Meta's AI Agent Faces Trust Questions as Users Push Back

Meta's plan to put AI agents in front of its entire user base is colliding with a problem no amount of distribution can solve: people do not trust what the assistant tells them. The Wall Street Journal reported that the company's agent products are drawing scrutiny over reliability and credibility, a setback for a feature Meta has framed as central to its next stage of growth. The company has spent heavily to make an AI assistant a default presence inside Messenger, Instagram, WhatsApp and its Ray-Ban smart glasses, a footprint that reaches roughly 3 billion people daily.

The scale is what separates Meta's situation from that of other AI vendors. When an agent misfires at a typical software company, a few thousand users notice; when it misfires inside Meta's apps, the audience runs into the billions. Trust is not an abstract metric for agent products, because an agent acts on the user's behalf — drafting messages, answering questions, recommending content and increasingly taking actions inside accounts. If people cannot tell when the system is guessing, they either stop using it or stop believing it.

Meta has marketed the assistant as helpful and personal, and that is precisely where the criticism is landing: the distance between the promise and daily output. The company has previously tried to get ahead of such complaints by attaching labels to AI-generated content and broadening its AI disclosure rules. Labeling output, however, is not the same as earning confidence in an agent that speaks in the user's voice — a gap that matters as much to advertisers and regulators as it does to users. The Federal Trade Commission has already signaled interest in how AI systems present themselves to consumers, and a high-profile failure inside Meta's apps would hand those inquiries fresh material.

Meta is not alone in absorbing these complaints. OpenAI, Google and Anthropic have each shipped agents over the past year and each has fielded criticism about confident wrong answers, invented citations and erratic behavior in long conversations. Meta's case is distinct because it does not have to persuade anyone to download anything. Its agents reach users who never opted into an AI product and have no baseline for what the tool can and cannot do — a population far less equipped to catch a bad answer.

What to watch next is whether Meta changes the product itself rather than its messaging. That could mean tighter limits on what the agent will attempt, clearer signals when it is uncertain, or a narrower set of tasks it can perform without user confirmation. The company has the reach to make AI agents ordinary; the open question is whether it can make them trustworthy first. If the rollout stalls on trust, it slows Meta's plan to monetize AI through engagement and ads and gives rivals room to pitch their own assistants as the safer default.