The incident adds to a growing body of documented cases that researchers and regulators are using to argue for more rigorous, domain-specific evaluation standards before AI systems are used — formally or informally — in clinical or quasi-clinical contexts.
Read original article →