I am at #EMNLP2025🇨🇳 to present our main paper *MythTriage: Scalable Detection of Opioid Use Disorder Myths on a Video-Sharing Platform*! Come by to discuss details!

🏦 Location: Hall C
⏲️Time: 11AM-12:30PM
🔗 Paper: aclanthology.org/2025.emnlp-m...
📁 Repo: github.com/hayoungjungg...

📊Finding #3: YouTube’s recommendation continued surfacing myth-supporting content.

➡️12.7% of recs from myth videos led to more myths initially—rising to 22% at deeper levels.

⚠️ Moderation should target these rec pathways that reinforce harmful myths.

📊 Finding #2: How you filter your search results matters! Switching from “Relevance” to “Upload Date” or “Rating” increases exposure to myths—echoing the same patterns seen in my COVID-19 misinformation audit: ojs.aaai.org/index.php/IC...

😬A few clicks can change your exposure to myths!

🫶Thanks to MythTriage, we present the first large-scale study of OUD-related myths on YouTube!

📊 Finding #1: Nearly 20% of YouTube search results support OUD myths, while 30% oppose.

😰Despite more opposing, myth-supporting content is widespread—and risks shaping how people understand treatment.

⚙️So how does MythTriage perform?
📊 Achieves 0.68-0.86 macro F1 and defers only 5-67% of the examples to the costly LLM.

In practice, MythTriage:
💸 Cuts financial costs by 98% vs experts and by 94% vs LLM labeling
⏱️ Cuts time costs by 96% vs experts & by 76% vs LLM labeling

🚀 Our solution: MythTriage
👉 Uses lightweight DeBERTa for routine cases
👉 Defers harder ones to GPT-4o (high-performing but costly)

The trick? We distilled DeBERTa on GPT-4o’s synthetic labels—achieving strong performance without massive expert-labeled data.

🩺 To rigorously detect OUD myths in our datasets, we collaborated closely with clinical experts to:

✅Validate eight pervasive myths on OUD (see examples below!)
✅Create and refine annotation guidelines
✅Build a gold-standard dataset: 310 videos labeled across 8 myths (~2.5K expert labels).

To measure the scale and prevalence of myths on YouTube, we curated opioid and OUD search queries based on real-world search interests. Using these queries, we built two datasets on YouTube:

1️⃣ OUD Search Dataset: 2.9K search results
2️⃣ OUD Recs Dataset: 343K video recommendations

🚨YouTube is a key source of health info, but it’s also rife with dangerous myths on opioid use disorder (OUD), a leading cause of death in the U.S.

To understand the scale of such misinformation, our #EMNLP2025 paper introduces MythTriage, a scalable system to detect OUD myth🧵

8/ SA bots encountered significantly more misinformation in the default Relevance filter.

Since most users will likely engage with SERPs in this default settings, users in SA 🇿🇦 face a higher likelihood of misinformation exposure, raising public health risks.

6/ Alarmingly, two topics—Biological Weapon and Lab Leak Theory—were riddled with misinformation across all search filters. Interestingly, while 3 topics opposed misinfo under the default Relevance filter, switching to filters like Upload Date revealed SERPs packed with misinfo.

5/ The disparity was especially pronounced in topics like:
📡 5G Claims
💉 Vaccine Content Claims
💻 Bill Gates Claims
Effect sizes suggest 🇿🇦 users are at greater risk of encountering misinformative search results than 🇺🇸 users.

4/ We analyzed misinformation bias scores in search engine result pages (SERPs) across both countries. Higher scores reflect a greater prevalence of misinfo in the SERPs⚠️ Bots in SA faced significantly more misinformation in top-10 results, which account for 95% of user traffic

2/ To fill this gap, we conducted a large-scale geolocation-based audit of YouTube Search, comparing results in the US🇺🇸 and SA🇿🇦—countries heavily impacted by the pandemic in the Global North and South.

How does YouTube’s search algorithm handle COVID🦠 misinfo in the United States🇺🇸(US) and South Africa🇿🇦(SA)?

In our #icwsm '25 paper w/ @prerna6.bsky.social @tanumitra.bsky.social, we found bots in SA received significantly more misinfo in top-10 search results, which accounts for 95% of user traffic