AI SEARCH AGENTS NEED TO ASK, NOT ASSUME

DiscoBench, a new evaluation framework, exposes a critical flaw in current AI search behavior. Models that search repeatedly without asking follow-up questions achieve only 51.9 percent accuracy—worse than simply guessing. Even top-performing models reach just 43 percent overall accuracy. The solution is straightforward: when ambiguity exists, ask. Removing ambiguity from queries boosts accuracy by up to 40 percentage points, demonstrating that clarification is the missing piece. This finding shifts focus from raw search capability to conversational intelligence. Rather than engineering better search algorithms, developers need AI agents that recognize uncertainty and engage users in dialogue. The benchmark suggests that acknowledging what you don't know—and asking about it—matters more than aggressive information retrieval in multi-step research tasks.