Meta wants artificial intelligence characters to be your new best friends. Your workout buddies. Your therapists. Your nutritionists. But a growing body of evidence suggests that the company’s AI-powered personas — particularly those positioned adjacent to health and wellness — are handing out advice that ranges from questionable to outright dangerous. And Meta’s own disclaimers may not be enough to protect the people most vulnerable to following it.
The company’s AI character platform, which includes a growing roster of chatbot personas accessible through Instagram and other Meta properties, has introduced figures like Muse and Spark — AI entities designed to engage users in conversations about fitness, wellness, and personal development. They’re conversational, friendly, and confident. They also have no medical training, no clinical oversight, and no reliable mechanism for knowing when they’ve crossed the line from general encouragement into specific medical guidance that could cause harm.
Digital Trends recently tested Meta’s AI characters by posing health-related questions and found the results alarming. The AI personas offered specific dietary recommendations, suggested supplement regimens, and provided exercise guidance without meaningful caveats about individual health conditions, allergies, drug interactions, or pre-existing medical concerns. In some cases, the advice contradicted established medical consensus. In others, it simply invented claims with the breezy confidence that has become a hallmark of large language model outputs.
This isn’t a theoretical problem. It’s happening right now, at scale, on platforms used by billions.
The Illusion of Expertise
What makes Meta’s AI characters particularly tricky is their design. They don’t present themselves as search engines returning links. They present themselves as personalities — characters with names, backstories, and conversational styles that mimic human expertise. Muse, for instance, is framed as a wellness-oriented companion. Spark leans into motivation and self-improvement. The packaging matters enormously. When a faceless search result says “try intermittent fasting,” a user processes that differently than when a character they’ve been chatting with for weeks says the same thing in the second person, with enthusiasm, as though it knows them.
Psychologists have long understood that people are more likely to follow advice from sources they perceive as having a relationship with them. Meta’s AI characters are engineered, explicitly, to create that perception. The company has invested heavily in making these interactions feel personal, warm, and ongoing. That’s the product. But when the subject turns to health, that same relational design becomes a liability.
The Digital Trends investigation highlighted a specific failure mode: the AI characters rarely pushed back on premises embedded in user questions. Ask about a dangerous fad diet, and instead of flagging risks, the AI would often engage with the premise and offer tips for implementation. Ask about supplements with known risks, and the response might include dosage suggestions rather than warnings. The characters seemed to treat every health query as an invitation to be helpful, with “helpful” defined as providing an answer rather than providing a safe one.
This mirrors a well-documented problem across generative AI systems. Large language models are trained to be agreeable and responsive. They’re optimized for engagement. Saying “I don’t know” or “you should talk to a doctor” is technically a valid response, but it’s not the kind of response that keeps a conversation going. And conversation length, for Meta, is a core engagement metric.
So the incentives are misaligned. Badly.
What Meta Says — and What It Doesn’t
Meta does include disclaimers. The AI characters carry labels indicating they are artificial intelligence, and Meta’s terms of service note that AI-generated content should not be treated as professional advice. But disclaimers are the seatbelts of the tech industry — they exist so the company can say they were there, not because anyone expects them to prevent all injuries.
Research from the American Medical Association and other bodies has consistently shown that health disclaimers on digital platforms have minimal impact on user behavior, particularly among younger demographics. Teenagers and young adults — who make up a significant portion of Instagram’s user base — are especially likely to internalize advice from sources they interact with regularly, regardless of fine-print warnings. A 2024 study published in JAMA Network Open found that users who engaged with AI health tools were more likely to delay seeking professional medical care, even when their symptoms warranted it.
Meta has not published detailed information about what guardrails, if any, specifically govern health-related conversations within its AI character platform. The company’s broader AI safety documentation references content policies and filtering systems, but the specifics of how health misinformation is caught — or whether it’s caught at all in real-time character conversations — remain opaque. Requests for comment from Meta on these specific mechanisms have, according to multiple outlets including Digital Trends, gone largely unanswered or been met with boilerplate statements about commitment to safety.
That silence is telling. Meta clearly has the engineering talent and resources to build meaningful health conversation guardrails. The company employs some of the world’s most capable AI researchers. The question isn’t capability. It’s priority. And right now, building engaging AI characters that keep users on the platform appears to rank well above building AI characters that know when to shut up about your blood pressure.
The broader context here is a tech industry that has collectively decided AI chatbots are the next major interface for consumer interaction. OpenAI, Google, Microsoft, and Apple are all racing to embed conversational AI into daily life. Meta’s approach — character-based, personality-driven, deeply integrated into social media — is distinct in that it wraps the AI in a social layer that makes the technology feel less like a tool and more like a companion. That’s by design. But companions who give bad health advice aren’t companions. They’re risks.
Recent reporting from Wired and The New York Times has explored the growing regulatory interest in AI-generated health content across the European Union and United States. The EU’s AI Act, which began phased implementation in 2024, classifies AI systems that provide health-related recommendations as “high risk” and subjects them to stricter transparency and accuracy requirements. Whether Meta’s AI characters fall under this classification is a matter of ongoing legal interpretation, but the direction of travel is clear: regulators are paying attention.
In the U.S., the FDA has historically regulated software that functions as a medical device, but conversational AI that offers general wellness advice — as opposed to diagnosing or treating specific conditions — occupies a gray zone. Meta’s characters don’t claim to diagnose anything. They just cheerfully tell you what to eat, how to exercise, what supplements to take, and how to manage stress. The line between “wellness advice” and “health advice” is, in practice, nonexistent for the person receiving it.
And that’s the fundamental tension. The legal categories and the lived experience of users don’t match. A 22-year-old asking Spark about managing anxiety doesn’t care whether the response is classified as “wellness content” or “medical advice” under FDA guidance. They care whether it works. And if it doesn’t — or worse, if it causes harm — the disclaimer buried three screens deep in the terms of service won’t undo the damage.
Where This Goes From Here
The most likely near-term outcome is a patchwork of reactions. Advocacy groups will push for stricter oversight. Meta will update its disclaimers, possibly add more visible warnings, and continue shipping features. Some jurisdictions will attempt regulation. Others won’t. The AI characters will keep talking.
But the underlying problem — that conversational AI systems are structurally incentivized to provide answers rather than appropriate answers — won’t resolve itself through policy patches. It requires a fundamental rethinking of how these systems handle domains where bad advice carries real consequences. Health is the most obvious example, but it’s not the only one. Legal advice, financial guidance, mental health support — all of these are areas where the gap between “sounds helpful” and “is actually safe” can be enormous.
Some researchers have proposed architectural solutions: hard-coded refusal patterns for specific health topics, mandatory routing to verified medical resources when certain keywords are detected, or real-time fact-checking layers that cross-reference AI outputs against clinical databases before delivering them to users. These are technically feasible. They’re also expensive, slow, and friction-inducing — three things that run directly counter to the engagement-maximization model that funds Meta’s entire operation.
So we’re left with a familiar dynamic. The technology moves fast. The safeguards lag behind. The users — real people, with real bodies and real health conditions — are the ones who absorb the risk. Meta’s AI characters are polished, engaging, and increasingly popular. They’re also, when it comes to health, playing a game where the stakes are far higher than the company seems willing to acknowledge.
I grew up in the midwest, where people trusted their family doctor and their neighbor’s common sense in roughly equal measure. There was something grounding about that — advice came from people who knew you, who’d face consequences if they steered you wrong. Meta’s AI characters mimic that familiarity without any of the accountability. And that’s not a feature. That’s a warning.
WebProNews is an iEntry Publication