Study highlights how AI models take potentially dangerous 'shortcuts' in solving complex recognition tasks