Warning Message Content Increases Help Seeking in a Large-Scale Dark Web CSAM Intervention

Technical Summary: Warning Message Content Increases Help Seeking in a Large-Scale Dark Web CSAM Intervention

Problem Statement
Digital networks have significantly expanded the accessibility of child sexual abuse material (CSAM), creating a public health crisis where millions of children are exploited annually. While offline crime prevention struggles to reach individuals at the moment of risk, digital interventions offer a scalable mechanism to interrupt CSAM-seeking behavior. However, there is a significant gap between high-risk online activity and the uptake of therapeutic resources. Although warning messages are widely deployed on search engines and social media to divert users toward support services, large-scale field evidence regarding the efficacy of specific message content—particularly in high-anonymity environments like the Dark Web—remains limited. Existing studies often rely on self-reports or small-scale experiments, leaving a lack of empirical data on how message framing and thematic content influence actual help-seeking behavior in live operational settings.

Methodology
This study conducted a large-scale field experiment on Ahmia.fi, a Tor search engine, over a 140-day period (June 2025 to November 2025). The study utilized a quasi-experimental design combining a randomized field experiment with a campaign-wide evaluation.

• Data Scope: The researchers observed 19,611,744 total searches. Of these, 3,105,800 searches (15.84%) contained banned CSAM-related terms that triggered a warning intervention.
• Intervention Design: During the 70-day intervention period, users triggering a warning were presented with one of nine message variants or a neutral control message. The messages were rotated daily, with a specific "randomized-warning day" occurring every tenth day to allow for random assignment of message conditions.
• Message Variants: The active messages were categorized into four thematic areas, each with a negatively framed and a positively framed version:
  1. Legality/Consequences: Emphasizing risks of arrest and personal costs.
  2. Harm: Highlighting the impact of the behavior on victims.
  3. Control/Self-Efficacy: Emphasizing the individual's capacity to change and access support.
  4. Distress/Psychological Need: Prompting reflection on emotional drivers and consequences.
  • Neutral Condition: A standard informational message about the ReDirection self-help program without specific persuasive framing.
• Outcome Measures: The primary outcome was the Click-Through Rate (CTR) to the ReDirection anonymous self-help program. A secondary outcome was the rate of warning-triggering searches (deterrence).
• Analytic Strategy: Analyses were conducted using binomial logistic regression models on aggregated daily counts. The study employed three analytical approaches:
  1. Randomized-Day Analysis: Comparing message conditions on days where messages were randomly assigned, controlling for day fixed effects.
  2. Campaign-Period Analysis: Examining CTR across all campaign days, adjusting for day of the week and linear time trends using Newey–West heteroskedasticity and autocorrelation consistent (HAC) standard errors.
  3. Interrupted Time Series (ITS): Comparing pre-campaign (70 days) and campaign (70 days) data to assess immediate level changes and slope changes in CTR and warning-trigger rates.

Key Results

• Impact on Help Seeking (CTR): The study found that message content consistently influenced engagement with help resources.
  • Overall Increase: The platform-wide CTR increased from a pre-campaign baseline of 8.73% to 15.67% during the campaign.
  • Message Efficacy: All active message variants significantly outperformed the neutral message. In the randomized-day analysis, odds ratios for clicking through ranged from 1.12 to 1.76 relative to the neutral condition.
  • Thematic Performance: Negatively framed harm-focused messages produced the strongest effects on CTR. This was followed by positively framed harm messages, and then legality and control messages. Distress-focused messages showed smaller, though still reliable, effects.
  • Framing: While negatively framed harm messages performed best, differences between positively and negatively framed messages within the same theme were modest, with overlapping confidence intervals.
• Impact on Deterrence (Warning-Trigger Rate): Unlike the CTR results, the study found no evidence that message content affected the rate of warning-triggering searches. The campaign did not significantly alter the volume of searches that triggered a warning compared to the neutral condition or pre-campaign levels. This suggests the intervention primarily diverted users to help rather than deterring the search behavior itself at the aggregate level.
• Robustness: The findings were consistent across randomized-day comparisons, the full campaign period, and the interrupted time-series models, indicating that the effects were driven by message content rather than temporal confounds or seasonal variations.

Significance and Claims
The authors claim this study provides the first large-scale field evidence systematically examining the impact of warning message content on help-seeking engagement in a high-anonymity Dark Web environment.

• Ecological Validity: The study's primary contribution is its ecological validity, capturing behavioral responses in situ at the point of risk rather than relying on hypothetical scenarios or self-reports.
• Message Content Matters: The findings demonstrate that even in an anonymous environment where traditional deterrence cues (like IP tracing) are less credible, the specific content and framing of a warning message significantly shape user behavior.
• Adaptive Intervention: The results argue against a "one best message" approach. While harm-focused messages were most effective in this specific context, all active messages improved engagement over a neutral baseline. The authors suggest that warning messages should be viewed as part of an adaptive intervention system where content is iteratively refined based on user behavior and context.
• Mechanism of Action: The study indicates that in high-risk digital environments, the primary mechanism of successful warnings may be the disruption of goal-directed search and the immediate provision of a pathway to support, rather than the deterrence of the search itself.
• Limitations and Generalizability: The authors modestly note that the inability to track individual users limits the analysis of demographic or psychological subgroups. Furthermore, they caution that the specific effectiveness of these messages on Ahmia.fi (a Tor search engine) may not generalize directly to clear web platforms or other digital services, though the underlying mechanism of interruption and redirection is shared.

The paper concludes that optimizing warning message content is a practical, scalable means to increase help-seeking behavior, helping to close the gap between high-risk online activity and engagement with therapeutic resources.