Contradictory wiki data won't make AI agents hallucinate — but it will break them

A developer built a controlled testbed to examine how ingest quality affects an AI agent that navigates a wiki using search and read tools, rather than relying on vector databases or RAG. The agent scored 4 out of 4 on questions when given a clean wiki, but its correct-answer rate dropped to near zero when contradictory or stale duplicate pages were introduced. Crucially, the agent did not hallucinate confidently wrong answers — instead, it detected conflicting information and refused to commit to a single response. The real cost of poor ingest quality is therefore not misinformation but a collapse in the wiki's authority, forcing the agent to read multiple pages, hedge its answers, and consume more resources per query. The researcher concluded that this failure mode is easy to overlook because no single answer is technically wrong, yet the system's core value — delivering one trusted answer — is effectively destroyed.