Auditing cognitive drift in AI-driven recommendation: a responsible AI methods protocol with a health case demonstration

We propose a protocol to detect and track cognitive drift caused by algorithmic curation. We confirm that the protocol is interpretable, sensitive, reproducible, and portable across domains. It is especially suited to cognitive and neurocognitive research. We build a Cognitive Drift Index (CDI), confirm its three dimensions, and use a small calibration run to set reasonable ranges. We then map CDI to governance action bands. Using health short videos as a case, we estimate path effects with a weighted least squares ANOVA and test robustness. The steps lead to the same pattern, which supports the protocol's design and practical use. We show each component alongside the composite and sort totals into action bands so practitioners and responsible-AI teams can choose proportionate actions. We also consider using the method, as a conceptual tool, with eye-tracking or EEG to enable multimodal validation.