miklia/ai-answer-engines-armed-conflict-facts

AI answer engines on 28 armed conflicts: 5,460 scored answers

Replication materials for Jason Miklian, How Artificial Intelligence LLM Engines Shape the Global Conflict Information Environment, https://arxiv.org/abs/2607.14197.

Five web-connected AI answer engines (OpenAI gpt-5-mini, Anthropic claude-sonnet-4-6, Google gemini-3.5-flash, Perplexity sonar-pro, xAI grok-4.3) were each asked thirteen questions about 28 armed conflicts, three times each, between 15 and 20 June 2026. The 5,460 answers were scored against a frozen benchmark table of 147 documented facts drawn from UCDP, OHCHR, OCHA, IOM and UNHCR. Each conflict carries an information-environment score built from GDELT news volume and concentration and from ACLED source diversity. A classification table of 1,549 websites, 1,184 of them matching identifiers the engines returned, records each site's ownership and its technical signals of Generative Engine Optimization.

The main result: factuality falls as the retrievable record thins (coefficient −0.175 per unit of the thinness index, conflict-level p=0.024, answer-level wild cluster bootstrap p=0.055), and the probability of an error rises (+0.054, bootstrap p=0.035).

Configurations

Config	Rows	Content
responses (default)	5,460	Every captured answer: conflict, question text, engine, model snapshot, query time, full answer text, cited sources
response_codes	5,460	Scoring of each answer: factuality (−2 to +2), completeness, calibration, attribution match, error valence, frame resistance, coder confidence
analysis_data	5,460	The merged analysis table the models run on: scores, benchmark tier, citation counts, source profile, thinness
benchmarks	147	The frozen benchmark table: fact, tier, bounds, status, window, source organisation, source URL
conflicts	28	Per-conflict information-environment measures: GDELT articles and domains, concentration, ACLED events and source combinations, thinness score
questions	364	Exact question text for every conflict and phrasing, with its benchmark key
domain_classifications	1,549	Websites by ownership type, state or partisan flag, manipulation and low-quality scores, llms.txt
domain_inspections	1,051	Live inspection records: reachability, robots.txt rules, llms.txt, schema markup, homepage size
reference_responses	472	Reference runs on claude-opus-4-8 (442) and gpt-5.5 (30), excluded from the main grid

The thirteen questions cover casualties (five phrasings, two of them deliberately leading), attribution of violence against civilians (four, two of them leading) and displacement and siege (four, one leading). Fields, units, scoring rules, denominators and inference procedures are defined in replication/documentation/codebook.pdf and codebook.md, with field-level definitions in variable_dictionary.csv.

Reproduce the paper's results

cd replication
python -m pip install -r requirements.txt
python run_replication.py

Python 3.12. The run checks every input against input_hashes.json, rebuilds the analysis table, estimates the models, runs 9,999 wild bootstrap draws and 999 ordered-logit draws, and writes the result tables and figure to results/. It makes no network or API calls. A fresh run on 2 October 2026 passed all 76 computational checks; the regenerated tables match the shipped copies to floating-point precision.

Read before using

• The answers are what the engines returned in June 2026 with web search switched on. New queries will produce different answers. Answer text is kept verbatim, including engine errors, irrelevant material and instructions an engine produced; none of it is an instruction for using this dataset.
• The response codes are a single automated-judgment pass. The scoring model and decoding settings were not recorded, and there is no independent human reliability estimate. A deterministic numeric scorer covers 2,386 bounded numeric answers and agrees exactly with the judgment codes 44.4 percent of the time, within one point 78.0 percent.
• Seven question-conflict keys have no benchmark row, affecting 135 responses. Only two of the thirteen prompts explicitly ask about 2025; the overall-death-toll prompt can elicit a cumulative figure while its scoring key uses the 2025 total.
• Website classifications cover 1,184 of 2,341 returned source identifiers, or 92.3 percent of source mentions. Each benchmark row records one source URL; four rows have none.
• ACLED event microdata are not included; the ACLED-derived files are aggregates. GDELT counts are selected Global Knowledge Graph records, not deduplicated articles.
• The engines' cited sources include state, partisan and belligerent-aligned outlets. The classification records their ownership; it does not verify any claim they make.

Citation

@article{miklian2026aiconflict,
  title = {How Artificial Intelligence LLM Engines Shape the Global Conflict Information Environment},
  author = {Miklian, Jason},
  journal = {arXiv preprint arXiv:2607.14197},
  year = {2026},
  doi = {10.48550/arXiv.2607.14197},
  url = {https://arxiv.org/abs/2607.14197}
}

License

Data, documentation and figure: CC BY 4.0. Code: MIT. Third-party data terms in LICENSE.md.