Have an evidence conversation, then check the claim
Use AI to find and interrogate research. Open the source before a finding becomes a recommendation.
Start with a question you can answer
Specify the population, intervention, comparator and outcome. For example: in postmenopausal women, what human evidence links menopause symptoms with disrupted sleep? Ask for studies and guidelines, with links, limitations and a clear separation between association and causation.
Follow up conversationally: who was excluded, how was sleep measured, how large was the effect, and what remains uncertain? Ask what source material the assistant actually accessed. If it has only an abstract, it should say so; provide a paper you are allowed to share rather than accepting invented full-text details.
A real citation can accompany a wrong claim
The Her Longevity talk uses a deliberately incorrect draft claim about creatine increasing bone density. The cited 2023 trial is real: 237 postmenopausal women received creatine or placebo alongside exercise for two years. It found no added bone mineral density benefit, although some bone geometry measures improved. That does not establish fracture prevention.
This is an authored teaching example, not a captured model failure. Its lesson is to compare the exact sentence with the study’s population, comparison and measured outcome. A plausible citation alone does not validate the sentence.
Make the review visible
Keep a small claim log: proposed sentence, source link, supporting passage or table, limitations and reviewer decision. Mark claims supported, overstated or unverified. Revise the wording before using it in a handout, talk or care discussion.
Research support does not settle an individual treatment decision. The clinician still evaluates the patient, alternatives, feasibility and uncertainty.