Meta AI gave bad advice

A general chatbot is built to answer almost anything, from vacation plans to algebra, and that is exactly why doctors worry when the same tool starts acting like a lab interpreter. Meta’s new Muse Spark did that this week inside the Meta AI app, where a Wired reporter was prompted to paste in blood pressure readings, glucose data, and lab results for analysis. (wired.com) Meta launched Muse Spark on April 8 as the first model from Meta Superintelligence Labs, the new artificial intelligence group led by Alexandr Wang after Meta’s $14.3 billion Scale AI deal. Meta said the model is “small and fast by design” but still able to reason about “science, math, and health.” (about.fb.com) (cnbc.com) The health pitch was not buried in fine print. Meta’s own launch post said Muse Spark can identify high-protein snacks from a photo and compare products, and Wired found the bot explicitly offering to “calculate trends, flag patterns, and visualize” data from a lab report or glucose monitor. (about.fb.com) (wired.com) That sounds useful until you remember what raw lab data actually is. A cholesterol panel or blood count is less like a school test with one right answer and more like a dashboard full of warning lights that only make sense with age, symptoms, medications, and medical history. (wired.com) In Wired’s test, Muse Spark did not just stay vague and cautious. The report says the bot produced poor guidance, missed important context, and gave answers that were unreliable enough that a real clinician would not want a patient acting on them alone. (wired.com) Meta says it worked with more than 1,000 physicians to curate training data for more factual health responses. That claim shows the gap in this story: better training examples can make a chatbot sound more medical, but they do not turn a broad consumer assistant into a regulated diagnostic tool. (wired.com) (about.fb.com) The privacy problem sits right next to the accuracy problem. TechCrunch noted that Muse Spark requires a Meta account login, and Wired’s test involved the bot asking for exactly the kind of sensitive numbers people usually share only with a clinic, insurer, or patient portal. (techcrunch.com) (wired.com) This is the part critics keep circling: the danger is not a dedicated medical device pretending to be a doctor. The danger is a friendly all-purpose assistant, sitting next to your social accounts and shopping prompts, casually inviting you to hand over blood work as if it were calorie counts. (wired.com) (techcrunch.com) Meta plans to push Muse Spark beyond the app to WhatsApp, Instagram, Facebook, Messenger, and its artificial intelligence glasses in the coming weeks. A tool that reaches billions of people does not need to be disastrously wrong every time to create a problem; it only needs to sound confident often enough that people stop double-checking. (about.fb.com)