Generative AI enables non-experts to rapidly produce papers, code, figures, repositories, quality-assurance documents, and submission metadata. This capacity can widen access to knowledge production, but it also creates a risk that artifact completeness is mistaken for domain validity. This manuscript presents a reflexive case study of two AI-assisted research attempts conducted by the author: a lithium-ion battery-model audit and an A10-derived Hubble-tension scaffold audit. In the battery case, an initially technology-oriented project was reduced to a residual-drift audit of simplified PyBaMM models against external discharge time series. Claims about safe charging, practical battery-management-system readiness, degradation reduction, lithium-plating avoidance, and real-cell validation were explicitly prohibited. In the Hubble case, a cosmological scaffold failed to establish a first-principles physical E(z)/H(z) bridge and therefore stopped before real-data likelihoods, MCMC, posterior comparison, or evidence evaluation. Across two distinct domains, the same pattern appeared: AI inflated research artifacts, but strict grounding checks reduced them to claim-boundary and failure-boundary records. This manuscript does not argue that AI-assisted research is useless, nor does it evaluate expert-led AI science. It argues that non-expert AI-assisted research requires explicit domain-grounding, prior-art, external-data, implementation-or-experiment, expert-review-worthiness, and public-claim-boundary gates before public claims are made.