AI Tool 'Waldo' Detects Hidden Health Risks in Social Media Posts

Researchers at UC San Diego have developed an artificial intelligence system capable of sifting through social media posts to identify health harms that people report about consumer products—a gap that traditional safety monitoring has largely missed. The tool, called Waldo, was tested on Reddit discussions about cannabis-derived products and achieved 99.7% accuracy when compared against human reviewers who manually identified adverse events in the same posts. When the team ran Waldo across a dataset of 437,132 Reddit posts about cannabis products, it flagged nearly 29,000 potential reports of harm. When researchers manually checked a random sample of these flagged posts, they confirmed that 86% represented genuine adverse events.

The motivation behind Waldo reflects a real gap in how the United States monitors product safety. The FDA's current system for tracking adverse events relies on voluntary reports submitted by doctors and manufacturers—a process that works reasonably well for prescription medications and medical devices that have gone through formal approval. But the market for consumer health products has expanded rapidly in recent years: cannabis-derived products, dietary supplements, and other items sold without FDA oversight now reach millions of people. These products generate no mandatory safety reporting, which means serious harms can persist undetected for months or years. Waldo was designed to capture the safety signals that people naturally share online, treating social media not as noise but as a legitimate source of real-world evidence about what products do to people's bodies.

The technical achievement matters. Waldo is built on a machine learning model called RoBERTa that was carefully trained to recognize the specific language people use when describing health problems. When researchers compared it to ChatGPT—a general-purpose AI chatbot—Waldo substantially outperformed the larger, more famous model. This suggests that specialized training on a specific task produces better results than relying on a general system, no matter how sophisticated. The researchers have released Waldo as open-source software, meaning regulators, clinicians, researchers, and public health officials can download it and use it without paying licensing fees or waiting for permission.

Karan Desai, the study's lead author, framed the finding this way: the health experiences people share online contain valuable information about safety. By capturing these voices systematically, researchers can surface real-world harms that traditional reporting systems never see. John Ayers, a senior researcher on the project, emphasized that the work demonstrates how digital tools can reshape post-market surveillance—the ongoing monitoring of products after they reach consumers. Vijay Tiyyala, another team member, noted that the accuracy of the model was encouraging, suggesting that careful training can produce tools that outperform state-of-the-art alternatives.

The immediate application was cannabis products, but the researchers suggest Waldo's approach has broader reach. Any consumer health product that lacks regulatory oversight could potentially be monitored this way. The team's decision to make the tool open-source reflects a deliberate choice to democratize access to safety monitoring technology. Rather than keeping Waldo proprietary or limiting it to a single institution, they've made it available to anyone who wants to use it. This move is intended to accelerate what researchers call open science—collaborative, transparent research that spreads knowledge and tools widely—while simultaneously improving the safety net for patients who use products that fall outside traditional regulatory channels.