Ask the evidence
An AI model answers from PubMed studies, clinical trials and 48 official guidelines, grades the evidence and links every claim to its source.
How it works
Your question becomes a search of PubMed, the US National Library of Medicine's index of medical research, Europe PMC's open-access full texts and a library of official guidelines.
Meta-analyses and randomized trials in people come first, then other studies. Trials still running are found on ClinicalTrials.gov.
The evidence is graded with GRADE, the method Cochrane and the World Health Organization use, before a word of the answer is written.
An AI model writes the answer using only those studies. Each claim has a numbered link to its source.
How we rate evidence
Grades run from A to D, each a GRADE certainty, the scale used in Cochrane reviews and clinical guidelines. U means no study answers the question.
- AStrongHigh certainty
- Good randomized trials in people that agree, with no serious weaknesses. New studies are unlikely to change the answer.
- BModerateModerate certainty
- Good evidence in people with one real weakness. New studies could change the answer.
- CLimitedLow certainty
- Studies in people with real weaknesses: small, short or conflicting trials, or observational studies only.
- DPreliminaryVery low certainty
- Very uncertain: animal or cell studies, or studies in people with several serious weaknesses.
- UInsufficientNo evidence
- The studies found don't answer the question.
How a grade is worked out
Every answer's grade is reached the same way, and the answer shows each step under "How we graded this". An AI model reads the studies and judges each check; the steps are then counted the same way every time, so the same judgements always give the same grade.
1 It starts by design
- Randomized trials in peopleHigh
- Long-term studies of what a test predictsHigh
- Observational studies in peopleLow
- Animal or cell studiesVery low
2 It drops a step or two for each problem
Problems that overlap are counted once, so trials in people drop two steps at most unless one problem is severe.
- Study flawsRisk of bias
- Trials that weren't blinded, lost many people or were run by the maker.
- DisagreementInconsistency
- Studies of the same thing found different results.
- Fit to the questionIndirectness
- Different people, doses or outcomes from the ones asked about.
- PrecisionImprecision
- Too few people or events to be sure of the size.
- Missing studiesPublication bias
- Signs that studies with no effect went unpublished.
3 Observational evidence can rise
- Large effect
- Risk roughly halved or doubled, again and again.
- Dose-response
- More of it, more effect.
- Bias would hide the effect
- Any likely bias would make the effect look smaller than it is.
Limits
- Answers are written from study abstracts and, where a paper is free to read, its full text. They can contain mistakes, so check the studies they cite.
- Answers describe what studies found in groups of people. Talk with your clinician before starting, stopping or changing a treatment.
- Longevity University is part of Platform Health, which sells some of the things people ask about. Answers never name products or stores, ours included.
- Questions go to PubMed, Europe PMC, ClinicalTrials.gov and Anthropic, which writes the answers. We keep questions and answers, without anything that identifies you, so repeated questions are answered faster. Don't include your name or other personal details.