{
    "@context": "https://schema.org",
    "@type": "ScholarlyArticle",
    "@id": "https://miklian.org/papers/how-ai-llm-engines-shape-the-global-conflict-information-environment#article",
    "headline": "How Artificial Intelligence LLM Engines Shape the Global Conflict Information Environment",
    "name": "How Artificial Intelligence LLM Engines Shape the Global Conflict Information Environment",
    "author": [
        {
            "@type": "Person",
            "name": "Jason Miklian",
            "@id": "https://miklian.org/#person",
            "sameAs": [
                "https://orcid.org/0000-0003-1227-0975",
                "https://scholar.google.com/citations?user=RHlevGEAAAAJ&hl=en",
                "https://www.researchgate.net/profile/Jason-Miklian",
                "https://www.wikidata.org/wiki/Q47107618",
                "https://en.wikipedia.org/wiki/Jason_Miklian",
                "https://www.globe.uio.no/english/people/aca/jasontm/",
                "https://www.prio.org/people/5833",
                "https://jasonmiklian.com"
            ],
            "affiliation": {
                "@type": "Organization",
                "name": "Centre for Global Sustainability, University of Oslo",
                "url": "https://www.globe.uio.no/"
            }
        }
    ],
    "datePublished": "2026",
    "dateModified": "2026-07-16",
    "creativeWorkStatus": "Working paper",
    "isPartOf": {
        "@type": "CreativeWorkSeries",
        "name": "Working Paper"
    },
    "abstract": "Ask an AI answer engine about Ukraine and it returns the documented record. Ask about a forgotten conflict and it guesses. This article tests how five leading AI LLM answer engines (OpenAI, Anthropic, Gemini, Grok, and Perplexity) handle armed conflict facts, asking each a thirteen-question battery about twenty-eight conflicts, repeated three times, for 5,460 answers scored against documented evidence ranges from UCDP, OHCHR, IOM, and UNHCR. Three findings emerge. First, the engines are least accurate on the least-covered conflicts: error rates run about 20 percent on Ukraine but near 50 percent on forgotten conflicts like South Sudan, Lake Chad, and western Congo, and the accuracy loss falls on settled, documented facts rather than on unanswerable questions. Second, the engines almost never hesitate. Just two of 5,460 answers refused a question, disputed or bounded facts are stated as settled in about one answer in five, and directional attribution errors favor the state side of conflicts. Third, the thinnest conflict records are the most structurally exposed to Generative Engine Optimization: coverage of forgotten conflicts collapses onto a handful of websites, about one in six cited domains carries AI-courting technical markers, and the heaviest optimizers include state organs and belligerent-aligned outlets. All leading AI LLMs amplify the world's neglect of most conflicts, delivering confident answers assembled from degraded records. The failure mode is amplification rather than hallucination, and repairing it means funding the information commons of forgotten conflicts rather than waiting for better models.",
    "keywords": [
        "artificial intelligence",
        "large language models",
        "AI answer engines",
        "armed conflict",
        "forgotten conflicts",
        "generative engine optimization",
        "state-based disinformation",
        "conflict data",
        "data voids"
    ],
    "about": [
        "artificial intelligence",
        "large language models",
        "armed conflict",
        "forgotten conflicts",
        "generative engine optimization",
        "disinformation"
    ],
    "url": "https://miklian.org/papers/how-ai-llm-engines-shape-the-global-conflict-information-environment",
    "mainEntityOfPage": "https://miklian.org/papers/how-ai-llm-engines-shape-the-global-conflict-information-environment",
    "inLanguage": "en",
    "isAccessibleForFree": true,
    "license": "https://creativecommons.org/licenses/by/4.0/",
    "encoding": [
        {
            "@type": "MediaObject",
            "encodingFormat": "text/markdown",
            "contentUrl": "https://miklian.org/papers/how-ai-llm-engines-shape-the-global-conflict-information-environment.md"
        },
        {
            "@type": "MediaObject",
            "encodingFormat": "text/html",
            "contentUrl": "https://miklian.org/papers/how-ai-llm-engines-shape-the-global-conflict-information-environment"
        }
    ],
    "mentions": [
        {
            "@type": "DefinedTerm",
            "name": "AI Conflict Amplification",
            "url": "https://miklian.org/concepts/ai-conflict-amplification",
            "inDefinedTermSet": "https://miklian.org/concepts/"
        }
    ],
    "subjectOf": [
        {
            "@type": "WebPage",
            "url": "https://miklian.org/ai-governance",
            "name": "AI, Democracy, and Society"
        },
        {
            "@type": "WebPage",
            "url": "https://miklian.org/fragile-states",
            "name": "Fragile States and Conflict"
        }
    ],
    "keyMessages": [
        "Across 5,460 answers about 28 conflicts, all five leading AI answer engines lose accuracy as a conflict's information record thins: error rates rise from 28.1 percent on watched conflicts to 36.9 percent on forgotten ones, and mean accuracy falls from about +1.1 to about +0.8 on a five-point scale. The gradient holds at the conflict level (p=0.024) and within every engine.",
        "The engines miss settled, documented facts precisely where coverage is thin. Accuracy drops sharply on facts with a documented answer as coverage thins (slope −0.259, p=0.001), while on facts with no agreed figure coverage makes no measurable difference. The answers exist and could be retrieved by any competent researcher; the engines cannot find them because the surrounding record is too sparse to retrieve from.",
        "The engines almost never hesitate. Only two of 5,460 answers refused a question, disputed or bounded facts are stated as settled in 18.5 percent of answers, and when a leading question implied a death toll was higher than documented the engines went along 79 percent of the time, against under 8 percent when it implied a lower toll. Engines can easily be talked up on death figures and almost never talked down.",
        "Attribution errors tilt toward the state: of the 390 errors that favor a side, 61 percent favor the state side of the conflict (binomial p=0.0001), which is the route by which undercounted state killings reach readers as state-favoring answers.",
        "Forgotten conflicts are the most structurally exposed to Generative Engine Optimization. Their coverage collapses onto a handful of websites (western Congo's five largest sources account for nearly three-quarters of everything written), about one in six cited domains publishes an llms.txt file courting AI citation, and fourteen of the seventeen most heavily optimized domains are state organs or belligerent-aligned outlets. The beginning stages of GEO information warfare are visible.",
        "Kashmir shows a second failure mode: a dense record saturated with contested, state-aligned material degrades answers much as absence does. Volume without verification does not protect accuracy."
    ]
}