Seeking Help, Facing Harm: Auditing TikTok's Mental Health Recommendations
A 7-day audit of 30 accounts shows engagement drives 45% of recommendations, even for harmful content.
A team of researchers from the University of Pennsylvania and Yale, led by Pooriya Jamie, Amir Ghasemian, and Homa Hosseinmardi, published a study titled 'Seeking Help, Facing Harm: Auditing TikTok's Mental Health Recommendations.' The research, accepted at ICWSM 2026, conducted a rigorous 7-day audit using 30 fresh TikTok accounts controlled by LLM-guided agents. These agents simulated different user behaviors: some started with distress-related searches, others with help-seeking queries, and then followed engagement, avoidance, or passive viewing strategies.
The study analyzed 8,727 recommended videos and found a critical flaw: TikTok's algorithm is primarily driven by engagement metrics, not user intent. Accounts that actively engaged with mental health content saw their feeds rapidly saturate, with approximately 45% of daily recommendations becoming mental health-related. In contrast, accounts that avoided such content still saw 11-20% of recommendations in this category. While help-initiated searches yielded slightly more supportive material, the algorithm failed to adequately filter harmful content, including videos in the Suicide/Self-Harm category, which persisted at low but non-zero levels.
The findings suggest TikTok's recommender system has 'limited sensitivity to user intent signals' for sensitive topics. The initial search framing (seeking help vs. expressing distress) mainly shifted the composition of the content, not its volume. This indicates the platform's safeguards are insufficiently context-aware, potentially exposing vulnerable users to harmful material even when they signal a desire for support. The research underscores the need for more nuanced algorithmic design that can interpret the subtle differences between a cry for help and a descent into a harmful content spiral.
- Engagement is the primary driver: Accounts that engaged with mental health content saw ~45% of their daily feed become related, demonstrating the algorithm's focus on watch time and interaction.
- Avoidance is ineffective: Even accounts programmed to avoid mental health topics still received 11-20% related recommendations, showing the system's limited respect for negative feedback.
- Intent signals are ignored: The algorithm showed limited ability to distinguish between 'help-seeking' and 'distress expression' searches, allowing potentially harmful content to reach vulnerable users.
Why It Matters
This reveals a fundamental flaw in how social platforms handle sensitive content, potentially worsening mental health crises instead of providing support.