2026 Lebanon Conflict Misinformation Study
Data and analysis code for a descriptive computational study of fact-checked misinformation claims related to the 2026 Israel-Hezbollah/Lebanon conflict. The study classifies 142 unique Arabic- and English-language fact-checked claims by manipulation type (authentic content reused/recontextualized vs. AI-generated/AI-manipulated), reports inter-rater reliability on that classification, tests the sensitivity of the main distribution to the corpus's reliance on one fact-checking organization, and reports whether any claims were identified as having penetrated professional news coverage.
Full methodology, taxonomy definitions, and coding rules are in
docs/Codebook_v1.1.docx and supplementary/Supplementary_Material.docx.
Repository layout
data/ Final coded dataset (CSV) + the complete frozen workbook
scripts/ Analysis pipeline (see below)
figures/ Figures 1-3 (SVG)
supplementary/ Supplementary material referenced in the paper
docs/ Study codebook (taxonomy, coding rules, outlet roster)
Data
| File | Rows | Description |
|---|---|---|
data/claims.csv |
142 | One row per unique claim: normalized description, media type, manipulation-type code, observed language(s). |
data/factchecks.csv |
152 | One row per fact-check article (a claim can have more than one). |
data/newsroom_search_log.csv |
142 | One row per claim: newsroom-penetration search method, date, and result. |
data/newsroom_coverage.csv |
3 | The professional-news articles identified as category-2 penetration, with outlet, URL, and correction details. |
data/irr_sample.csv |
28 | The blinded inter-rater reliability sample: both coders' manipulation-type labels. |
data/full_dataset_with_triage_and_correction_log.xlsx |
— | The complete frozen workbook, including the pre-triage candidate pool, exclusion decisions, and a log of one post-freeze correction. Included for full transparency; claims.csv/factchecks.csv above are the derived tables actually used in the analysis. |
What isn't fully scriptable. Turning retrieved candidates into the final
142-claim corpus involved human coding: deciding topical eligibility,
merging multiple fact-checks of the same underlying claim, and assigning a
manipulation-type code from a direct reading of each fact-check (Codebook
Sections 1.3-1.4, 4.1, 4.6). This is standard for content-analysis research
and is not a bug in the pipeline — 01_collect_candidates.py reproduces the
retrieval step, and the coded output of the human-judgment step is what's
provided directly in data/.
Scripts
Run from inside scripts/. Requires Python 3.10+ and pip install -r ../requirements.txt.
01_collect_candidates.py— Retrieves candidate fact-checks from the Google Fact Check Tools API and Misbar's listing API. Needs a Google Fact Check API key (copy.env.exampleto.envin the repo root and add your own). Output will differ from the frozen corpus over time, since it depends on live API/site state and does not perform the human-coding step described above.02_statistical_analysis.py— Reproduces Table 1 and every descriptive statistic in the Results section (manipulation-type distribution with Wilson 95% CIs, the three Misbar sensitivity analyses, newsroom-penetration rates, source composition) directly from the CSVs indata/. No API key needed.03_irr_analysis.py— Reproduces the inter-rater reliability results (raw agreement and Cohen's kappa, fine-grained and grouped) fromdata/irr_sample.csv.04_generate_figures.py— Regenerates Figures 1-3 as SVG directly fromdata/claims.csvanddata/factchecks.csv.
Each script's output can be checked against the numbers reported in the paper.
Manipulation-type taxonomy (summary)
- AI-generated — wholly synthetic image, video, or audio.
- AI-manipulated — real media digitally altered with AI tools.
- recycled old content — genuine media from a real, different event or date, presented as current.
- false context / wrong location — genuine, contemporaneous media misattributed to the wrong place or actor.
- fabricated quote / fabricated document / other — no manipulated media involved.
For the primary analysis these are grouped into authentic-content
misuse, AI-related, and other (mutually exclusive, by primary
mechanism — see docs/Codebook_v1.1.docx Section 4.1 and the Methods
section of the paper). Full definitions and the complete
major-international-outlet roster used for the newsroom-penetration
analysis are in supplementary/Supplementary_Material.docx.
Status
This repository accompanies a manuscript currently in preparation. Citation details and a license will be added on publication.