Gate the claim, not the tool: verification for AI-accelerated research in nutrition

Artificial intelligence (AI) and agentic workflows can accelerate nutrition science across living systematic reviews, dietary assessment and cohort analyses, yet they have also fuelled a surge of formulaic papers mined from open datasets. Journals and funders have responded largely by restricting how AI may be used, especially in writing. Because AI use leaves no reliable trace in a single paper, such rules cannot be enforced; they drive use out of sight, penalize honest authors who disclose it, and fall out of date within months as models improve. A gate works only when what it checks is visible at submission. When three publishers stopped accepting unvalidated papers built on the US National Health and Nutrition Examination Survey (NHANES), their NHANES publications fell by 72% in the first nine months of 2026, showing that visible gates work, though a rule that blunt can also turn away sound epidemiology. Rather than policing the tool, nutrition science should use AI openly under human scientific judgement and hold every result to four verification gates scaled to its consequence: depositing data, code and agent logs so others can rerun the work; naming an accountable investigator who understands the study design and answers for each claim made; independently verifying findings that could change clinical practice or dietary policy; and disclosing what AI and humans did without penalty. As output grows, papers will likely become concise accounts for human readers resting on machine-auditable records, shifting prestige from the journal name to whether a claim matters when verified.