Key claims
-
Radicalization is a process, not a binary choice. Exposure to counter-narratives matters most during the early questioning phase, before ideology hardens. A 23-year-old ISIS recruit defector in Iraq reported he would have changed his mind if he had been exposed to disconfirming information before joining, but was resistant once fully indoctrinated (within 6 months of arrival).
-
ISIS's recruitment success derives from understanding human psychology, not technological superiority. ISIS translates propaganda into 14+ languages (English, Arabic, German, Russian, French, Turkish, Kurdish, Hebrew, Mandarin, sign language) because they understand what drives different audiences. A 13-year-old British schoolgirl who attempted to join ISIS thought she was joining an "Islamic Disney World" where she'd meet a "jihadi Brad Pitt"—ISIS carefully crafted emotional appeals to her desires for belonging, romance, and adventure.
-
Counter-narratives from credible voices are underutilized. Defectors, local videographers, and peaceful religious scholars have authentic, emotionally compelling counter-narratives that outweigh recruitment propaganda, but they lack the marketing reach of extremist organizations. They often risk their lives to speak up yet fail to reach vulnerable audiences.
-
Online advertising can bridge the gap between at-risk users and credible counter-voices. The Redirect Method uses behavioral targeting (based on search queries, not profiles) to serve counter-propaganda ads to users actively seeking extremist content. This respects privacy while reaching the highest-intent audience.
-
Online harassment works through psychological pressure, not just threats. Harassers learn what makes targets angry, afraid, or insecure, then weaponize that. When harassment goes unchecked, it silences speakers, collapses comment sections, and forces remaining online spaces into ideologically homogeneous echo chambers—which then become vulnerable to disinformation and polarization.
-
AI can predict emotional impact of language at scale. Machine-learning models trained on human judgments can distinguish whether a comment (e.g., "break a leg") is encouraging or threatening, capturing nuance humans recognize instantly. This enables content moderation at scale and real-time writer feedback on emotional impact.
-
Empathy is scalable through technology. If AI can understand emotional impact, it can give commenters real-time feedback as they type—mimicking how facial expressions in face-to-face conversation allow adjustment before harm occurs.
Solutions presented
The Redirect Method — Partnership with Moonshot CVE (2016). Uses search-based ad targeting to reach users searching for extremist keywords and invite them to watch counter-narrative videos (defectors, clerics, local footage). Pilot (8 weeks, English + Arabic): reached 300K people with jihadist sympathy. Now global, targeting Islamist, white supremacist, and other extremist groups.
Perspective — Collaboration with Google's Counter Abuse team, Wikipedia, and the New York Times. Machine-learning models predict which comments will make readers leave the conversation. Real-time feedback for commenters; improved moderation efficiency. NYT result: comment-enabled articles increased from 10% to 30%.
Notes
This talk exemplifies the field's shift from purely technical content moderation (flagging/removing violating posts) toward psychological and behavioral approaches. Green's insight—that extremists and harassers exploit human vulnerabilities rather than technological gaps—aligns with research on motivated reasoning, in-group identity, and moral psychology. The Redirect Method's emphasis on early intervention during the "questioning phase" of radicalization echoes de-radicalization research showing that exit moments from extremist communities correlate with exposure to alternative worldviews and community reintegration pathways.
The Perspective tool demonstrates a pragmatic application of NLP to a genuine problem: online toxicity drives away speakers, particularly marginalized groups who face disproportionate targeting, thereby reducing comment diversity and enabling information bubbles. However, the talk notably does NOT address the algorithmic recommendation systems that concentrate extreme content in the first place—Perspective treats toxicity as an isolated content-level problem rather than a symptom of platform-wide incentive misalignment (a critique articulated by Wardle and others in the moderation literature).
Both interventions focus on the supply side (intercepting radicalization pathways, reducing harasser effectiveness) rather than demand-side interventions (inoculation against extremist recruitment, building resilience to harassment). This reflects Jigsaw's institutional positioning within a tech company: technical interventions deployable at platform scale. The field increasingly recognizes that psychological and community-based interventions (offline mentorship, family reintegration, social support networks) are often more effective than technology alone, but scale is a fundamental constraint.