LLMs hallucinate most where source records are thinnest

Hallucination rates across 5,460 answers to questions about 28 conflicts follow a clear gradient, according to a study posted to arXiv. The thinner the retrievable source record on a given conflict, the more five leading answer engines invented, misattributed, or miscounted. The finding is not a general warning about LLM unreliability. It is a structural map of where unreliability concentrates, and that map has direct consequences for any organisation whose mandate or reputation depends on how AI answers describe it.

The researchers asked a battery of questions about 28 conflicts across five answer engines, generating 5,460 scored responses. Accuracy degraded in proportion to the sparseness of the underlying source record. Well-documented conflicts, those with thick corpora of news coverage, academic literature, and institutional reporting, returned more accurate answers. Conflicts with thin records returned more invented ones. The mechanism is straightforward: models trained on retrieved content can only be as accurate as the content they retrieve. When the retrieval pool is shallow, the model fills gaps with statistical inference rather than evidence.

Where thin records become a security problem

The more consequential finding is not the hallucination itself but the exploitation pathway it opens. The study identifies what it calls structural exposure to Generative Engine Optimization. Thin source records are not merely incomplete; they are cheap to warp. A small volume of strategically placed, well-structured content can shift how an engine characterises a conflict, an institution, or a policy position, because there is so little competing material to push back.

This is GEO as information warfare, and the arXiv authors are right to treat it as a category distinct from ordinary SEO manipulation. Traditional search manipulation requires displacing high-authority incumbents. LLM manipulation in thin-record environments requires only filling a vacuum. The barrier is far lower, and the output, a confident prose answer rather than a ranked list of links, is far harder for a casual reader to interrogate.

The implication for multilateral institutions and policy bodies is serious. Organisations such as UNDRR, the World Bank, or the various UN agencies working on conflict, climate, or health crises are frequently asked about by analysts, journalists, and decision-makers who now reach for ChatGPT or Perplexity before they reach for a library. If those organisations have thin, fragmented, or inconsistent digital source records, particularly around their specific programmes, field operations, or data methodologies, they become exactly the kind of subject where models invent and adversaries can cheaply intervene.

The same logic applies to industrial groups and financial institutions operating in contested or under-reported geographies. A major infrastructure company working in a conflict-adjacent region, a development finance institution funding projects in fragile states, a philanthropic body whose grant-making touches politically sensitive terrain: each of these has domain-specific source records that are, in all likelihood, thinner than their communications teams assume. The test is not whether the organisation is well-known in aggregate. The test is whether specific claims that models are likely to make about specific programmes or positions are corroborated by sufficient, structured, retrievable content.