AIasaScienti?cCollaborator

Frombiologytoblackholes,ChatGPTisacceleratingresearchJanuary2026

Introduction:WhyAIforScienceMatters

8.4million

averageweeklymessages

onadvancedtopicsinthehardsciencesandmathematics

Roughly

1.3million

weekly

ChatGPTusers

focusonadvancedmath&sciencetopicsworldwide

Thenumberofmonthly

advancedsciencemessages

grewnearly50%

lastyear

OpenAIisbuildingtoolstohelpresearchersgenerateinsights,acceleratescienti?cdiscoveryandtranslatethoseinsightsintoreal-worldimpact.AcrossChatGPT,researchers,students,STEMfacultyandengineersalreadyuseAItoreadandsynthesizetechnicalliteratures,debugandwritecode,analyzedata,andplanexperiments.Eachweek,ChatGPTseesalmost8.4millionmessagesonadvancedtopicsinthesciencesandmathematics.Thesenowcomefromroughly1.3millionweeklyusersworldwide.

Onlyabout0.1percentoftheglobalpopulationidenti?esasscientists,accordingto

UNESCO

,andyettheyhaveanoutsizedimpact.Scienti?cresearchdrivestheengineofprogresstowardahealthier,moreprosperous,andmoreresilientfuture.Newmedicines,newtechnologies,andnewindustriescomefromnewknowledgeputtopracticaluse.Asmallgroupofearlytwentieth-centuryphysicistslaidthefoundationsofquantummechanicsthroughabstractresearchthat,decadeslater,wouldunderpinmuchofthemoderndigitaleconomy,nowmeasuredinthetensoftrillionsofdollars.Basicresearch–workdonebeforethepayo?isclear–wasthesourceofthatknowledge.In1947,scientistsatBellLabscreatedthe?rstworkingtransistor,

based

oninsightsfromquantumphysics.Thetransistorbecameabuildingblockofcomputers,phones,andtoday’sdigitaltechnology.TheGlobalPositioningSystem(GPS)reliesonEinstein’s

insightsintorelativity

toguideourcarsandkeepatomicclocksaligned.

2

Yetinmanydomains,itisgettinghardertokeepmakingprogress.Economistsandresearchanalystspointtofalling“researchproductivity,”meaningmorepeople,time,andmoneyarerequiredtoproducethesamenumberofinsights.Semiconductorso?erawell-knownexample:sustainingMoore’sLawtodoublethenumberoftransistorsinanintegratedcircuiteverytwoyearshasrequiredadramaticincreaseine?ort,withthenumberofresearchersneededtodayestimatedatmorethan

18times

whatwasneededintheearly1970s.Asknowledgegrowsmorecomplex,eachnewgenerationofresearchersfacesaheavierburdenjusttoreachthefrontier,whichlengthenstheirtrainingtimeandnarrowstheirspecializations.Institutionally,researchhasshiftedtowardlargerteams,withgrowingoverheadforgrantproposals,compliance,reporting,andcoordinationcosts.

Inmedicine,scienti?cadvanceshavesavedcountlesslives.Worldwide

lifeexpectancyrose

fromroughly32yearsin1900toabout73yearsin2023(andtomorethan78yearsintheUnitedStates).Buttheremainingburdenofdiseaseisheavy.ButtheWorldHealthOrganizationreportsthatnoncommunicablediseasessuchasstroke,heartdisease,cancer,anddiabetesstillaccountforabout

74%ofglobaldeaths

.Evenwhenprogressisrapid,turningnewideasintoavailabletreatmentstakestime.Onaverage,ittakes

10-15years

fromtargetdiscoverytoregulatoryapprovalofanewdrugintheUnitedStates,alagimposedonpatientswhoneednewandbettertreatments.

Makingprogressfasterwillsavelivesandimprovethem.AIisalreadyhelpingtoaddressthebottlenecksthatslowsciencedown.Modernresearchisfragmentedacrossdisciplinesandconstrainedbylimitsthatarebothcognitiveandlogistical:readinganddigestingenormousliteraturestodeterminewhatisknown,translatingideasintomathematicsandcode,settingupanalysesandsimulations,checkingcalculations,searchinghugedesignspaces,anddecidingwhichfutureexperimentsarethemostpromising.Usedwell,AIcanserveasahigh-throughputpartnerforthought,computation,andstructuredreasoning,shorteningthecyclefromhypothesistotestandincreasingthecapacityofresearchersworkingaloneandinteams,evenacrossdisciplinarybarriers.

KevinWeil,VPofOpenAIforScience,describestheopportunitythisway:“AIisincreasinglybeingusedasascienti?ccollaborator,andwe’reseeingitsimpactgrowinrealresearchsettings.Moreresearchersareusingadvancedreasoningsystemstomakeprogressonopenproblems,interpretcomplexdata,anditeratefasterinexperimentalwork.Thatusagehasbeengrowingquicklyoverthepastyear,andtheresultsarestartingtoshowupacross?elds.We’restillearly,butthepaceofadoptionandthequalityoftheworksuggestscienceisenteringanewaccelerationphase.”

OpenAIisproudtoworkwithresearchpartnersacrossgovernmentagencies,nationallaboratories,academia,andmedicine,includingtheU.S.DepartmentofEnergy,LawrenceLivermoreNationalLaboratory,theU.S.CentersforDiseaseControlandPrevention,HarvardUniversity,MassachusettsInstituteofTechnology,theUniversityofOxford,TexasA&MUniversity,andBostonChildren’sHospital.

3

Thisreportdetails:

1.HowAItoolsarealreadybeingusedinday-to-dayresearchwork?ows,includingliteraturesynthesis,codegenerationanddebugging,dataanalysis,simulationsupport,andexperimentplanning

2.WhatearlyresultssuggestaboutAI’spotentialtosupportnewbreakthroughs

3.HowindividualscientistsacrossmultipledisciplineshaveusedChatGPTtomakeprogressintheir?eld

4.PolicysuggestionstosupportcontinuedAIprogressinscienceandmath

4

Thescaleofscienti?candmathematicalworkonChatGPT

AcrossChatGPT,asmallbutconsequentialcohortofresearchersusesOpenAI’sAImodelsforsophisticatedtasksrangingfromtechnicalderivations,advancedmathematics,engineeringsimulationandmodeling,andotheradvancedproblem-solving.ThisincludesscientistsandmathematiciansspanningPhDcandidatesandpost-docstoworkingresearchersandSTEMfaculty.

BasedonaninternalanalysisofafullrandomsampleofanonymizedChatGPTconversationsfromJanuarythroughDecember2025,averageweeklymessagecountsonadvancedscienceandmathtopicsgrewabout47%,from5.7millionmessagestonearly8.4millionmessagesoverthecourseof2025.AsofJanuary2026,therearenearly1.3millionweeklyusersdiscussingadvancedtopicsinscienceandmath.

Together,thesesignalsshowhowChatGPTisacceleratingadvancedresearch:withtensofmillionsofadvancedhard-scienceandmathpromptseachmonth,generatedbyalargeandgrowingcohortusingthesystemforseriousscienti?candengineeringworktobene?tsocietyandsupporteconomicgrowth.

Here’showusagebreaksdownacrossdisciplinesamongourcohortofresearch-focusedusers:

5

WhatscientistsandmathematiciansactuallydowithChatGPT

Scientists,mathematicians,andengineersuseChatGPTasahighlyavailabletechnicalcollaborator:atoolwithwhichtheycaniterateoncalculations,translateideasintocode,interrogateassumptions,andcompresscomplexmaterialsintoworkablementalmodels.InOpenAI’sanalysis,“advanced”hardsciencepromptsarede?nedasthosebeingorientedtowardresearch,andrequiringgraduate-levelorresearch-levelexpertisetoanswercompetently.Withinthatcohort,behaviordi?ersfromtypicalusersinwaysthatmapdirectlyontomodernresearchwork?ows.

Researchtasksclusterindomainssuchascoding(drafting,reworking,anddebuggingcode),dataanalysis(cleaningandmergingdatasets,runningstatistics,interpretingresults),mathematicalreasoning(derivations,proofstrategies,algebraicchecking,longcalculations,translatingbetweenformalisms),andliteraturereviewandsynthesis(?ndingreferences,understandingrecentwork).

ContrastedwithtypicalusersofChatGPT,advancedscienceandmathusers:

●Sendroughly3.5×moremessagesthanthebaseline

●Sendcoding-relatedmessagesnearly12×moreoften

●Average9informational-overviewpromptsperweekvs.1.5prompts

6

Categorizedbymostcommontasks,hereishowthiscohorttendstouseChatGPT,:

7

AIatthefrontierofmathandscience

FrontierAIcapabilitiesinmathematics

Recentprogress

Overthelasttwoyears,largelanguagemodelshaveprogressedfromearly,unevenperformanceonbasicarithmetictohandlingmulti-stepmathematicalreasoningthatcanbeusefulinrealmathematicalwork.Muchofthatimprovementcamefrommethodsthatencouragestep-by-stepreasoning,andfromtighterintegrationwithtoolslikecalculatorsandcodeexecutionforexactcomputation.Asmodelsimproved,benchmarkingalsoshiftedtowardhardertestsdesignedtomeasuredeeperreasoningwhilereducing“patternmatching”wins.

In2025andearly2026,thegreatestimpacthascomefromtest-timecomputescaling,or“slowthinking.”Insteadofcommittingquicklytoonepath,amodelwillspendmorecomputationexploringalternativesandself-checking.Atthesametime,approachestotrainingthatrewardveri?ableoutcomes,suchasproducingacorrect?nalanswerorexecutablecode,havepushedmathandcodingtobecomemorereliable,andcorrectoftenenoughtobeusefulwithhumanguidance.

OnesignofthisshiftcamewithInternationalMathematicalOlympiadcoverageinJuly2025,whenanOpenAImodelachieved

gold-levelperformance

onthe2025problemsetalongsideDeepMind.

Currentcapabilities

GPT-5.2’ssteadyadvanceinmathematicalcapabilitiesstemsfromstrongerlong-horizonreasoning,moresystematicveri?cationhabits,andbetteruseofcheckabletools.AIME,theAmericanInvitationalMathematicsExamination,isdesignedtotestmulti-stepproblemsolving:GPT-5.2ThinkingachievedaperfectscoreonAIME2025withoutexternaltools.TheGPT-5.2series(e.g.ThinkingandPro)hasprogressedpastcompetition-levelperformancetowardmathematicaldiscovery,includingworkonestablishedopenproblems.

Onresearch-stylebenchmarking,the“Google-proof”FrontierMathproblemsethasbeenconstructedtobeaccessibleonlytotrueexperts;i.e.evenasmartPhDstudentinmathcannotsolvetheminafewhoursofwork.Onthatbenchmark,GPT-5.2Thinkinghassolved

40.3%ofproblems

inTiers1–3.

8

Performancestilldropsonthehardesttier,whereGPT-5.2Prohasscored

31%onFrontierMathTier4

onasetofproblemsthatcanbedescribedas“miniresearchprojects.”

AsecondmajorcapabilityleaphasoccurredasGPT-5.2isincreasinglypairedwithformalveri?cationwork?ows.Inoneprominentintegration,GPT-5.2generatesnaturallanguageproofsandusesAristotle,athird-partyLLM,toformalizethoseproofsinLean,whichisaproofassistantwhereproofsarewritteninaformacomputerchecksstepbystep,withthesystemdetectingandcorrectinggapsduringformalization.

Theseintegrationsmatterbecauseonelongstandingfailuremodeforlanguagemodelsinmathematicshappenswhenasolution“l(fā)ooksright”,butisn’t:i.e.argumentsthatappearplausible,butcontainsubtlegaps.Lean-checkedproofssubstantiallyraisethestandardofcon?denceinaproofbyforcingexplicit,mechanicallycheckedstepsunderastatedformalization.

Erd?sproblems,AIsolutions

PaulErd?s(1913–1996)wasaglobe-trottingHungarianmathematicianwholivedoutofasuitcase,movingfromcampustocampusashesoughtoutfellowmathresearchers.The“Erd?sproblems”arethevastsetofquestionsandconjecturesheposed,rangingfromdeceptivelysimplepuzzlestoproblemsthatstillresistthebesttechniqueswehave.Theproblemshaveactedliketrailmarkersformodernmathematics:clarifyingwhatwedon’tyetunderstand,seedingwholeresearchprograms,anddrawinggenerationsofmathematicianstowardtheedgesoftheknown.

Inearly2026,GPT-5.2hascontributedtosolutionstoseveralopenErd?sproblemswiththehelpoftoolslikeAristotleandLean,andwiththesolutionsvalidatedbyTerenceTao.Problems

#281

,

#728

,

#729

arenowlistedasproved,and

#397

asdisproved.WhilemathematicianscautionthatErd?sproblemsvaryenormouslyindi?culty,thesesolutionspointtotheincreasingcapabilityofOpenAImodelstodorealmathematicalworkandmakenovelcontributionswithminimalguidance.

Near-termpotential

Somesigni?cantmathematicaldiscoveriescantaketheformofstitchingknownmethodstogetherto?ndthecorrectargument.GPT5.2candothisnowinmanycases.Otherdiscoveriesinvolveinventingentirelynewkindsofmath,asNewtoninventedcalculustounderstanddynamics(theforces,massandenergythatexplainchangesinmotion).ThatisbeyondcurrentAImodels.Butathirdtypeofsigni?cantdiscoveryinvolvesestablishingconnectionsbetweentwo?elds,andbringingtheknownmachinery,resultsandtoolsofone?eldtotheother(e.g.algebraicgeometry,whicharosefromabstractalgebraandclassicalgeometry).ModestexamplesofthishavealreadyoccurredwithAI,andwebelievethesigni?canceofthoseconnections,andthesub?eldscreatedfromthem,willincreaseinthenearfuture.

Atthesametime,muchofAI’snear-termvaluewillbeintransformingwork?ows.GPT-5.2canproposeanditerateonsolutionpaths,whileexternaltoolsenforcecorrectnessthroughexactcomputationorformalchecking.Thisalignswithbroadertrendstowardhybridapproachesandauto-formalization,

9

whereinformalmathistranslatedintoformallanguageslikeLeansothatcorrectnesscanbeveri?edmechanically.GPT-5.2isalreadyusefulforliteraturereview(surveyingwhatisknown)tolocatetheedgeofknowledgeandsurfaceunexpectedorobscurereferences.Itisusefulincomingupwithproofs,critiquingthem,simplifyingthemandsuggestingproofstrategies.

Ifthistrajectorycontinues,GPT-5.2’snear-termimpactislikelytoshowupasabroadproductivityupgradeformathematicalresearchers,aswellasscienceandengineeringteams(sincemathisacorecomponentofmuchscienti?candengineeringwork):thiswillmanifestasfastertranslationfromamessyproblemdescriptiontoacleanmathematicalstatement,fewerdroppedconstraintsinmulti-stepderivations,morereliabledebuggingofcalculationsandproofs,andagrowingshareofresultsthatcanbebackedbyformalveri?cation.

10

FrontierAIcapabilitiesinscience

Acrossdisciplinessuchasphysics,chemistry,andbiology,ChatGPT-classLLMsincreasinglysupporttechnicalreasoningandtool-mediatedresearchwork?ows,aswellasscienti?cwriting.BenchmarkslikeGPQA,agraduate-levelsetof“Google-proof”questionsauthoredbydomainexperts,initiallyshowedasigni?canthumanadvantage,withexpertsreaching65%accuracywhiletheGPT-4baselinereached39%.OpenAInowreports

GPQADiamondaccuracy

of93.2%forGPT-5.2Proand92.4%forGPT-5.2Thinking(withnotoolsenabledandatmaximumreasoninge?ort),suggestingahigherbaselineforgraduate-levelscienti?cquestionansweringacrossmanyscienti?cdisciplines.Inparallel,automatingtheunspokendrudgeryofresearch–referencehunting,bibliographyassembly,androutineadministrativereporting–freesscientists’scarceattentionforhigher-valuework.Pairedwithinformationretrievalandexecutabletoolsthatenablecheckablecalculationsandstepwisevalidation,thesemodelsarebecomingreliablework?oworchestratorsforscienti?cplanning,analysis,anddocumentation,teeingupaccelerationacrossmany?elds.

Physics

Lastmonth,OpenAIannouncedamemorandumofunderstandingwiththeUSDepartmentofEnergytosupportcollaborationonAIandadvancedcomputinginordertoadvanceDOEinitiativesincludingtheGenesisMission,withapplicationsinenergysuchasfusionresearch.

Inphysics,LLMsarebeingusedacrossmajorfacilities,includingmany

USnationallabs

,asaunifyinglayerovercomplexoperationsstacksandinternalknowledgebases,acceleratinganalysisanddecision-makingunderstrictconstraintsalongsideexistingmachine-learningtoolsforsimulation,real-timedatareduction,andexperimentalcontrol.LLMscandigestshiftlogsandalerts,answerquestionsfrominternaldocumentation,andhelprouteworktotherightanalysis,simulation,orcontroltool,allunderstrictsafety,timing,andresourceconstraints.Thisaugmentsalongertrackrecordofspecializedmachinelearninginphysics:neural“surrogate”modelsthatapproximateequation-governedsimulationswhenfullcomputationistooslow,real-time?lteringandreconstructioninparticledetectorsthatseetensofmillionsofcollisionspersecond,andmachine-learningsupportedcontrollersthatcoordinatethemanyelectromagnetsintokamakfusionexperimentswhilestayingwithinhardwareandsafetylimits.

Near-termgainsinphysicsarelikelytoconcentrateinhigh-throughput,decision-heavysettingswhereexpertattentionandturnaroundtimearethebottlenecks.AIassistantsthatcanreferencealab’sinternaldocumentationandrunautomatedcheckscanturnliveexperimentalerts,notes,andlog?lesintoprioritizednext-stepresearchplansandrepeatableanalyticoutputs,suchasnotebooks,scripts,andreports.

11

Intheoreticalphysics,LLMswillcontinuetodelivervalueasthoughtpartnerscompressingresearchers’cognitiveoverheadwhileexpandingthespaceofexploration.Whenacomplexcalculationhitsaroadblock,LLMscansurfacewaystoreframetheproblem,suggestintermediatesteps,andprovidequickconsistencychecksthatreducetimethatphysicistsspendbeing“stuck.”Sometimesthisproducesinsightslikeamissingcondition,acleanerformulation,orausefulrelationshipbetweenexpressionsthatcanbecomeameaningfulingredientofapaper.Thelargermultipliermaycomefromsynthesizingresearchatscale,wheremodelsscantheliteratureacrosspapersandsub?eldstosurfacenewconnectionsautomatically.

Chemistry

ChatGPT’sapplicationsinchemistryhavemovedpastone-shotquestionansweringtowardmulti-stepwork?owsthattranslatebetweennaturallanguageandchemicalrepresentations,andrelyonexternaltoolsforveri?cationandretrieval.

ChemBench

,publishedinNatureChemistryin2025,curatedmorethan2,700expert-writtenquestionsandfoundthatleadingmodelsoutperformedhumanchemistsonaverage,whilestillstrugglingonsomebasictasksandproducingovercon?denterrors.

LeadingAIsystemsinchemistryincreasinglyuseahybridwork?ow:ageneral-purposeLLMhelpstoplanmulti-stepworkandcoordinatetools,whilespecializedmodelsthatunderstandmolecularstructurehandlepredictionandsimulation.Akeyexampleisstate-of-the-artgraphneuralnetworks(GNNs):thesemodelstreatamoleculelikeanetwork,withatomsasnodesandbondsasconnections,sothesystemcanlearnhowlocalchangesa?ectthewholestructure.NewerGNNsaredesignedsotheirpredictionsremainconsistentwhenamoleculeisrotatedorshiftedin3Dspace,whichmakesthemwellsuitedforlearningtheenergyrulesneededtorunfast,accuratemolecularsimulations.Asmoleculargraphmodelsscalewithmoredataandpretraining,resultscontinuetoimprove,buttougherbenchmarksforchemicalreasoning,includingorganicreactionmechanismtaskssuchas

oMeBench

,underscoretheneedforhumanoversight.

12

Biology

ChatGPT’sapplicationsinbiologyincreasinglyextendintomulti-stepwork?owsthatcombinenatural-languagequestionswithstructuredscienti?csourcessuchasgenomicsdatabases,proteinrepositories,andthebiomedicalliterature,oftenwithcodeandretrievaltoolsusedfortraceabilityandveri?cation.

GeneTuring

,a2025genomicsbenchmarkinBrie?ngsinBioinformatics,curated1,600questionsacross16tasktypesandmanuallyevaluated48,000answersfrommultiplemodelcon?gurations.Thestrongestresultscamefromatool-augmentedsetupthatpairedageneral-purposemodelwithdirectaccesstoNationalCenterforBiotechnologyInformation(NCBI)APIs,reinforcingthatreliabilityimproveswhenlanguagemodelsareconnectedtoauthoritativereferencedataandcanshowtheirwork.

Aswithchemistry,state-of-the-artAI-enabledresearchinbiologyreliesonhybridstacks:general-purposelanguagemodelshelpplanandcoordinateanalysis,whilespecializedfoundationmodelstrainedonbiologicalsequencesandstructurespowerpredictionanddesign.Inproteinscience,

AlphaFold3

representsasteptowarduni?edbiomolecularmodelingbypredictingthejoint3Dstructureofcomplexesthatcanincludeproteins,DNAandRNA,andsmallmoleculeswithinadi?usion-basedarchitecture.

13

14

Usecasepro?les

ErnestRyu-Mathematician

ErnestRyupickedupChatGPToutofcuriosityin2023,andsawitadvanceuntilitcouldgenerateapublishableresultlastyear.Ryu’sacademicworkhasfocusedonoptimization:themathbehinde?cient,reliablealgorithmsthatsupportmoderneconomies,fromplanninglogisticstokeepingaircraftwingsstable.

Whenlargelanguagemodelsweresurginginpopularityin2023,Ryubeganhis?rstexperiments:couldamodeltranslatereal-world“wordproblems”intopreciseoptimizationmodels,includingallthehiddenconstraints,andthenhandthemtoasolver?Schedulingabaseballseason,forexample,requireshardconstraints(e.g.noteamplaystwogamesatthesametime)andsofterones(e.g.travelrestdaysthatcanbeviolatedifnecessary).Thatearlymodelstruggledwiththecarefulconstrainthandlingthisworkdemands,sometimesomittingconstraintsandfailingonlarger,realisticschedules.

Thein?ectionpointcamelastyear,afterthearrivalofreasoningmodelsandOpenAI’swinninggoldattheInternationalMathematicalOlympiad.ThesameclassofschedulingproblemsRyuhadtestedbeforewerereliablysolved.ThatsuccessledRyutoapplyLLMstoeverydaymathematicalwork:whilewritinglectures,RyubeganaskingChatGPTforproofsofresultsheknewweretruebutdidn’thavetopofmind.

Finally,hetrieditonresearch.RyuchoseaproblemrelatedtoNesterovacceleration,awell-knowntechniqueforspeedingupoptimization,andpickedaversionoftheproblemthatwasopenlongenoughthatothershadattemptedit,yetsimpleenoughthatashortproofmightexist.Forthreeconsecutiveevenings,afterhissonwenttobed,heworkedfrom8pmtomidnight,andbythethirdnighthehadguidedAItothepointwhereitcrackedtheproblem.

Theircollaborationlookedlikerealresearch.Themodelproducedaninitialproofwithacalculationmistake,soRyubegantoiterate:hecorrectedtheerror,preservedthecorrectintermediatestepsinagrowingprompt,abandoneddeadends,andpushedthemodelintootherapproaches.Ryudescribesthisasmazerunning,whereyouturndowncorridorsandopendoors,sometimesonlyto?ndthemempty,whilekeepingamentalmapofwhatfailsandwhatseemspromising.ChatGPThelpedRyuacceleratehowfastheranthemazeby3xto10x.

15

Onthethirdnight,themodelmadeasmallbutmeaningfulleapthat“l(fā)ookeddi?erent”enoughtounlocktheproof.Ryusaidhe“morethantriple-checked”theargument,thenhadastudentverifyit,beforeheshareditpubliclytoanoptimizationcommunitythatreactedwithsurpriseandexcitement.Fromthere,thecontinuous-timeresultwastranslatedintothediscrete-timealgorithmstatementwithasingleprompt,leavingashort,onenovelcorethatmetthestandardforapublishableadvance.

Sincethen,RyuhasjoinedOpenAI’ssyntheticdatateam,wherehiscorefocusisimprovingthemodel’smathematicalcapability.

AlexLupsasca-Physicist

AlexLupsascacametoAIthewaymanyphysicistscometoboldclaims:withpoliteskepticismandabunchoftests.Inearly2025,hetriedChatGPT,andhefounditusefulfortheroutineadministrativetasksthatregularlypopupinacademia.Buthedidnotseeitasatoolforthehardpartofthejob:turningthelawsofphysicsintoconcreteandveri?ablepredictions.

Acommonrealityofacademicpublishing,Lupsascasays,isthataresearchprojectmayoftenresultinadraftequationsandroughconnectivetissue,andaprofessor’sworkbecomesthatofacopyeditor,especiallywhencollaboratorsarewritingoutsidetheir?rstlanguage.ChatGPTcantakearough