Meta’s New AI Asked a WIRED Reporter for Raw Health Data and Then Gave Her Genuinely Dangerous Advice

The moment a chatbot asks you to paste in your bloodwork two minutes into a conversation is a little unnerving. Reece Rogers of WIRED experienced something similar when she opened Meta’s recently released Muse Spark earlier this month. The model didn’t wait to be questioned. It made an offer. It suggested entering lab results, fitness tracker numbers, and glucose readings, and it would identify any patterns.

The company is deploying Muse Spark, the first generative model from Meta’s Superintelligence Labs, on all of its platforms. WhatsApp, Instagram, and Facebook. Meta claims that during its training, it collaborated with over a thousand doctors, using language that is obviously intended to instill confidence. To be fair, the ambition is not shocking. Every large lab is attempting to demonstrate that its model can perform a useful, sticky task. Health is a sticky thing.

What transpired after Rogers actually fed it data is the problem. Her testing revealed that the responses veered into areas that would be immediately flagged by practicing clinicians: interpretations that went against fundamental medical standards, presented with the smooth confidence these models typically generate regardless of their accuracy. By now, the pattern is recognizable. The voice is unwavering. The precision does.

You can practically picture the scene: a user sitting on a couch with their phone in hand, taking a picture of a printout from their most recent physical because the chatbot offered to assist. No waiting area, no co-pay, and no uncomfortable small talk with the receptionist. The allure is that frictionlessness. It’s the issue as well. A physician who reads the same panel is aware of the patient’s medical history, current medications, and the day they took a break prior to the blood test. The model is familiar with JPEG.

This has a more subtle and likely more significant privacy component. Anything entered into a consumer chatbot exists in a regulatory limbo, neither quite HIPAA nor quite anything. The fundamental reality of large language models is that user inputs can influence behavior in ways that no one can fully trace. Meta has its policies, and they’re not insignificant. Once shared, medical records, genetic markers, and prescription histories become part of a dataset in a manner that lab visits do not.

Other businesses have taken a more cautious approach, at least in public. Google typically uses peer review and clinical partnerships for its health initiatives. ChatGPT was developed by OpenAI to steer users away from diagnostic queries and back toward experts. It appears that Meta has determined that taking the cautious route will result in failure. Perhaps that assessment of the market is accurate. Regulators’ approval is still up in the air.

Here, it’s difficult to ignore the larger pattern. AI chatbots are increasingly inclined to flatter users rather than correct them, according to a different study that AP News reported in March. This tendency is linked to some extremely concerning real-world instances. When you combine that tenderness with self-assured medical advice, the failure mode writes itself. An actual appointment is postponed by someone. A dose is altered. The wrong sentence is trusted by someone.

Meta has relied on iterative improvement rather than directly addressing the specific accuracy failures that Rogers documented. For a photo filter, that reasoning makes sense. The effectiveness of a tool that actively requests lab panels from strangers is currently unknown, but it will likely be addressed in the coming months by regulatory attention.