Mapping Information Warfare: The Syrian Coastal Massacres
A first-of-its-kind AI-powered conflict analysis pilot: mapping a conflict narrative at scale and speed, transforming months of manual coding into days of computational analysis.
Lebanon Office
In March 2025, massacres erupted on Syria's coast, with thousands of the Alawite minority falling victim to sectarian killings. Within hours, social media fragmented into competing realities: blame split along sectarian lines, hate speech flooded platforms, casualty figures ranged from 1 to 2,000,000. Real violence became a strategic instrument of information warfare.
BrainBridge was contracted to pilot a first-of-its-kind AI-powered conflict analysis tool, mapping conflict narrative at scale and speed, transforming months of manual coding into hours of computational analysis while revealing patterns invisible to conventional methods.
We scraped 100,000 posts and comments from YouTube, Facebook, and Twitter. After filtering for conflict-relevance, we built a knowledge graph modelling the entire narrative ecosystem.
The Findings, Visualised
Actor polarisation: which narrative battles are settled, and which are still fluid
Blame/support ratio across 512 actors · 0.0 = uncontested support · 1.0 = uncontested blame
HTS is the only narrative battle still open: 64.8% blamed, 35.2% supported. Evidence introduced now can still shift which account consolidates. Counter-messaging on Assad's settled narrative (0.974) would be wasted resources.
Hate speech is the dominant mode of discourse, not the margin
Automated classification across the full 100,000-post corpus
The first quantified baseline for this conflict's discourse, turning "hate speech is prevalent" into a measurable threshold that can trigger escalation alerts.
188 competing casualty claims: variance as a weapon
All death-toll claims tracked simultaneously · logarithmic scale
High claims correlate with sectarian framing (r=0.64) and hate speech (r=0.58): inflation is strategic, not random error. The variance itself prevents verification by design, until it is tracked systematically.
Five Technical Innovations
Actor Polarization Scoring Algorithm
Automated calculation of blame/support ratios across 512 actors, producing 0.0–1.0 polarisation scores. HTS: 0.296 (the only contested actor: 64.8% blamed, 35.2% supported). Assad: 0.974 (near-universal blame). Israel: 0.167 (weaponised by both sides). Reveals which narrative battles are still fluid versus settled, and where evidence could still shift outcomes.
Emergent Pattern Detection via Machine Learning
AI auto-detected 524 DiscoursePattern nodes and 373 NarrativeType nodes through structural co-occurrence analysis, not pre-defined by researchers. 22 distinct variants of "Sectarian Hate Campaign" emerged. Proves information warfare operates through structure, not just content, separating spontaneous rage from coordinated strategy.
Systematic Hate Speech Detection at Scale
Automated classification flagged 71.7% of posts for hate speech across the entire corpus (combined with 93.9% sectarian/aggressive tone). Hate speech is the dominant mode, not marginal. Establishes a quantifiable baseline for threshold-based monitoring: >80% = high escalation alert.
Casualty Claim Variance Tracking
Tracked all 188 competing death toll claims simultaneously, ranging from 1 to 2,000,000 deaths (median: 300). High claims correlate with sectarian framing (r=0.64) and hate speech (r=0.58). Proves inflation is strategic, not random error. Variance prevents verification by design.
Multi-Dimensional Network Graph
8 node types × 12 relationship types encoding the full discourse structure. 78% geographic concentration on 4 real massacre locations, yet 71.7% of posts about those locations contain hate speech. Real violence and systematic toxic weaponisation rendered simultaneously visible: information warfare architecture made legible.
What the Project Delivered
Chaos became intelligence in days, not months
Decision-ready intelligence in 10 days: 71.7% hate speech prevalence, HTS identified as the only contested actor, 22 distinct hate-campaign variants. Policymakers can now be briefed with quantified evidence, not qualitative impressions.
Information warfare structure made visible
Analysts could observe "sectarian rhetoric is prevalent" but couldn't prove coordination. The graph revealed 524 AI-detected patterns and the correlation between casualty inflation and sectarian framing (r=0.64). This is strategic deployment, not spontaneous rage.
The one narrative battle still fluid
HTS polarisation of 0.296 means neither blame (64.8%) nor support (35.2%) has won. A strategic intervention window: evidence about HTS's actual role could still shift which narrative consolidates, while counter-messaging on settled narratives (Assad: 0.974) would be wasted.
A replicable early-warning capability
71.7% hate speech + 93.9% toxic tone establishes a quantified escalation-risk baseline where none existed. Future monitoring can trigger alerts when hate speech crosses defined thresholds.
From reactive analysis to 7–14 days advance notice
Scaled to real-time monitoring, the methodology enables hashtag-cascade detection within 24 hours of a campaign launching, geographic-shift alerts, and actor role-reversal detection. Replicated across conflicts (Ukraine, Gaza, Lebanon, Yemen), it supports cross-conflict pattern recognition and predictive modelling: moving early warning from "this looks bad" to "72% probability of escalation within 14 days." Systematic documentation of hate-speech deployment also builds a chain of evidence usable in accountability proceedings.
Explore the Knowledge Graph
Interact with the live data: search actors, explore narrative clusters, and trace relationships across 3,087 nodes and 22,376 relationships. Use the panels within the dashboard to navigate. Contact us for free access to the full graph.
Questions About This Project
What did the project analyse?
100,000 posts and comments scraped from YouTube, Facebook, and Twitter following the March 2025 coastal massacres. After conflict-relevance filtering, the corpus was modelled as a knowledge graph of 3,087 nodes and 22,376 relationships covering actors, events, narratives, locations, and discourse patterns.
How fast was it compared to traditional analysis?
Ten days to decision-ready intelligence, versus 12–24 months for traditional manual coding: an order-of-magnitude cost reduction.
Can the methodology be applied to other conflicts or domains?
Yes. It is a demonstrator of one way of doing the work; the same approach extends to other conflicts and to adjacent domains such as legitimacy risk and narrative-driven market risk. The wider ambition, pending funding, is real-time monitoring of entire information ecosystems.
Who gets access to the full graph?
The embedded dashboard above is public. Full access (including query capability across the complete dataset) is available to collaborators, researchers, and stakeholders on request. Contact us to discuss access, extensions, or partnerships.