- Information warfare operates through structure, not just content, which is why it must be modelled as a network, not a stream of posts.
- Four metric families make it measurable: polarisation scores, prevalence baselines, emergent pattern detection, and claim-variance tracking.
- Quantification revealed strategy where observers saw chaos: casualty inflation correlated with sectarian framing (r=0.64).
- Measurement converts information warfare from an invisible advantage into an observable, counterable phenomenon.
The problem: a war you can see but not measure
When massacres erupted on Syria's coast in March 2025, the violence was immediately followed by a second conflict, fought in feeds and comment sections. Blame split along sectarian lines. Hate speech flooded platforms. Casualty figures ranged from 1 to 2,000,000. Every observer could see that information warfare was underway. No one could measure it.
That gap matters. Unmeasured, information warfare is deniable ("it's just organic anger"), unactionable ("what exactly would we counter?"), and self-concealing (the noise it generates is precisely what defeats verification). The pilot BrainBridge conducted for the Konrad Adenauer Stiftung (documented in full in the case study) was designed to close that gap.
The method: model the ecosystem, not the posts
The core methodological decision was to treat the discourse as a relational system rather than a pile of texts. From 100,000 scraped posts and comments (YouTube, Facebook, Twitter), filtered for conflict-relevance, we built a knowledge graph: 3,087 nodes across 8 types (events, massacres, actors, discourse patterns, narrative types, hashtags, locations) connected by 22,376 relationships across 12 types.
Why a graph? Because the analytically decisive properties of information warfare (coordination, propagation, convergence) are properties of relationships, not of individual posts. A single hateful comment is noise. Twenty-two structurally distinct variants of a sectarian hate campaign, detected through co-occurrence analysis rather than researcher-defined categories, is architecture.
Four families of measurement
1. Polarisation scoring: which battles are still fluid
Automated blame/support ratios across 512 actors produced 0.0–1.0 polarisation scores. Assad scored 0.974: near-universal blame, a settled narrative. HTS scored 0.296, the only genuinely contested actor (64.8% blamed, 35.2% supported). This single metric carries strategic weight: counter-messaging on settled narratives is wasted money, while contested narratives are where evidence can still shift outcomes.
2. Prevalence baselines: from impression to threshold
Hate-speech classification across the full corpus found 71.7% prevalence (93.9% sectarian/aggressive tone). The point is not the alarming number; it is the baseline. "Hate speech is prevalent" becomes "71.7%, up from 60% last week: high escalation alert." Impressions cannot trigger protocols. Thresholds can.
3. Emergent pattern detection: structure reveals strategy
Machine learning auto-detected 524 discourse patterns and 373 narrative types, none pre-defined by researchers. The patterns separated spontaneous rage from coordinated deployment, the distinction that matters most for attribution and response.
4. Claim-variance tracking: weaponised uncertainty, quantified
All 188 competing death-toll claims were tracked simultaneously (range: 1 to 2,000,000; median: 300). High claims correlated with sectarian framing (r=0.64) and hate speech (r=0.58), statistical evidence that casualty inflation was strategic, not random error. The variance itself is the weapon: it prevents verification by design. Tracked systematically, it becomes evidence instead.
What measurement changes
The deliverable was not a report describing chaos; it was a set of numbers with consequences:
- Briefings carry quantified evidence. Policymakers can now cite polarisation scores and prevalence rates no other methodology produces.
- Coordination became provable. Before: analysts could observe sectarian rhetoric but not demonstrate orchestration. After: 524 detected patterns and inflation-framing correlations prove information warfare is operational, not latent.
- Intervention windows became visible. The HTS narrative battle is still fluid, meaning evidence introduced now can still shape which account consolidates.
- A replicable early-warning capability exists. Quantified baselines support threshold-based monitoring; scaled, the methodology targets 7–14 days advance notice of escalation, hashtag-cascade detection within 24 hours, and cross-conflict pattern recognition (Ukraine, Gaza, Lebanon, Yemen).
The longer-term implication is the most important: information warfare loses its invisibility advantage when it becomes measurable. Systematic documentation of hate-speech deployment builds chains of evidence usable in accountability proceedings; variance data gives fact-checkers systematic ammunition instead of ad-hoc debunking; and policy decisions can rest on what discourse reveals rather than which narrative shouted loudest.
Lessons for any organisation facing narrative threats
The Syria pilot was a conflict case, but the method generalises. Boycott campaigns, coordinated reputation attacks, and regulatory backlash all share the same anatomy: structured narrative deployment masquerading as organic sentiment. The questions worth asking are identical: Which narratives are converging? Is the activity coordinated or spontaneous? Which battles are settled and which are still fluid? Those are measurement questions, and they are answerable. See our Narrative & Legitimacy Intelligence and Strategic Market & Geopolitical Intelligence programmes for how this is applied.
Frequently asked questions
Is the knowledge graph publicly accessible?
Yes: an interactive dashboard is embedded in the case study page, covering all 3,087 nodes and 22,376 relationships. Full query access is available to collaborators and researchers on request.
Could this methodology run in real time?
That is the design goal of the next phase: continuous monitoring with hashtag-cascade detection within 24 hours of campaign launch, geographic-shift alerts, and actor role-reversal detection, providing 7–14 days advance notice of escalation rather than retrospective analysis.
How are the AI classifications validated?
Through expert review, the hybrid model BrainBridge is built on. AI surfaces patterns at scale; domain experts validate, contextualise, and trace the reasoning, keeping the analysis both scalable and defensible. More on this in The Third Intelligence.
