Overview
This series separates two questions about AI research agents: whether their ideas broaden beyond seed literature and whether experiments actually ran. It traces exploration breadth, seed distance, frontier alignment, and question-versus-method novelty, then examines execution logs, synthetic results, and claim scope. The supplied studies show local concentration and unverified outputs in tested settings, but do not prove agents cannot make important discoveries or validate a finished two-gate architecture.
Learning outcomes
- In the supplied studies, AI-generated ideas are more concentrated and closer to seed literature than the human comparisons.
- Novelty instructions and elaborate orchestration do not by themselves demonstrate broad exploration.
- A paper cannot turn a failed or unverified run into empirical evidence; execution provenance must support the claim.
- A two-gate model is a useful synthesis, not an architecture validated directly by the supplied studies.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What does an exploration gate test in an AI research agent?
It tests whether a collection of ideas spreads beyond its seed literature and connects with later directions, rather than merely sounding novel or being large.
Why is a novel method not automatically a new research question?
A method can change how an established problem is studied while leaving the problem unchanged; the supplied study found 90.4% method novelty versus 10.5% new questions under its coding scheme.
What evidence is needed before an AI research paper can make an empirical claim?
The claim needs a traceable execution record showing what ran, what outputs were produced, and how those outputs support the analysis; fluent prose is not enough.
Does the evidence prove that AI research agents cannot make important discoveries?
No. The studies report narrower, locally anchored ideation and unverified results in tested settings, but they do not show that every AI idea is derivative or that agents cannot make important discoveries.