Deterring Searches for Child Sexual Abuse Material on Google Search and Promoting Help-Seeking

Technical Summary: Deterring Searches for Child Sexual Abuse Material on Google Search and Promoting Help-Seeking

Problem Statement
The production, distribution, and consumption of Child Sexual Abuse Material (CSAM) have been exacerbated by global internet access, creating significant harm to victims and communities. While Electronic Service Providers (ESPs) like Google have long focused on "supply-side" prevention—using classifiers, hashing, and ranking protections to remove content and block access—there is an urgent need to address the "demand side." Specifically, there is a gap in understanding how to effectively intervene at the precise moment a user attempts to access CSAM to deter the behavior and redirect the individual toward therapeutic services. Existing digital interventions, such as warning messages, have varied in their framing (e.g., focusing on reporting vs. help-seeking), but large-scale field evaluations of their behavioral impact on actual search sessions have been limited.

Methodology
This study evaluates a strategic shift in Google Search's "Onebox" feature—a prominent UI element displayed at the top of search results—triggered when users query for CSAM. The research employs a quasi-experimental Difference-in-Differences (DID) design using internal, anonymized Google Search logs.

• Intervention: The Onebox messaging was revised in February 2025. The previous iteration focused heavily on reporting mechanisms (e.g., directing users to the CyberTipline or IWF). The revised iteration maintained the illegality warning but shifted the primary focus to the negative consequences of CSAM and prominently featured links to anonymous, confidential therapeutic resources (e.g., "Stop It Now!", "Kein Täter werden").
• Data Collection: The dataset included a country-stratified random sample of 10% of search sessions containing at least one CSAM-seeking query. The study analyzed 17 weeks of data (3 weeks pre-launch, 14 weeks post-launch) from January 17, 2025, to May 20, 2025.
• Groups:
  • Treatment Group: Users in 9 countries (e.g., Australia, Canada, Germany, US) who received the new Onebox design.
  • Convenience Comparison Group: Users in 7 countries/languages (e.g., Austria, Denmark, Sweden) with high internet penetration and similar legal frameworks who did not receive the new Onebox (receiving standard filtered results instead).
• Outcome Variables:
  1. Deterrence: Whether a user issued a subsequent CSAM-seeking query within the same search session after the initial query (binary outcome).
  2. Help-Seeking Engagement: The click-through rate (CTR) on hyperlinked buttons directing users to therapeutic resources.
  3. Traffic Impact: Interrupted time series analysis of aggregated traffic data provided by partner helplines to measure changes in website visits.
• Robustness Checks: The authors conducted robustness analyses, including restricting the sample to European countries only and comparing treated vs. untreated languages within the same countries, to ensure the validity of the parallel trends assumption.

Key Results

1. Reduction in Re-offending Behavior: The revised messaging resulted in a statistically significant 3.8 percentage point reduction in the rate of users issuing follow-on CSAM queries within the same session compared to the control group. In a robustness check limited to European countries, this reduction increased to 5.6 percentage points. The effect stabilized approximately seven weeks after launch.
2. Help-Seeking Engagement: Across all treated regions and modalities (click-to-call, website links), the average click-through rate to therapeutic resources was 0.73%. While numerically small, this represents approximately 5,000 sessions in the sampled dataset. In the UK, where a pre/post analysis was possible due to a prior version of the Onebox, CTR rose from 0.5% to 1.7% following the wording update.
3. Helpline Traffic: Interrupted time series analysis of helpline data from five regions (Australia, Canada, Germany, Ireland, UK) showed an immediate and significant jump in traffic attributable to the intervention. For example, Germany saw an increase of approximately 58 daily visitors, while Ireland saw an increase of roughly 1 daily visitor.

Significance and Claims
The paper claims to provide convergent evidence from large-scale, real-world behavioral data that digital interventions can effectively deter CSAM-seeking behavior and redirect users to help services.

• Deterrence Mechanism: The study suggests that shifting the messaging from a purely punitive/reporting focus to one that emphasizes consequences while offering a "way out" (therapeutic resources) creates a deterrent effect beyond the baseline established by previous warning messages.
• Feasibility of Help-Seeking: The results challenge the assumption that individuals seeking CSAM are unreachable; a non-trivial subset (0.73% of sessions) actively engaged with help resources when presented with low-barrier, anonymous options directly within the search interface.
• Policy Implications: The authors argue that this supports a socio-ecological approach to prevention, suggesting that ESPs have a unique capacity to deliver "just-in-time" interventions. They advocate for the expansion of such localized, therapeutic-focused interventions to more countries and languages, noting that early intervention can reduce victimization and alleviate strain on reporting ecosystems (e.g., NCMEC, IWF).

Limitations and Modesty
The authors explicitly acknowledge several limitations:

• Scope: The intervention only reaches users on the "open web" via Google Search, missing sophisticated users on the dark web or peer-to-peer networks.
• Measurement: The study measures within-session desistance, not necessarily long-term behavioral change or sustained abstinence. There is a possibility of displacement (users moving to other search engines or platforms).
• Data Constraints: The reliance on a convenience comparison group rather than a randomized controlled trial (RCT) introduces potential confounding variables (e.g., news events in Denmark affecting control group search patterns).
• Underestimation: The reported click-through rates likely underestimate total help-seeking, as they exclude phone calls and text messages initiated directly by users without clicking the UI button.

The paper concludes that while the effect size is modest, the intervention represents a viable, scalable component of a broader strategy to prevent child sexual abuse by addressing demand-side behaviors.