Where does AI assistance become unreliable in research?

The roadmap's first finding is that AI capability is not uniformly distributed across research work — it is sharply stage-dependent. Where tasks are structured, externally checkable, and tool-mediated (literature retrieval, drafting, figure generation, review support), AI is reliable. Where tasks demand genuine novelty, implicit domain knowledge, long-horizon reasoning, or scientific judgment (open-ended ideation, research-level experiments), capability drops sharply and autonomy becomes unreliable.

This is more useful than a blanket "AI is/isn't good at research" claim because it predicts where to draw the human-machine boundary rather than whether to draw one. The survey documents the failure pattern concretely: generated ideas often degrade after implementation, research code lags far behind pattern-matching benchmarks, and end-to-end autonomous systems have not consistently reached major-venue acceptance standards.

The counterpoint is that the boundary moves — yesterday's "unreliable autonomy" zone (e.g. coding) keeps shrinking. But the boundary's shape is stable even as it shifts: it always tracks checkability. Tasks with an external oracle to verify against fall on the reliable side; tasks requiring judgment with no ground truth stay on the unreliable side. Therefore the design principle is durable even though the specific task assignments are not — which is why this pairs naturally with the lifecycle verification gap: the boundary is exactly the line where verification becomes impossible.