How an experiment involving a made-up skin condition exposes the risks of increasingly popular AI medical advice