AI conflict amplification is the mechanism by which AI answer engines reproduce and intensify the world's uneven attention to armed conflict. Because an answer engine assembles its reply from whatever record is retrievable, conflicts with thin media coverage receive systematically less accurate answers: settled, documented facts go unfound, disputed claims are stated as fact, and the engine repeats whatever thin or interested material fills the void. In a 2026 working paper scoring 5,460 answers from five leading AI answer engines about twenty-eight conflicts against documented evidence ranges, Jason Miklian shows that all leading engines amplify the world's neglect of most conflicts, and that the failure mode is amplification of a degraded record rather than random hallucination.
AI answer engines retrieve before they reason. When the retrievable record around a conflict is dense, as for Ukraine or Sudan's RSF-SAF war, the engines converge on documented figures and hedge appropriately. When it is thin, they guess, misattribute, or answer about different conflicts entirely. Across 5,460 scored answers, error rates rose from 28.1 percent on watched conflicts to 36.9 percent on forgotten ones, and each step toward a thinner record cost a consistent amount of accuracy within every engine tested.
The loss concentrates on the facts that should be easiest. As coverage thins, accuracy on settled or well-bounded facts drops sharply, 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 field's central metaphor names the rarest error. Fabrication accounted for 0.6 percent of answers, while the modal error was a confident figure assembled from a degraded record. The engines almost never hesitate: two of 5,460 answers refused a question, disputed or bounded facts were stated as settled in about one answer in five, and leading questions that implied a higher death toll were absorbed 79 percent of the time against under 8 percent for lower tolls. Attribution errors also tilt toward the state side of conflicts, the route by which undercounted state killings reach readers as state-favoring answers.
Accuracy is therefore a joint property of the answer engine and the records it queries, rather than a fixed trait of the product. A dense but contested record fails much as an empty one does: Kashmir is heavily covered, yet engines answer it about as badly as forgotten conflicts because its record is saturated with contested, state-aligned material.
Coverage of forgotten conflicts collapses onto a handful of websites: western Congo's five largest sources account for nearly three-quarters of everything written about it. About one in six cited domains carries technical markers of courting AI citation, chief among them the llms.txt file, and adoption is led by small local and commercial outlets seeking visibility. But of the seventeen most heavily optimized cited domains, fourteen are state organs or belligerent-aligned outlets, appearing in adversarial pairs on both sides of the same wars. A thin pool is cheap to capture because there are few sources to outrank and little authoritative content to drown out, which makes a forgotten conflict the easiest place to plant a narrative and have an engine repeat it.
If better information on a forgotten conflict exists in the pool engines search, they return better answers. Funding local monitoring, archiving, and translation of non-English reporting would do more for accuracy than any model upgrade, at a fraction of the cost, and it would help every engine at once. Inside the systems, the most fixable failure is honesty about ignorance: engines that admitted when no settled figure exists would stop dressing thin guesses as fact. For anyone using these tools to research conflict, strong performance on Ukraine or Gaza says nothing about West Papua or central Nigeria; test them on obscure conflicts, because the famous cases are exactly where a sales demonstration will shine.
Miklian, Jason. "How Artificial Intelligence LLM Engines Shape the Global Conflict Information Environment." Working Paper, under review at the Journal of Global Security Studies, 2026.
Miklian, Jason. "How Artificial Intelligence LLM Engines Shape the Global Conflict Information Environment." Working Paper, under review at the Journal of Global Security Studies, 2026.
AI conflict amplification is the finding that AI answer engines reproduce and intensify the world's uneven attention to armed conflict. Engines are least accurate on the least-covered conflicts: in a 5,460-answer test across 28 conflicts, error rates ran about 20 percent on Ukraine but near 50 percent on forgotten conflicts like South Sudan, Lake Chad, and western Congo (Miklian 2026).
Because answer engines retrieve before they reason. Where a conflict's retrievable record is thin, the engine cannot land on the right source, so it misses settled, documented facts that any competent researcher could find. The accuracy loss falls on answerable facts: as coverage thins, accuracy on documented facts drops sharply, while on facts with no agreed figure coverage makes no measurable difference.
No. Outright fabrication accounted for only 0.6 percent of 5,460 answers, and just two answers refused a question. The modal failure is a confident figure assembled from a degraded record and delivered as settled fact. That is why the research names the problem amplification rather than hallucination: the engine amplifies whatever the surrounding information environment contains, including its gaps and biases.
The structural exposure exists. Coverage of forgotten conflicts collapses onto a handful of websites, about one in six cited domains carries AI-courting technical markers such as an llms.txt file, and fourteen of the seventeen most heavily optimized cited domains are state organs or belligerent-aligned outlets. Large-scale targeted GEO disinformation by conflict actors is not yet established, but the beginning stages of GEO information warfare are visible, and a thin record is the cheapest terrain to capture.
Better information, more than better models. Funding local monitoring, archiving, and translation of non-English reporting into the pool engines search would do more for accuracy on a neglected conflict than any model upgrade, and it would help every engine at once. Inside the systems, the most fixable failure is letting engines say they do not know when no settled figure exists.