Large language models produce fluent but factually unreliable statements, and the mainstream routes to factuality—retrieval augmentation and alignment fine-tuning—presuppose that authoritative external knowledge is obtainable. When that assumption fails, a distinct family applies: adversarial interaction, in which a model cross-examines another model’s claims, several agents debate an answer, or a skeptical prompt interrogates its own reasoning. This review surveys representative adversarial factuality mechanisms published from 2022 to 2026 and organizes them into a two-level taxonomy, partitioned by evidential basis and adversarial structure. Two findings stand out. First, the surveyed mechanisms populate the design space unevenly, no structure is uniformly superior, and a single failure mode recurs across structures: false consensus—premature convergence on a shared but incorrect answer. Second, the field’s diagnosis of false consensus is mature while its countermeasures remain almost entirely unverified, and evaluation is biased against internal-consistency methods because established benchmarks presuppose external ground truth. The review closes with concrete directions: standard evaluation protocols, a verification convention for countermeasures, conformity-aware mechanism design, cost reporting, and a game-theoretic formalization of adversarial examination.