ChatGPT:fiveprioritiesforresearchEvaA.M.vanDis,JohanBollen,RobertvanRooij,WillemZuidema&ClaudiL.BocktingConversationalAIisagame-changerforscience.Here’showtorespond.AchatbotcalledChatGPTcanhelptowritetextforessays,scientificabstractsandmore.SinceachatbotcalledChatGPTwasreleasedlatelastyear,ithasbecomeapparentthatthistypeofartificialintelligence(AI)technologywillhavehugeimplicationsonthewayinwhichresearcherswork.ChatGPTisalargelanguagemodel(LLM),amachine-learningsystemthatautonomouslylearnsfromdataandcanproducesophisti-catedandseeminglyintelligentwritingaftertrainingonamassivedatasetoftext.ItisthelatestinaseriesofsuchmodelsreleasedbyOpenAI,anAIcompanyinSanFrancisco,California,andbyotherfirms.ChatGPThasperspectives.However,itcouldalsodegradethequalityandtransparencyofresearchandfundamentallyalterourautonomyashumanresearchers.ChatGPTandotherLLMsproducetextthatisconvincing,butoftenwrong,sotheirusecandistortscientificfactsandspreadmisinformation.Wethinkthattheuseofthistechnologyisinevitable,therefore,banningitwillnotwork.Itisimperativethattheresearchcommunityengageinadebateabouttheimplicationsofthispotentiallydisruptivetechnology.Here,weout-linefivekeyissuesandsuggestwheretostart.HoldontohumanverificationLLMshavebeenindevelopmentforyears,butcontinuousincreasesinthequalityandsizeofdatasets,andsophisticatedmethodstocalibratethesemodelswithhumanfeed-back,havesuddenlymadethemmuchmorepowerfulthanbefore.LLMswillleadtoanewgenerationofsearchengines1thatareabletoproducedetailedandinformativeanswerstocomplexuserquestions.ButusingconversationalAIforspecializedresearchislikelytointroduceinaccuracies,biasandplagiarism.WepresentedChatGPTwithaseriesofquestionsandassignmentsthatrequiredanin-depthunderstandingoftheliteratureandfoundthatitoftengeneratedfalseandmisleadingtext.Forexample,whenweasked‘howmanypatientswithdepressionexperiencerelapseaftertreatment?’,itgen-eratedanoverlygeneraltextarguingthattreatmenteffectsaretypicallylong-lasting.However,numeroushigh-qualitystudiesshowtha...
