AI and Coder Employment: Compiling the Evidence

AI and Coder Employment: Compiling the Evidence LelandD.Crane* PaulE.Soto* March20,2026 Abstract WeevaluatewhetherLLMshavehadanydiscernibleimpactontheaggregatelabor marketsofar. Wefocusonoccupationsthatarecomputerprogramming-intensive,motivated by data showing that coding is one of the most LLM-exposed tasks. Linking O*NET to CPS we find that aggregate employment of coders has decelerated sharply since the introduction of ChatGPT. Using a novel control variable for industry-level shocks we show that the deceleration is not attributable to the exposure of coders to slowing industries, suggesting instead that coders experienced an occupation-specific shockaroundtheintroductionofChatGPT.Coderemploymenthascontinuedtogrow inrecentyears,thoughmuchmoreslowlythanitdidpre-2022.Wevalidatetheindustrylevelcontrolvariablebyexamininghistoricalexamplesofoccupationsthatexperienced either occupation-specific or industry-level shocks. We also provide statistics on the agreementratesbetweendifferentmeasuresofAIexposure. *BoardofGovernorsoftheFederalReserveSystem WethankGabriellaGalassi,DaveByrne,conferenceparticipantsattheSanFranciscoFedandtheBankof Canada, andbrownbagparticipantsattheFederalReserveBoardandSogangUniversityforinsightfulcomments.WethankNicoleHoffmannforexcellentresearchassistance. OpinionsexpressedhereinarethoseoftheauthorsaloneanddonotnecessarilyreflecttheviewsoftheFederal ReserveSystemortheBoardofGovernors.

1 Introduction RecentprogressinthefieldofAIhasfueledspeculationthatlabormarketdisruptionsmay beimminentoralreadyunderway. Bloombergassertsthat“TheAIHiringPauseIsOfficially Here”,andprominenttechnologycommentatorsspeculatethatfundamentaltransformation maybe1-2yearsaway.1 Suchascenariowouldbeaseriouschallengeforpolicymakers;in thispaperwebringtheavailabledatatobearonthequestion. Recentlyreleasedreal-worldAIusagedatashowsthatcomputerprogrammingisoneof the predominant uses of generative AI. The Anthropic Economic Index (AEI, Handa et al. 2025)publishesanonymizeddataonhowpeopleuseAnthropic’sflagshipchatbot,Claude,a majorcompetitortoOpenAI’sChatGPT.Thisisconsistentwithanecdotalevidencethatcoding assistants are one of the more successful applications of the tools. Motivated by this— andotherevidencedevelopedbelow—wefocusoncoding-intensiveoccupations. Ourview isthatifgenerativeAIistosubstantiallyaffectthejobmarket,theeffectsshouldbeapparent herefirst. WeuseOccupationalInformationNetwork(O*NET)datatoidentifyoccupationswhere computer programming is an important skill. We track monthly employment of “coders” in the CPS. Our analysis is essentially an event study; we explore whether employment in coding-intensive jobs had a markedly different trend after the introduction of ChatGPT in November 2022.2 Simple regressions show that the growth of coder employment did indeed slow down post-ChatGPT. This is more clear when we control for industry-level shocks (discussed below.) Controlling for factors that affect industry employment but not its composition, we find robust evidence that annual coder employment growth is about 3 percent lower now than it was pre-ChatGPT. The interpretation of this estimate is subject to several caveats: among other issues, we are not controlling for AI’s effects on aggregate 1See“BehindtheCurtain:Awhite-collarbloodbath”,“SituationalAwareness”andAI2027forexamples. 2Of course, the introduction of LLMs for coding is not a single event. GPT 3.5 and Github Copilot were availablebeforeNovember2022,andsubsequentmodelreleaseshaveimprovedcodingproficiencysincethen. But November 2022 is the most natural breakpoint, and likely the time when most managers and business leaderswouldhavebecomeawareofthepossibleproductivityimplications. 1

labordemandorprices,occupationaltaskmixmaybechanginginawaythatoverstatesthe effects on coders as a skill group, and there may be ample room for coders to match well in other occupations. Nonetheless, the results suggest that AI is having a measurable and potentiallyconsequentialimpactonemploymentforsomegroups. Animportantstepwetakeistodevelopcounterfactualemploymentseriesforcoders— thatis,coderemploymentabsentanoccupational-specificshockcausedbyLLMs—basedon industryexposure. Thecounterfactualisderivedfromawithin-industry/between-industry decomposition of occupational employment growth: We explore which employment fluctuationsareattributabletoindustry-levelfactorsandwhichareoccupation-specificshocks. The counterfactual variable is the sum of industry employment growth rates weighted by the distribution of coders across industries. This is the between-industry component of occupationgrowth;ittellsuswhatcoderemploymentwouldhavedoneifitstayedaconstant fractionofemploymentwithineachindustryasthoseindustriesgrewandshrank. Theintuitionandkeyidentifyingassumptionisthatindustry-levelshocks(likedemandfortheoutputproductandchangesinindustryTFP)shouldscaleindustryemploymentupanddown homothetically—not affecting composition—while occupation-specific shocks will change the composition of employment within industries. Comparing the counterfactual to actual coder employment within the information sector shows that some of the slowing in coder employment is attributable to industry-level dynamics rather than an occupation-specific shock,thoughasubstantialfractionisstillattributedtooccupation-specificfactors.3 Aside from the main results on coder employment growth and the properties of the counterfactual, we uncover several other interesting facts. For example, while coders are widely dispersed across industries, about 40 percent of coders work in computer systems designandrelatedservices(NAICS5415). Thisindustrycoversmanysoftware/ITcontractor activities. It is remarkable that the modal coder is not working at a Silicon Valley tech firm,noratastartup,norasanin-housedeveloperinotherindustries,butisdoingcontract 3NotethatthecounterfactualhastheformofaBartikinstrument, butwearenotusingittoobtaincausal variationquantities;hereitplaystheroleofacontrolvariable. 2

software development. Note that this industry is not in the information sector (NAICS 51) oftenusedtoproxythetechindustry. Inaddition,weexplicitlycomparethegenerativeAIexposuremeasuresfromEloundou et al. (2024) and Handa et al. (2025). The large and growing literature studying generative AIlargelyreliesonthesemeasuresbutlittlehasbeendonetoexplicitlycomparethem. We find notable disagreement in the measures, though both agree that coders are among the mosthighlyexposedoccupations. Before turning to the data and results it is important to clarify the scope of the paper. An important question is what we should expect LLMs to do to coder employment. LLMpowered coding assistants appear to be complements in production to coders: The coding assistantincreasesthemarginalproductofacoderbylettingthemcompletetasksfaster. If thedemandforcodingservicesisinelasticthiscouldleadtoafallinemployment,asfewer coders are required to satiate demand. On the other hand, elastic demand for coding services could result in increased coder employment as the more efficient coders are able to serve a much larger market for low-cost coding. This effect seems more likely in the long run than in the short run, as businesses have time to adapt and develop new products. Ex ante it is not clear which world we live in, though there is some work on the topic (Humlum, 2021; Acemoglu and Loebbing, 2026). In the longer run other important dimensions are the introduction of new work and the reorganization of occupational tasks (Acemoglu andRestrepo,2019). LLMsmayleadtotheintroductionofnewcoding-intensiveproducts, and individuals from other occupations may begin to do LLM-assisted coding tasks. This meansthatboththeshort-runandlong-runeffectsofAIoncoderemploymentareempirical questions. Relatedly, our focus is on generative AI as an occupation-specific shock. Our emphasis onexposedoccupationsmoreorlessforcesustoignoreothermargins,suchastheautomation of a wide range of a firm’s business processes by LLM-using coders. Our view is that initial effects of AI are most likely to show up as occupation-specific shocks, since larger- 3

scale automation and new businesses take time to develop. While these margins are likely todominateinthelongruntheymightnotbeinformativethisearlyinthediffusionprocess. We also cannot address general equilibrium effects, where automation raises productivity and labor demand for all workers. These are important and difficult questions that are beyondthescopeofthispaper. 2 Literature Whilefutureeffectsonproductivityareuncertain,generativeAIisshowingsignsofbeinga generalpurposetechnology(GPT)andaninventionofthemethodofinvention(IMI),which could have longer lasting impacts on productivity growth (Baily et al., 2025). AI adoption rates in the workplace have also been rising steadily (Bonney et al. 2024, Bick et al. 2024, Craneetal.2025),andstudieshavefoundthatmanyjobsareexposedtogenerativeAI(see Eloundouetal.2024andFeltenetal.2023). Thesefindingshaveraisedquestionsaboutthe macroeconomic impact of generative AI, see, e.g., Acemoglu (2025) and Korinek and Suh (2024). Focusing on the labor market, Humlum and Vestergaard (2025) find little imprint of generativeAIuseonworkerwagesandemployment, whileBrynjolfssonetal.(2025)finds a decline in the employment of young workers relative to older workers within the occupationsmostexposedtoAI.Ourworkislargelycomplementaryinthatweseektoidentify changes in total occupation employment for a large, highly-exposed occupation, while Brynjolfsson et al. (2025) identify relative changes in the age composition of employment for a broader group of exposed occupations. Brynjolfsson et al. (2025) are able include firmlevel controls from their proprietary data which is not possible in the public CPS we use, thoughwedevelopusefulindustry-levelcontrols. Ourviewisthatthefirst-orderquestion is whether labor demand (in a given occupation grouping) is increasing or decreasing as a result of AI. A larger literature is addressing this and related topics, including Lichtinger and Hosseini Maasoum (2025), Eckhardt and Goldschlag (2025), Gimbel et al. (2025), Is- 4

cenko and Millet (2026), Brynjolfsson et al. (2026), Dominski and Lee (2025), Atkinson and Yamco(2026),AhnandCarollo(2026),andMassenkoffandMcCrory(2026) The literature has largely focused on indirect generative AI exposure measures developed by researchers: Eloundou et al. (2024) and Felten et al. (2023) calculate exposure by asking humans (or LLMs) to judge AI exposure of individual tasks. Soto (2025) estimates firm-levelAIusagethroughconferencecalltranscripts. Thesemethodologiesareusefuland especially valuable when tools are new, but we now have some real-world data on usage from the Anthropic Economic Index (Handa et al. 2025). These data are our starting point. We go further than much of the literature in comparing the Handa et al. (2025) exposure metric to Eloundou et al. (2024) and showing that while there is substantial disagreement, bothagreethatcoderoccupationsarehighlyexposed.4 Similar to Chandar (2025) our methodology provides timely readings of employment trends related to AI based on public data. Another contribution of our paper is to develop a control for industry-level shocks which is especially useful when firm-level data are not available. The industry-level control lets us cleanly separate occupation-specific shocks(whichisourfocus)frompossiblycorrelatedcyclicalorsecularindustryfactors. The control is based on simple within-between decompositions like those used by Davis and Haltiwanger(1992), JaimovichandSiu(2020)andothers, thoughourapplicationis—asfar asweknow—original. Aseparateliterature(e.g.,Eisfeldtetal.(2023))examinesgenerativeAIthroughthelens of financial markets. They find that markets expect firms with more-exposed workforces to deliver more value in the future. Similarly, Wiles and Horton (2025) find that while AI technologycanreduceprivatesearchcostsforfirmhiring,thereislittleevidencetosuggest thatthisleadstoimprovedlabormarketefficiency. Tomlinson et al. (2025) also analyze real-world LLM use, in their case the Bing Copilot tool. We suspect that the usage patterns they document may be dominated by the fact that 4SeeCottieretal.(2023)forananalysisofthedifferencesbetweenvariousexposuremeasures,priortothe releaseoftheAEIdata,andGimbeletal.(2025)foracomparisonmoresimilartoours. 5

the tool is integrated into a search engine, so many queries are of the type “looking for information.” Nonetheless,ourmethodscouldbeusedwiththeirdataaswell. 3 Motivation and Coders as an Occupational Group In this section we first provide evidence that coders are likely the most generative AIexposedoccupationalgroupandthenconstructadefinitionofcoders. 3.1 AnthropicEconomicIndex Anthropic is one of the largest firms training LLMs and its Claude models are competitive with models offered by OpenAI and Google. Beginning in February 2025 Anthropic began releasing the “Anthropic Economic Index” which provides data on the composition of queries to their Claude models.5 The data are organized by O*NET tasks; an instance of Claude is shown a user’s query (in a privacy-preserving fashion) and asked which O*NET task it matches best. The resulting dataset shows which tasks are most commonly representedininteractionswithClaude.6 Figure1shows,bybroadoccupationalgroup,theshareofClaudequeriesrelatedtothat occupation’stasks(red)andtheshareofthatoccupationintheworkforce(blue). Strikingly, computer and mathematical occupations account for more that 1/3 of Claude queries, despite comprising only 3.4% of the workforce. Handa et al. (2025) show that these queries are essentially all computer programming-related. This simple fact motivates our focus on codingfortherestofthepaper: Codersareclearlyaveryhighlyexposedgroup. Thisisalso consistent with the survey data from Bonney et al. (2024) and Bick et al. (2024), which find thehighestuseratesamongcomputerandmathematicaloccupationsandintheinformation andprofessionalandbusinessservicessectors. 5The analysis in this section is based on the first report from February, 2025. For the complete set of AEI releases,pleaseseehttps://huggingface.co/datasets/Anthropic/EconomicIndex 6ThedataexcludesAPIusers,whichmightoverrepresentcoders,aswellasbusinesslicenses: “Becausewe focusonstudyingpatternsinindividualusage,theresultssharedinthispaperexcludesactivityfrombusinesscustomers (i.e.,Team,Enterprise,andallAPIcustomers).” 6

Architecture and Engineering Arts, Design, Entertainment, Sports, and Media Building and Grounds Cleaning and Maintenance Business and Financial Operations Community and Social Service Computer and Mathematical Construction and Extraction Educational Instruction and Library Farming, Fishing, and Forestry Food Preparation and Serving Related Healthcare Practitioners and Technical Healthcare Support Installation, Maintenance, and Repair Legal Life, Physical, and Social Science Management Office and Administrative Support Personal Care and Service Production Protective Service Sales and Related Fraction of US workforce Transportation and Material Moving Fraction of LLM queries 0 .1 .2 .3 .4 Note:FractionofUSworkforceandClaudequeriesassociatedwithmajorSOCoccupationgroups Source: Figure3ofHandaetal.(2025) Figure1: Claudequeriesandworkforceshares Note that we do not have data on the composition of work-related queries going to other LLM tools, like ChatGPT or Gemini. Claude’s consumer market share appears to be smaller that the major competitors, though the gap appears smaller for business users.7 However, to a first order the capabilities of these models are similar across providers so we should expect broadly similar patterns to hold.8 Massenkoff and McCrory (2026) show that with updated Anthropic data and an updated methodology computer programmers andsoftwaredeveloperscontinuetobehighlyexposed. IndependentoftheAEIdata,Bick etal.(2024)showthatcomputer/mathematicaloccupationshavethehighestgenerativeAI adoptionrates,andBonneyetal.(2024)findthattheinformationsectorrankshighestinAI exposure. 7ForconsumeruseseeFigureA.1here,forindicativeevidenceonbusinessuseseehere. 8Chatterjietal.(2025)reportthat4.2percentofChatGPTpromptsacrossalltypesofconsumerplansbetween 2024and2025wererelatedtocomputerprogramming,notablylowerthanwhatHandaetal.(2025)findintheir Claudedata. ThismayreflecthigheruseofChatGPTforpersonalpurposes. Handaetal.(2025)estimateless thanaquarterofqueriesintheirdatasetarenon-work. 7

3.2 Whatoccupationsare“coders”? Havingestablishedthatcomputerprogrammingisahighlyexposedskill,weneedaprincipledwaytodefinecomputerprogramming-intensiveoccupations. Thecomputerandmathematical occupations major group is not ideal because it includes some occupations that are fairly removed from coding, such as “Information Technology Project Managers” and “ComputerUserSupportSpecialists.” Inaddition,somecoding-intensivejobsarefoundin other groups like “Computer Numerically Controlled Tool Programmers” and “Statistical Assistants.” WerelyondatafromO*NET.O*NETisaU.S.DepartmentofLabor-sponsoreddatabase thatprovidesstandardized, detailedinformationonoccupations, includingrequiredskills, tasks, knowledge, and work activities. O*NET data—in particular the tasks and work activities—havebeenwidelyusedinlaboreconomics,seeAcemogluandAutor(2011)and Eloundou et al. (2024) among many others. The tasks and work activities are very detailed and often specific to single occupations or a small grouping. Separately O*NET defines broaderskillsandratestheirimportanceacrossmostoccupations. InparticularO*NEThas a computer programming skill rating, which measures the importance of the skill for the occupation on a 1-5 scale. This is a natural fit for our analysis. The skill variable is defined for about 900 of the roughly 1,000 O*NET occupations. For the remainder we impute the skill on basis of their tasks and related occupations using text embeddings and a random forest,seeAppendixAfordetails.9 In the next section we show the most coding-intensive CPS occupations, but Table A3 9One special case is the occupation “Software developers.” They are a large occupation—accounting for more than one percent of U.S. employment—and the occupation has been growing quickly. The importance ofsoftwaredevelopersisnotedinthisrecentarticle. O*NETdoesnotprovideacomputerprogrammingskill valueforsoftwaredevelopers,andourimputationbasedontaskdescriptionsassignsthemafairlylowvalue. ThisisbecausetheO*NETtasklistforsoftwaredevelopersmakesthemseemsliketheyonlydoalittlecoding: thetasksmentionedarethingslikedesigningsoftware,meetingclients,etc. Oursenseisthatmodernsoftwaredevelopersareverycoding-intensive.Anecdotallythepeoplethatreferto themselvesas“Softwaredevs”or“softwareengineers”arehands-oncodersandnot,forexample,plannersthat onlytellotherswhattocode. InAppendixAweprovideevidencefromself-reportedjobdutiesthatthisisthe case.Consequently,weignoretheimputedskillvariableforsoftwaredevelopersandincludetheminthelistof coding-intensiveoccupations. 8

CPSCode Title EmploymentShare Cum.Emp.Share CodingImportance 1010 computerprogrammers 0.288 0.288 4.750 1020 softwaredevelopers,applicationsandsystemssoftware 1.369 1.657 3.986 1100 networkandcomputersystemsadministrators 0.137 1.794 3.620 1060 databaseadministrators 0.087 1.880 3.473 7900 computercontrolprogrammersandoperators 0.049 1.930 3.120 5920 statisticalassistants 0.011 1.940 3.000 1760 physicalscientists,nec 0.279 2.219 2.880 1240 mathematicalscienceoccupations,nec 0.233 2.452 2.854 1700 astronomersandphysicists 0.017 2.468 2.815 1000 Comp.scientistsandwebdev. 1.235 3.703 2.780 0110 computerandinformationsystemsmanagers 0.482 4.186 2.750 1220 operationsresearchanalysts 0.086 4.271 2.620 1460 mechanicalengineers 0.255 4.526 2.543 1400 computerhardwareengineers 0.057 4.583 2.500 1350 chemicalengineers 0.047 4.630 2.380 Note:TopCPSoccupationsforcodingimportancesortedindescendingorder.Horizontallineshowsourthresholdforanoccupationbeingcoding-intensive. Employmentshareistheoccupation’sfractionof2022CPSemploymentinpercentagepoints. Cum. Emp. Shareisthecumulativeemploymentshare. CodingImportanceis theO*NETcomputerprogrammingskillmetric,averagedwithineachCPSoccupationwhentheCPSoccupationsareaggregationsofO*NEToccupations. Source: CPS,O*NET,authors’calculations Table1: Coding-intensiveCPSOccupations showscodingintensityusingthefinerO*NEToccupations. 3.3 LinktoCPS We link the O*NET data to CPS so we can define the programming skill variable on the universe of CPS occupations. CPS uses (a subset of) Census occupations codes, which can be crosswalked to Standard Occupational Classification (SOC) codes but are more coarse. O*NET uses occupation codes based on the SOC codes, though O*NET’s codes are more detailed. MoredetailsofthelinkingprocessarefoundinAppendixB. There are no CPS data for October 2025. We interpolate those industry-occupation employmentcountsfromtheSeptemberandNovemberdata. Table1liststhemostcoding-intensiveoccupationsintheCPS.TheCPScodeisocc2010, the longitudinally consistent occupation code provided by IPUMS. The third and fourth columns give each occupation’s employment share and the cumulative employment share 9

respectively. We see the top ranks are dominated by computer-oriented occupations. The Employment Share column makes clear the importance of software developers (“software developers, applications and system software”). Using a programming skill threshold of 2.76, we find that the coding-intensive occupations comprise about 3.7 percent of total employment. Thisthresholdmakessureweexcludetheexplicitlymanagement-oriented“computer and information systems managers” as well as mechanical engineers, both of which strikeusaslesscoding-intensivethanthehigher-rankedoccupations. 3.4 Industries Inadditiontodefiningacoding-intensivegroupofoccupationswedefineagroupofindustries where these occupations are over-represented.10 Table 2 shows a sample of the CPS industries sorted by their intensity of coder employment: the share of their workforce that are coders (“Coders share of ind.” in the table). The second column gives the industry’s share of 2022 national employment, and the third column is the industry’s share of coder employment. Finally,thelastcolumnshowsthecumulativeshareofcodersemployed. The top industry, “Computer systems design and related services” is more than 40 percent coders and accounts for more than 30 percent of national coder employment.12 The other top industries are a mix of software/computer/high tech industries, with a mix of other sectors. We set the threshold for coder-intensive industries at 10 percent of industry employment. ThiscoversnearlyhalfofallcoderemploymentintheU.S. TheindustriesinTable2arebasedonCensusindustrycodes. TableA4shows,for2022, the NAICS codes that are mapped into each industry. The “NAICS percent of group” column is the fraction of the Census industry code employment that is accounted for by that 10IPUMSprovidesan“ind1990”longitudinally-consistentindustrycodebasedonthe1990industrycodes, whichinturnarebasedonthe1987SICcodes. Thesecodesarenotagreatmatchgivenourfocusonrecent history and the tech sector.11 Instead, we develop longitudinally-consistent codes linking together the 2012, 2017and2022versionoftheCensusindustrycodes. Welinkcodesbycoarseningeachcodethatresultsina many-to-onematch,avoidingtheneedtoallocateemploymentacrosssplit/mergedcodes. 12SeeDeckerandHaltiwanger(2024)formoreonentrydynamicsinhightechandcomputersystemsdesign inparticular. 10

Title Codersshareofind. Ind.shareofemp. Ind.shareofcoders Cum.shareofcoders Computersystemsdesignandrelatedservices 44.99 2.68 32.61 32.61 Softwarepublishers 36.04 .12 1.12 33.74 Dataprocessing,hosting,andrelatedservices 32.78 .1 .85 34.59 Scientificresearchanddevelopmentservices 24.72 .53 3.54 38.13 Computerandperipheralequipmentmanufacturing 21.32 .07 .4 38.53 Othertelecommunicationsservices 15.29 .27 1.11 39.65 Newspaperpublishers 11.74 .57 1.81 41.45 Pharmaceuticalandmedicinemanufacturing 11.41 .43 1.33 42.78 Electricandgas,andothercombinations 11.11 .07 .21 43 Othergeneralgovernmentandsupport 10.19 .1 .26 43.26 Nationalsecurityandinternationalaffairs 10.17 .66 1.8 45.06 Aircraftandpartsmanufacturing 10.07 .54 1.46 46.52 Electroniccomponentandproductmanufacturing,n.e.c. 9.75 .41 1.08 47.6 —-Misc.—-Professional,Scientific,andManagement,andAdministrative 9.69 .06 .17 47.77 Soap,cleaningcompound,andcosmeticsmanufacturing 9.51 .09 .24 48.01 Bankingandrelatedactivities 9.35 1.39 3.51 51.52 Communications,andaudioandvideoequipmentmanufacturing 9.33 .06 .16 51.68 Non-depositorycreditandrelatedactivities 9.22 .77 1.93 53.61 Navigational,measuring,electromedical,andcontrolinstrumentsmanufacturing 8.88 .1 .24 53.84 Note: Top CPS industries in terms of coder intensity, sorted in descending order. Employment share is the occupation’sfractionof2022CPSemploymentinpercentagepoints. Source: CPS,O*NET,authors’calculations Table2: Industriessortedbyintensityofcoders NAICS. 3.5 OtherMeasuresofAIExposure Much of the recent literature on AI and the labor market uses the exposure measures calculated by Felten et al. (2023) and especially Eloundou et al. (2024). Eloundou et al. (2024) used GPT-4 to impute each occupation’s generative AI exposure based on ONET detailed workactivities. TheyconfirmedthatGPT-4’sclassificationslargelyalignwithcrowdsourced humanclassifications. Eloundou et al. (2024) is an impressive and influential early contribution. However, being developed so early it does have disadvantages. In particular, the exposure metrics predatethepublicreleaseofGPT-4,4o,allClaudemodels,subsequent“reasoning”models such as o3 and coding tools like Claude Code and Codex. Notably, the knowledge cutoff for GPT-4 was September 2021, well before even the release of ChatGPT and major image generation tools like Midjourney and Stable Diffusion. This means that GPT-4 had little in itstrainingdatatohelpitjudgeAIexposure,forcingittorelyheavilyonthepromptswhere 11

Eloundouetal.(2024)describegenerativeAIcapabilities. We believe that real-world AI use data—like the AEI—has some important advantages over other exposure measures. By its nature, the AEI shows how people are actually using AIintherealworld. Atthesametime,AEIalsohassomeblindspots. Itonlycoverscurrent use and perhaps can’t flag occupations which are likely to be highly exposed in the near future. ItisalsolimitedtotheusageofthepublicversionoftheClaudechatbot,whichcould omit genAI exposure through, e.g., customized customer-service representative assistants andspecializedimage-generationtools. ItcouldbethattheEloundouetal.(2024)metricis bettersuitedtounderstandingwhichoccupationsarelikelytoeventuallybeexposed,while theAEI measuretells uswhich havealreadybeen exposedand thusmay havemeasurable changesinoutcomes. To understand the measurement landscape we compare the “GPTs are GPTs” exposure measure (Eloundou et al., 2024)—which we call “GPTs exposure”—with the AEI exposure measure using ordinal rankings. For both metrics we divide occupations into (1) a mostexposedgroupingaccountingforabout20percentofemployment,and(2)theremainder.13 This is motivated by the literature (e.g. Brynjolfsson et al. 2025) that focuses on the top quintileofAIexposure. Table3showsthedistributionofemploymentacrossthesegroups. Focusingonthefirst column we see that just over half of high-GPTs exposure employment is also classified as high-AEI exposure. Likewise, in the first row we see just over half the workers which are high-AEI exposure are also high-GPTs exposure. While it is encouraging that there is significantoverlapbetweentheGPTsandAEImetrics,nonethelessitistroublingthatthetwo disagreeabouthalfthetimeastowhichoccupationsareinthehighestexposuregrouping. The first row of Table 4 shows what fraction of coding occupations are high-exposure undertheAEIandGPTsgroupings. Thecodingoccupationsareoverwhelminglyinthehigh 13This is complicated by the fact that Handa et al. (2025) use ONET data based on 2010 SOC occupation codesandEloundouetal.(2024)using2018SOCcodes. Wecrosswalkthetwotoacommonsetofmorecoarse occupationcodestakingtheunweightedaverageofexposurewithineachcoarsecode. 12

(1) (2) GPTsexposurehigh GPTsexposurelow AEIexposurehigh 10.01 8.79 AEIexposurelow 9.92 71.23 Note: CellsshowthepercentofU.S.privateemploymentineachclass. “GPTsexposure”usesEloundouetal. (2024)’sGPT-β metric, “AEIexposure”isbasedonHandaetal.(2025). Highexposureoccupationsarethose in the (approximate) employment-weighted top 20 percent of exposure, “low” exposure occupations are the remainingroughly80percent. Source: Eloundouetal.(2024),Handaetal.(2025),OES,authors’calculations Table 3: Distribution of Employment by generative AI exposure (Percent of total employment) exposuregroupaccordingtobothmetrics,withmorethan98percentofcoderemployment inthehighestquintiles. Recallthatthecodingoccupationsgroupingwasmotivatedinpart by examination of the AEI data, so it is not totally surprising that coding occupations tend tobehighlyexposedthere. Regardlessitisusefulconfirmationtoseethattheprogramming skill classification ends up agreeing both with the AEI metric and the (more independent) GPTsmetric. Forreference,thesecondrowofthetableshowsthatnon-codingoccupations are far less likely to be considered high exposure, with only about 17-18 percent of noncodersendingupinthosegroups. In summary, we find—like Gimbel et al. (2025)—significant differences between Eloundou et al. (2024) and Handa et al. (2025) in terms of which occupations are highly exposed togenerativeAIthoughbothmetricsagreethatourcoderoccupationsarehighlyexposed. (1) (2) GPTsexposurehigh AEIexposurehigh Codingoccupations 99.5 98.2 Non-codingoccupations 18.0 16.9 Note: Row one shows the percent of coding employment that falls into the high exposure groups reported above. Row two shows the percent of non-coding employment falling into those groups. “GPTs exposure” uses Eloundou et al. (2024)’s GPT-β metric, “AEI exposure” is based on Handa et al. (2025). High exposure occupationsarethoseinthe(approximate)employment-weightedtop20percentofexposure,“low”exposure occupationsaretheremainingroughly80percent. Source: Eloundouetal.(2024),Handaetal.(2025),OES,authors’calculations Table4: Percentofcoding/non-codingemploymentfallingintohigh-exposuregroups 13

1.1 1.05 1 .95 .9 xednI Total CPS employment CES 2014m1 2016m1 2018m1 2020m1 2022m1 2024m1 2026m1 Note:IndexedlevelsoftotalCPSemployment(NSA)andCESemployment(SA).VerticallineindicatesChatGPT releasedate Source: CPS,CES Figure2: CPSandCESindexedlevels 3.6 Normalizations Whenlookingatemploymentlevelsitisimportanttorememberthattrendemploymentin CPShassometimesbeenatoddswithdatafromothersources. Figure2showsanindexof (NSA) CPS employment with the published (SA) CES levels. Neither has been adjusted to accountfortherelativelysmallscopedifferencesbetweenthetwo. WhileCPSgrewslightly slowerthanCESpre-covid,alargergaphasopenedpost-CovidwithCPSonlygrowing0.3 percent between March 2023 and March 2024 while CES grew 1.2 percent during the same period. Note that CES is benchmarked to the QCEW comprehensive administrative data through March 2025 so it should not suffer from much error over this period (though we don’tyetknowhowaccurateitisforpost-March2025.) ToaccountforthisdivergencewecalculatetrendsasemploymentsharesinCPSandthen convert them to levels using CES employment totals. This fairly crude adjustment ensures 14

that our numbers are consistent with CES aggregate growth rates. In the absence of this correction many occupation groups would incorrectly show a deceleration in employment post-Covid. 4 Regression Specifications Our basic approach is to test whether coder employment growth changed with the introductionofChatGPTinNovember2022. Ourbaselinespecificationis lne = α+β ·1{t ≥ 2022m11}+β ·t+β ·1{t ≥ 2022m11}·t+ε (1) t 1 2 3 t wherelne islogmonthlyemploymentofcoders,β showsthelevelshiftinlogmonthly t 1 employmentpost-ChatGPT,β ispre-ChatGPTtrendgrowth,andβ capturesthechangein 2 3 growth post-ChatGPT. We normalize the dependent variable so that the coefficients can be interpretedasannualgrowtheffectsinpercentagepoints. We run these regressions both for all industries and separately for coder intensive and non-coderintensiveindustries. Inaddition,weshowrobustnessbydroppingtheCovidera (onthetheorythatlabormarketsweredislocatedinthisperiod)andalternativelydropping 2022 and 2023 (since any LLM/ChatGPT effect might not be dated precisely to November 2022.) InSection5.2wedevelopacounterfactualemploymentseries Z ,meanttocaptureonly t shifts in employment due to industry-level shocks, not occupation-specific shocks. Then lne −lnZ istheoccupation-specificshocktoemployment. TotestwhetherChatGPTgent t eratedanoccupation-specificshockwerun(1)withthenewdependentvariable: lne −lnZ = α+β ·1{t ≥ 2022m11}+β ·t+β ·1{t ≥ 2022m11}·t+ε (2) t t 1 2 3 t 15

7000 6000 5000 4000 3000 sboj fo sdnasuohT All Industries Coders 3000 2016−2019 trend 2500 2000 1500 2014m12016m12018m12020m12022m12024m12026m1 sboj fo sdnasuohT Coder Intensive Industries Coders 3500 2016−2019 trend 3000 2500 2000 2014m12016m12018m12020m12022m12024m12026m1 sboj fo sdnasuohT Non Coder Intensive Industries Coders 2016−2019 trend 2014m12016m12018m12020m12022m12024m12026m1 Note:EmploymentlevelofcoderoccupationsbasedonO*NETprogrammingskill.DashedlineshowsChatGPT releasedate(November2022.)Coderintensiveindustriesarethosewhereover10percentofworkersarecoders, non-coderintensiveindustriesaretheremainder. Source: O*NET,CPS,authors’calculations Figure3: Coderemployment 5 Results Figure 3 shows coder employment in all industries, coder intensive industries, and noncoderintensiveindustries. RecallthisisbasedonCPSemploymentshares,adjustedsothat total CPS employment tracks CES employment. The vertical dashed line marks November 2022, the release date of ChatGPT. Our question is whether the employment of coders has been notably different after that date. Focusing on the first panel, there does appear to be a kink in employment growth, with pre-ChatGPT rapid growth flattening out. While employment is still fairly close to the pre-Covid linear trend the kink clearly suggests a change in employment dynamics around the introduction of ChatGPT. Turning to coder intensive industries in the middle panel, this pattern is more stark. Coder employment in theseindustrieshasbeenessentiallyflatsincelate2022. Thepatterninnon-coderintensive industriesisdifferentwithemploymentveryclosetothepre-Covidlineartrend. 16

5.1 InitialRegressions Table 5 shows the regression results. The first three columns show results for all industries,usingthefullsample,droppingtheCovidera,anddropping2022-2023respectively.14 Thedependentvariableislogemploymenttimes1200,sothecoefficientscanbeinterpreted as annual growth rates. To account for autocorrelation we use Newey-West standard errors, allowing for 16 lags. This is long enough both to account for seasonality and rotation groupissues(individualsleavethesurveyafter4monthsin,an8monthbreak,andafinal4 monthsin.) Weseefromthetrendcoefficient(thelineartimetrend)thatcoderemployment averaged an annual growth rate of about 4.8 percent pre-ChatGPT. This is very fast; total privateemploymentonlygrewatapaceofabout1.3percentoverthistime. Thecoefficient ofinterest,“Post-GPT*Trend”,isthechangeintrendgrowthpost-ChatGPT,whichshowsa substantialslowingofgrowth. Columns2,5,and8showthattheresultsarenotsensitivetodroppingtheCovidperiod. Further,columns3,6,and9showthattheresultssurvivedroppingtheyears2022and2023. This last exercise is relevant given the uncertainty about when exactly an AI shock could be said to arrive: Coding tools based on GPT 3.5 (like Github Copilot) were available well before ChatGPT, on the other hand it is likely that business reactions would have not been instantaneous. Theexerciseshowsthattheresultsarenotsensitivetotheexacttimingofthe shock. Theresultsforcoder-intensiveindustries(columns4-6)areevenstrongerthanthosefor all industries. Pre-ChatGPT growth was about 6 percent annually, and became flat or negative post-ChatGPT (though again the No 2022/3 specification shows less impact.) Finally, thenon-coderintensivesectorshowssomewhatsmallerchangesingrowth. Takentogether theseregressionsconfirmtheimpressionfromFigure3. 14ThenoCovidsampledrops2020m1through2021m6. Mostadultswerevaccinatedby2021m6andmasking/distancingguidancewasbeingrelaxed. 17

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend 4.882*** 4.871*** 4.814*** 6.110*** 5.981*** 6.213*** 3.883*** 3.977*** 3.680*** (0.270) (0.304) (0.412) (0.464) (0.489) (0.631) (0.207) (0.204) (0.288) Post-GPT*Trend -3.884*** -3.872*** -3.339*** -6.248*** -6.118*** -4.063*** -1.970*** -2.064*** -2.759** (0.371) (0.410) (0.871) (0.680) (0.750) (0.805) (0.549) (0.593) (1.105) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note: Coefficientsareannualizedlogpoints,standarderrorsareNewey-West. Dependentvariableismonthly (log)coderemploymentintherelevantindustries.Post-GPTperiodstartsinNovember2022. Table5: Regressions,nocontrols 5.2 IndustryShocks The assumption implicit in Table 5 is that no other factors changed the trajectory of coder employmentaround2022. Thisisastrongassumption,especiallysincecoderemploymentis fairlyconcentratedincertainindustries(seeSection3.4). Industry-levelshocksarecommon, implying that occupational employment is partially a function of industry shocks. Indeed, the popular Bartik or shift-share instrument depends on industry-level shocks being fairly large in some sense (Borusyak et al., 2025; Goldsmith-Pinkham et al., 2020). It is possible that the industries coders work for happened to cut employment (growth) around 2022 for reasons unrelated to AI. If this were the case we should see coder employment track employmentofthenarrowindustriestheyworkin. Putdifferently,industriesshouldnot(on average) change the coder’s share of industry employment, even if some industries shrink employment totals. This suggests a counterfactual: calculating what coder employment growthwouldhavebeenonthebasisofpurelybetween-industryfactors;ifitwaspurelythe exposure-weighted average of industry growth. Then the remainder—the within-industry component—istheoccupation-specificshock. Formally,wecanstartwiththeaccountingidentity N ∑ g = s g (3) .,o,t i,o,t−1 i,o,t i=1 18

wheres = ei,o,t−1 isthefractionofoccupationo’semploymentthatfallsintoindustry i,o,t−1 e .,o,t−1 i in t−1, g is the growth rate of employment in occupation o and industry i, and g is i,o,t .,o,t the aggregate growth rate of occupation o. Adding and subtracting industry level growth weobtain N N ∑ ∑ g = s (g −g ) + s g (4) .,o,t i,o,t−1 i,o,t i,.,t i,o,t−1 i,.,t i=1 i=1 (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) Within-industry Between-industry Herethewithin-industrycomponentaddsupthedeviationsoftheoccupation-industry growth rates from their corresponding industry rates. In other words, it counts the degreetowhichindustriesbecomemoreorlessoccupationo-intensive. Thebetweenindustry componentcapturestheexposure-weightedaverageofindustrygrowth. Thisiswhatoccupation o’s growth would have been if all industries kept constant occupation shares. This type of analysis has been used by, among many others, Acemoglu and Autor (2011) for industry/occupation margins in the context of polarization, Katz and Murphy (1992) for analyzing wages, Davis and Haltiwanger (1992) for decomposing excess job reallocation, andJaimovichandSiu(2020)foraccountingforjobpolarization. Note that equation 4 is an identity—it holds by construction. Our identifying assumption is that in the absence of occupation-specific shocks the within-industry component is zero. Thisistrueif(intheabsenceofoccupationshocks)industriesgrowandshrinkhomothetically, keeping occupation shares constant. This appears to be a reasonable approximation,thoughaformaltestisbeyondthescopeofthispaper. InSection7wepresentevidence thatinseveralhistoricalepisodesthisdecompositionhasprovidedthecorrectintuition. Industry-levelshocksmaytakeanumberofforms: changesindemandforindustryoutput,changesinindustryTFP,changestofinancialconstraints,andregulatoryortaxchanges are all candidates. Any of these factors would (generally) scale industry labor demand up or down but would not affect the optimal occupational mix. Occupation-specific shocks, 19

ontheotherhand,wouldbechangesintechnologyorpossiblyregulationsthatchangethe optimaloccupationalmix. We cumulate the counterfactual growth rate z = ∑N s g into an index and .,o,t i=1 i,o,t−1 i,−o,t normalizeittomatchtheobservedseriesinNovember2022: τ=t Z ∝ ∏ (1+z ) (5) o,t .,o,τ τ=1 (cid:32) (cid:33) τ=t N ∝ ∏ 1+ ∑ s g (6) i,o,τ−1 i,−o,τ τ=1 i=1 Notethatthisisnotexactlythebetween-industrycomponentoftheaccountingidentity above: here we sum up the industry growth rate excluding occupation o, g , rather than i,−o,t theactualindustrygrowthratesg . Thisconservative“leaveoneout”approachavoidsthe i,.,t mechanicalpositivecorrelationbetweenz andg whenoccupationshockscaninfluence .,o,t .,o,t industry growth.15 Figure 4 shows the results. Focusing on the first panel, the counterfactual matches observed employment remarkably well up to the start of the Covid era. At that point the series diverge some, and—importantly—diverge further post-ChatGPT. This suggeststhatpost-ChatGPTcoderemploymenthasnotkeptpacewithemploymentgrowth intherelevantindustries. Astrongerversionofthispatterncanbeseeninthesecondpanel. Heretheuniverseisrestrictedtoindustrieswhereatleast10percentofworkersarecoders (“coder intensive industries”) and the exposure and counterfactual calculations are subject tothesamerestriction. Theinterpretationisthatpost-ChatGPT,eveninthiscoder-intensive group of industries, industry employment grew faster than coder employment. The third panelshowsthatfornon-coderintensiveindustriesthecounterfactualappearstofollowobservedemploymentbetter. However,notethattheslopeofthecounterfactualisstillgreater thantheobservedslopepost-ChatGPT. 15Thedecompositionisalsodependentonthedefinitionoftheindustriesused.Toavoidnoisewerequirethat anindustryhaveatleast10CPSrespondentsineachmonth. Industriesfailingthiscriterionarelumpedinto supersector-level“miscellaneous”groupings.TheAppendixshowstheresultsarenotsensitivetothethreshold. 20

7000 6000 5000 4000 3000 sboj fo sdnasuohT All Industries Coders 3500 Counterfactual 3000 2500 2000 1500 2014m12016m12018m12020m12022m12024m12026m1 sboj fo sdnasuohT Coder Intensive Industries Coders 3500 Counterfactual 3000 2500 2000 2014m12016m12018m12020m12022m12024m12026m1 sboj fo sdnasuohT Non Coder Intensive Industries Coders Counterfactual 2014m12016m12018m12020m12022m12024m12026m1 Note:EmploymentlevelofcoderoccupationsbasedonO*NETprogrammingskill.DashedlineshowsChatGPT releasedate(November2022.)Coderintensiveindustriesarethosewhereover10percentofworkersarecoders, non-coderintensiveindustriesaretheremainder. Source: O*NET,CPS,authors’calculations Figure4: Coderemploymentwithcounterfactual Table 6 shows the regression results for specification (2). The dependent variable is the difference between observed log employment and the counterfactual. This difference captures occupation-specific changes in employment: changes that cannot be explained by exposure to industry dynamics. Strikingly, the post-ChatGPT coefficients are all significant andclusteredbetween-3and-4percent,withourbaselinespecification(column1)showing an estimate of -3.23. The interpretation is that an occupation-specific shock has been suppressing coder employment growth by (conservatively) about 3 percent per year since the introduction of ChatGPT. This is consistent with ChatGPT leading to losses of (potential) coderjobsoverthatperiod. Onepossibilityisthatotherchangestotheeconomycausedthegrowthrateofemployment to change. After all, 2022 was still in the midst of post-Covid reopening. To address thispossibilitywerunaplacebo,pickingagroupunlikelytobeaffectedbyLLMsandseeing iftheiremploymentgrowthappearstochange. Weusethebottom(employmentweighted) quintile of AI-exposed jobs from Handa et al. (2025) as this group. Table 7 shows that the post-GPT slope term is generally insignificant and has mixed signs across specifications. Onlyinthesampleexcluding2022and2023isitconsistentlynegativeandsignificant. Even 21

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -0.962*** -1.184*** -0.849* -2.359*** -2.797*** -1.967*** 0.616*** 0.579*** 0.533* (0.331) (0.300) (0.437) (0.642) (0.568) (0.743) (0.206) (0.211) (0.287) Post-GPT*Trend -3.222*** -3.000*** -4.427*** -3.561*** -3.123*** -5.943*** -3.016*** -2.978*** -3.410*** (0.635) (0.598) (0.850) (1.072) (0.998) (1.162) (0.414) (0.420) (0.818) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table6: RegressionsControllingforCounterfactual Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -3.880*** -3.782*** -3.853*** -2.997*** -3.140*** -2.614*** -3.893*** -3.779*** -3.883*** (0.172) (0.222) (0.236) (0.521) (0.555) (0.520) (0.172) (0.217) (0.242) Post-GPT*Trend 0.328 0.229 -0.955** -1.435 -1.292 -2.897** 0.409 0.296 -0.864* (0.408) (0.487) (0.385) (0.889) (0.949) (1.411) (0.408) (0.486) (0.449) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table7: RegressionsControllingforCounterfactual: Placebogroup(lowAIexposure) inthosespecificationsthecoefficientishalfthesizeorlessthanthecorrespondingcoefficient fromTable6. 5.2.1 Discussion The3percenteffectsizeislarge. Wecautionagainstinterpretingthisasasimplecausaleffect given the complexity of AI’s potential economic effects and the measurement and identification challenges. However, to give some context we can do some back of the envelope calculations. Cumulating over the roughly 3 years since November 2022 and using 5.735 millioncoderjobsasthebasevalue,theimplicationisthatroughly500,000additionalcoder 22

jobswouldhaveexistedintheabsenceoflarge-scaleLLMuse.16,17 Brynjolfssonetal.(2025) do not provide an estimate of aggregate job losses due to AI, and their methodology is focused on measuring relative gaps between young and old workers. Nonetheless, treating theireffectsizesasjoblosseswecanapproximateanaggregateeffect. Theyfindthatamong 22-25 year olds employment in the top two quartiles of AI exposure fell about 12 percent relative to employment in the bottom quartile. Starting from total private employment of 130 million, and assuming about 7.6 percent of the workforce is 22-25 years old (based on the CPS), a 12 percent job loss for two quartiles works out to about 475,000 jobs lost. In this sense a crude interpretation of our estimates are consistent with a crude translation of Brynjolfssonetal.(2025)’sestimates(whichagainareonlyestimatesofrelativejoblosses.) For a number of reasons, we do not interpret the results as evidence that AI has eliminated500,000jobsfromtheeconomy. First,manycoderswouldhavefoundjobs—andpossiblygood,well-matchedjobs—inotheroccupations. Ifoccupationaltaskmixanddemand forotheroccupationsisstablethenthosejobsmaybenotasgoodmatchesonaverage. But AI may be altering the task composition of occupations, such that a potential coder today may go into a management or other occupation that now uses more of their coding skills. Inaddition, theseestimates—likethoseinotherstudies—donotattempttoaccountforthe effectsofautomationonaggregateproductivityandlabordemand. Instandardmodelsthe average worker is better off after a positive productivity shock, both from cheaper output and usually from increased aggregate labor demand, even if the displaced workers suffer persistentearningslosses. Thatnotedoccupationalemploymentmayfall,especiallyifAIis asubstituteforlabororthedemandfortherelevantoutputisinelastic. Even if coder employment is well below where it would have been, the estimates in the first three columns Table 5 show that coder employment has continued to grow post- ChatGPT. In this sense the raw statistics do not support a view that coder employment is 16Nonfarmemploymentwas158millioninNovember2022,ifthecodershareis3.7thatmeanstherewere about5.8millioncoders.A3percentannualdeclinefor3yearsimplies5.8·0.973 ≈5.29millionjobsremaining. 17Obviously,wearenotlikelysee500,000additionalunemployedcodersinthedata. Mostofthoseworkers wouldhavebeenabsorbedintootheroccupations. 23

collapsing. This is consistent with data from Indeed showing job openings for software developers mostly flat from 2024 and even edging up recently (after falling more than 50 percentover2022-2023). If firms have reduced coder employment relative to the counterfactual, it is not clear whether it is due to observed productivity gains or anticipated gains. In a model with fixed hiringandfiringcostsafirmthatanticipatesAIsubstitutingforlaborinthenearfuturemay freeze hiring and stay in the band of inaction for a time, even before actual productivity increases. Thus,decliningemploymentgrowthdoesnotimplyactualproductivitygainsin theshortrun,thoughitstronglysuggeststheyareattheminimumanticipated. Examining the first panel of Figure 4, it can be seen that the gap between coder employment and the counterfactual does not open immediately after the release of ChatGPT. Indeed,thegaponlywidenssignificantlyinthemiddleof2024. Thispatternappearsplausible. Itisunlikelythatmanyfirmswouldhaveinstantlychangedemploymentpolicyupon the announcement of ChatGPT; more likely it would have taken time and observation of model progress to commit to changes. However, the apparent timing is not quite consistent with a formal Bai-Perron test for structural breaks (Bai and Perron, 2003), as seen in Figure A1 The Bai-Perron test detects three breaks in the difference between monthly (log) coder employment and the counterfactual (log) employment series. This includes 2022m6 (justbeforetheChatGPTrelease),and2020m3(reflectingCovid). Thetestalsofindsabreak at2017m10, whichmightreflectsamplingnoiseintheCPSandlacksaclearinterpretation. WhilethebreakduringCovidisexpected,thebreakin2022m6islessintuitive. Takenatface valueitcouldbeevidencethatfirmswereadjustingtotheavailabilityofGithubCopilot(released October 2021) and the initial version of OpenAI Codex (released August 2021). Our senseisthatthisisunlikely,anditismoreplausiblethattheBai-Perrontestisconflatingan end-of-Covid break with a post-ChatGPT break. The noise in our relatively small samples makeithardtodiscernmuchmore. One question is to what extent AI is a positive demand shock for coders, and insofaras 24

training and deploying LLMs require coders. We think this isn’t likely to be a major factor yet. Taken together, employment at OpenAI, Anthropic, and Google Deepmind is likely under 15,000, and many of those workers are not coders.18 Even if we multiply that by six toaccountforstartupsandtheAIgroupsatMeta,Microsoft,andelsewherewewouldstill onlyaccountforlessthantwopercentofU.S.coders. 5.2.2 Taxlawchanges The counterfactual controls for industry-level factors that may affect coder employment. However,aremainingthreattoidentificationisoccupation-specificfactorsunrelatedtoAI. Onecontemporaneouspolicychangewastheimplementationofthe2017TaxCutandJobs Act(TCJA).Startingin2022thislawmeantthatR&Dexpensescouldnolongerbeexpensed immediatelyandinsteadhadtobeamortizedoverseveralyears. ThischangeraisedtheeffectivecostofR&Dspending,whichisrelevantbecausesoftwaredevelopment—including developer salaries—is typically considered an R&D activity. The Information sector accountsforaboutonequarterofR&Dspending.19 TheTCJAimplementationthusmayhave differentiallyaffectedcoderhiringwithinthesetofrelevantindustries. Evidenceonthischannelissparse. Cowxetal.(2025)findthelawreducedR&Dspending,whileDuandLi(2025)—usingabroadersampleoffirmsandalongertimerange—find noreductionininternalR&DandrisingexternalR&D.MuchofsoftwareR&Dspendingis done by the industries that are coder-intensive according to our classification. With this in minditisencouragingthatourresultssurviveonthenon-coder-intensiveindustrysample, where the TCJA changes would perhaps be less salient. Regardless, more research is likely needed. 18PublicreportsputOpenAI’semploymentaround4,000,Anthropicat2,500,andDeepMindat6,000. 19SeeNSFdatahere. 25

5.2.3 Parametricassumptions Ideally we would like to use a more local and/or non-parametric approach to identifying theeffectofAIonjobs. Unfortunately,Covidlikelyrendersthisunworkable. TheCovid-era labormarketdistortionsmeanthattheperiodjustpriortoNovember2022ishighlyunusual, andunusualinwayswewouldexpecttobetransitory. Giventhesefactsthenextbestoption is to examine the pre-Covid period. Encouragingly for our approach, coder employment was growing more of less linearly pre-Covid; even better coder employment had reached itspre-Covidtrendbylate2022. Thisissuggestiveevidencethatcoderemploymentwason itshistoricaltrajectoryandwouldhaveremainedsointheabsenceoftheLLMshock. While lessthanideal,wethinkitbettertolooktothelonger-run,pre-Covidtrendforidentification ratherthanrelyonlocaltrendsaround2022whichwillbedistortedbyCovid. 5.2.4 Wages If AI is decreasing demand for coders it may also be visible in wages paid. However, to the extent that the composition of workers changed (the main point in Brynjolfsson et al. (2025)andLichtingerandHosseiniMaasoum(2025))thosewagechangesmaybeswamped byhavingalargershareoftheworkersbeolderormoreexperienced. Lookingthatthelog average wages of coders (winsorized at 1 percent) no break in 2022 is visible. While more workonthisislikelynecessary,weconcludethatthemaineffecthasbeenonthenumberof codersemployed,nottheirwages. 5.2.5 Robustness TheresultsinTable6showrobustnessacrossindustrygroupsandsampleperiod. However, thereareotherpossiblyrelevantdimensions: Definitionof“coders” We adjust the threshold for coding-intensive jobs to make coders approximately25percentsmallerandlargerasashareoftheworkforce. Inthestricter 26

case this amounts to dropping a single large occupation from the group. Tables A5 and A6 show the results. Nearly all the signs remain negative, though statistical significance is lost in about the half the specifications. We take this as validation of the methodology. As described above we took considerable care to construct a grouping ofoccupationsthatissufficientlylargetoallowforestimationandsufficientlyfocused to show any treatment effects, so we don’t necessarily expect the results to survive changes to the definition of coders. To explore further we halve and double the size of the coder group in Tables A7 and A8. These tables show a similar pattern, with statisticalsignificanceoftengonebutmostsignsremaining. Softwaredevelopers Aswenoted,softwaredevelopersarebothalargeoccupationalgroup andonewheretheO*NETtaskdescriptionsdonotindicateveryhighcodingintensity, thoughthesurveyrespondentjobdutywriteininformationdoesindicatehighcoding intensity. Giventhepotentialambiguityofthisgroup,TableA9showstheresultsifwe exclude software developers from the coders grouping. The results are qualitatively unchanged. Levelsvs. shares The main results are for CPS employment levels, scaled to match CES totals. Table A11 shows results for shares of CPS employment. In other words employmentineachoccupation(oroccupation-industrypair)isfirstconvertedtoashare of total CPS employment for the month, then the totals and counterfactuals are calculated. The results are exactly identical to the baseline, since the dependent variableisthelogdifferencebetweenobservedemploymentandthecounterfactualseries. Conversion to shares means that employment and the counterfactual have both been divided through by a month-specific term, which cancels out when we log and take differences. Thusthedependentvariableonlydiffersbyaconstant. Growthcalculations Recall that equation 6 used a leave-one-out growth calculation, excluding coder employment when calculating industry level growth. In Table A10 we 27

use total employment instead of leave-one-out and the results are qualitatively unchanged. Smallindustries As noted in Section 5.2 the decomposition depends on the industry definitions. We experiment with the threshold below which industries get lumped into catch-all supersector groupings. The baseline threshold is to have at least 10 respondents in every month; we try setting the threshold to zero and 50 in Tables A12 and A13showtheresultsarequalitativelyunchanged. E-commerce The 2022 NAICS code revision eliminated the direct selling NAICS code and reclassified those establishments under the same NAICS as brick and mortar stores sellingthesameproducts. Thischangeeffectivelyswitchese-commercefromitsown NAICS industry to being pooled with a variety of physical retailers. Given that ecommerce is associated with the tech industry and coders it is possible this change would affect our results. Since the 2022 NAICS code change was implemented in CPS in 2025, we truncate the data in 2024 and reconstruct time-consistent industries throughthatyear. TableA14showstheresultsarequalitativelyunchanged. ComputerSystemsDesign As noted above, 40 percent of coders work in computer systems design and related services (NAICS 5415). To check if the results are sensitive to the idiosyncracies of this industry we exclude it from the calculations. Table A15 showstheresultsarequalitativelyunchanged. CPSweights Each January the CPS updates population controls without reestimating the historical data, which creates a discontinuity. To see if this affects our results we use the smoothed weights developed by Coglianese et al. (2025). Table A16 shows the resultsarequalitativelyunchanged. 28

6 Coder intensive industry dynamics The previous section showed that industries substituted away from coders after the introduction of ChatGPT. The fact that narrow industries were substituting away from coders meansthatanindustry-levelshockcannotexplainthecoderemploymentdeclines. Nonetheless, it is informative to understand what industry-level influences have been in play recentlyandtowhatextentgrowthintherelevantindustriesisunusual. Recalling that the coder intensive industries are dominated by software, software design, and internet-centric industries, there are a number of other factors that could explain declining employment growth in the sector. Interest rates began to rise in March 2022, potentiallysqueezingover-extendedcompanies. Perhapsmoreimportantly,itappearsthatthe re-openingafterCovidcausedareassessmentofthesector,asspendingfromonlineservices movedbacktootherformsofconsumption. Forexample,advertisingrevenuefellatGoogle andMetain2022,consistentwiththere-openingmovingpeopleawayfromonlineservices. More tangentially, prior to 2022 there had been a wave of investor interest in cryptocurrencies but by 2022 the price of Bitcoin was falling, contributing to the collapse of FTX in November2022andgenerallyincreasedpessimismabouttheprospectsofblockchaintechnology. Finally,thetreatmentofR&Dfortaxpurposesalsochangedin2022—requiringdomesticR&Dspendingtobeamortizedoverfiveyearsratherthandeductedimmediately— thoughitisnotclearifthischangehadanymaterialimpact. Inshort, 2022wasayearthatbroughtmanychangestothecodingintensivesectorand estimating the contribution of each factor is beyond the scope of this paper. However, we cansaysomethingaboutlevelofemploymentrelativetothepre-Covidtrend. For this exercise we use BLS’s Occupation, Employment and Wage Statistics (OEWS), a survey of businesses that estimates the joint distribution of occupations across industries. Unlike CPS, it is low frequency: conducted semiannually and the published data averages 3yearsworthofsurveys. OEWSlikelyhasmoreaccurateestimatesbothbecauseofalarger effectivesamplesize(sinceitcantargetestablishmentsthatarelarger)andbecauseitrelies 29

Industries Pre-Covid 2020m2-2022m10 Post-ChatGPT Total private 1.62 .67 .83 Information (NAICS 51) 1.21 2.41 -2.76 Coder-intensive agg. 3.78 4.66 -1 Comp. Design (NAICS 5415) 3.25 3.87 -1.25 Software (NAICS 5132) 8.79 10.42 .05 Table8: CESindustrygrowth on the firm to report both occupation and industry (worker reporting of industry may be noisy). OEWSalsoreportsdatawithsomewhatmoreindustrydetailthanCPS.IntheOEWS we approximate our coder occupation group. Similar to our CPS methodology we define thecoderintensivesectorasthoseNAICSwithaworkforceofmorethan10percentcoders. WithourcoderintensivesectordefinedwelinkthedatatotheCES.CESreportsmonthly employmentbydetailedNAICScode. ThesampleisfarlargerthanCPSmakingitabetter gauge of employment levels.20 CES also reports with time-consistent NAICS codes. Figure 5 shows indexed employment for a variety of industry groupings. It is evident that our coder intensive aggregate (green) has been been growing much faster than total privateemployment. Theinformationsectorisoftenusedasaproxyforhigh-tech/computercentric activity—this comparison shows the limits of that approach. Information (in red) wasgrowingslowerthantotalemploymentpre-Covid. ThatreversedduringtheCovidera, butdeclinessince2022nowputinformationbelowits2019level. Finally, forreference, the software industry (which is a component of the coding-intensive sector) was growing extraordinarily fast up until 2022. Table 8 shows growth of over 8 percent pre-Covid for the softwareindustryfollowedbynearlyflatgrowthpost-ChatGPT. We can explore this further in the CES data. Within CES we partition employment into 3 digit NAICS industries and a handful of 4 digit industries. Figure 6 shows a scatterplot of average post-2020 employment growth for these industries against average 2016-2019 20CESisalsobenchmarkedannuallytothecomprehensiveQCEWdata,meaningthatexceptformostrecent monthsthetrendsshouldbecorrect. 30

1.4 1.2 1 .8 .6 8102 ni 1= ,xedni tnemyolpmE Total private Information sector Coder−intensive aggregate Software (NAICS 5132) 2014m1 2016m1 2018m1 2020m1 2022m1 2024m1 2026m1 Note:DashedlinemarksNovember2022 Source: CES,OES,authors’calculations Figure5: NormalizedCESEmployment 31

10 5 0 −5 −10 egareva dezilaunna ,htworg 7m5202−0202 3 digit industries Coding intensive sector 45 degree line Linear prediction −10 −5 0 5 10 2016−2019 growth, annualized average Note:AnnualizedCESgrowthof3digitNAICSindustriesandthecoding-intensiveindustrygroup. Source: CES,OEWS,authors’calculations Figure6: Industrygrowthscatterplot growth. Thelinearpredictionlineshowsthatthereissomemeanreversion: Industriesthat grewquicklypre-Covidcontinuedtogrowfastpost-Covid,thoughnotquiteasfast. Thered dot marks our coder intensive sector aggregate: it is exactly where we expect an industry to be given its pre-Covid trend. In other words coder intensive industry employment is right in line with where it “should” be, based on how industry employment evolved post- Covid. If the industry continues to have sluggish growth this will cease to be the case, but forthemomentthelevelofemploymentisconsistentwithareturntotrendfollowingCovid overhiring. 7 Evaluating the Decomposition Our results depend on the between-industry control variable correctly capturing industrylevel shocks but not occupation-specific shocks. We can examine historical episodes to es- 32

tablish the reliability of the control. We use the data from Deming (2017), which is based on Census data from 1980 through 2012. The Census sample sizes are larger than the CPS. Importantly,thedatahavemoreorlessconsistentoccupationandindustrycodingschemes overtime. For several example occupations we repeat the exercise from Section 5. The top left panelofFigure7showstheresultsforbanktellers. AsdiscussedinBessen(2015)andAutor (2015),theATMsubstitutesformuchoftheworkbanktellersoriginallydidandonemight expect the proliferation of ATMs to decrease bank teller employment. Instead, bank teller employmentwasflatfromabout1980. Thishasbeenattributedtochangesinregulationand thecostsavingsassociatedwithATMs,whichmadeopeningnewbranches(whichstillemploysometellers)moreappealing. Whatisstrikinginthechartistheimplicationthatbank tellers suffered a major negative occupation-specific shock. Basically, if bank tellers maintained their share of banking workforce they would have seen dramatic growth. This is completelyconsistentwithATMs(atechnologythatsubstitutesforlabor)havinganegative partialequilibriumimpactonbanktelleremployment. Whatthedecompositioncannotpick upisanygeneralequilibriumrelationshipbetweenoccupation-specificshocksandindustry growth. Insummary, thegraphcorrectlyshowsthattherewasatechnologicalsubstitution awayfrombanktellersintheproductionfunction,thoughthatsameshockandotherregulatoryfactorscombinedexpandedtheindustry. The second and third panels of Figure 7 show employment for bookkeepers and telephone operators. The literature on skill-biased technical change (e.g. Autor et al. (2003)) has argued that these types of routine cognitive occupations saw technological substitutioninrecentdecades. Thisisborneoutinthechart,wherethereisalargeimpliednegative occupation-specificshock. Thelowerrightpanelshowstheresultsfordataentrykeyers;this occupation initially grew quickly perhaps driven by computerization. But from the 1990s it suffered an occupation specific shock, possibly from the rise of the internet and the offshoringofthesejobs. Importantly,theimplicationisthattheindustriesthatemployeddata 33

800 600 400 sboj fo sdnasuohT bank tellers 3000 2500 2000 1500 1000 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth sboj fo sdnasuohT bookkeepers and accounting and auditing clerks 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth 500 400 300 200 100 sboj fo sdnasuohT telephone operators 600 550 500 450 400 350 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth sboj fo sdnasuohT data entry keyers 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth Source: ACS,Deming(2017),authors’calculations Figure7: Negativeoccupation-specificshocks entrykeyersmostlyremainedintheU.S.andonlyoutsourcedalimitedsetofoccupations. Takentogether,theexamplesinFigure7showthatthebetween-industrydecompositioncan correctlyidentifyoccupation-specificshocks. Figure 8 shows the analysis for examples of industry-level shocks. In the first panel we seesewingmachineoperatoremploymentfallingsteadily(astradeerodedthedomestictextile industry as a whole) and a smaller role for a negative occupation-specific shock. In the upper right panel we show petroleum, mining and geological engineers. This group saw employment first fall—as traditional oil production in the U.S. declined—and then reverse 34

700 600 500 400 300 200 sboj fo sdnasuohT textile sewing machine operators 55 50 45 40 35 30 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth sboj fo sdnasuohT petroleum, mining, and geological engineers 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth 350 300 250 200 150 sboj fo sdnasuohT pharmacists 1000 800 600 400 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth sboj fo sdnasuohT guards, watchmen, doorkeepers 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth Source: ACS,Deming(2017),authors’calculations Figure8: Examplesofindustryshocks 35

600 500 400 300 200 100 sboj fo sdnasuohT legal assistants, paralegals, legal support, etc 100 80 60 40 20 0 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth sboj fo sdnasuohT programmers of numerically controlled machine tools 1980 1990 2000 2010 2020 Observed Counterfactual based on industry growth Source: ACS,Deming(2017),authors’calculations Figure9: Positiveoccupation-specificshocks courseasshalegasandshaleoilbecameprofitable(seeDeckeretal.(2016)andDeckeretal. (2024) for more on shale oil). The counterfactual series attributes most of the employment dynamicstoindustryshocks,consistentwithchangesinoildemandandindustry-leveltechnologyshocksthatareneutralwithrespecttothetypesoflaboremployed. Thebottomrow ofFigure8showspharmacistsandguards,twootheroccupationsthathavenotexperienced largeoccupation-specificshocks. Finally, Figure 9 shows a couple examples of positive occupation-specific shocks. Since the 1980s legal assistants and paralegals have grown increasingly important in the legal domain. Similarly, computer numerically controlled (CNC) machine programmers have grown dramatically as the industry-based counterfactual would have predicted flat or declining employment. CNC machines are distinct from robots but their increased use is part of the continuing automation of manufacturing, especially in the motor vehicles and aerospacesectors. Theseexamplesillustratetheutilityofthebetween-industrydecomposition,anddemonstratesomecaseswherewecancrediblyidentifyotheroccupation-specificshocks. 36

8 Conclusion The growth of LLMs has sparked considerable debate regarding the potential to automate complextasks,potentiallyleadingtolastingimpactsonthelabormarket. Thispaperleverages occupation level data from O*NET and the CPS to measure the impact of ChatGPT’s releaseinNovember2022onemploymentincomputerprogramming,anoccupationheavily exposed to AI. We find robust evidence that coder employment growth fell after that release. After controlling for industry-level shocks we find that coder employment growth hasbeen3percentlowersincetheintroductionofChatGPT.Thismayreflectreallocationof tasks across occupations and we cannot control for all the relevant factors. Nonetheless, it suggeststhatAIishavingasignificantimpactonthisgroupofworkers. We believe it is important to control for industry-level shocks. Indeed, much of the debateoverAIandthelabormarkethascenteredonindustry-levelstoriessuchasinterest-rate sensitivity and post-Covid dynamics. Our counterfactual exercise shows that the relevant industries differentially substituted away from coder employment. This is evidence that codersspecificallyexperiencedashocknotsharedbytheircoworkersinotheroccupations. Ourhistoricalanalysisshowsmanycaseswheretheindustry-basedcounterfactualcorrectly identifiesthemixofindustryandoccupation-specificshocksaffectingemployment. There are many unanswered questions. If firms are reducing coder employment, is it because they see concrete productivity gains, or are they holding back on hiring anticipatinggains? Areremoteoroffshorecodingservicesprovidersmorevulnerable? Inthelonger run will the coder employment trend reverse, as new applications of (much cheaper) coding services develop? Are AI developments enabling the automation of other occupations (possiblywiththeinputofcoders)? TowhatextentareAIdevelopmentsaffectingaggregate productivityandlabordemand? Thecurrentpaper—suggestingaconcreteemploymenteffectonawell-specifiedgroupofworkers—isonlyafirststeptowardansweringtheseother questions. 37

References Acemoglu,Daron,“ThesimplemacroeconomicsofAI,”EconomicPolicy,2025,40(121),13– 58. andDavidAutor,“Skills,tasksandtechnologies: Implicationsforemploymentandearnings,”in“Handbookoflaboreconomics,”Vol.4,Elsevier,2011,pp.1043–1171. andJonasLoebbing,“AutomationandPolarization,”JournalofPoliticalEconomy,March 2026,134(3). ElectronicallypublishedJanuary22,2026. andPascualRestrepo,“Automationandnewtasks: Howtechnologydisplacesandreinstateslabor,”Journalofeconomicperspectives,2019,33(2),3–30. Ahn,HieJooandNickCarollo,“AIandLabor-MarketReallocation,”2026. Mimeo. Atkinson, Tyler and Shane Yamco, “Young workers’ employment drops in occupations withhighAIexposure,”January2026. Autor,DavidH.,“WhyAreThereStillSoManyJobs? TheHistoryandFutureofWorkplace Automation,”JournalofEconomicPerspectives,September2015,29(3),3–30. , Frank Levy, and Richard J. Murnane, “The Skill Content of Recent Technological Change: An Empirical Exploration*,” The Quarterly Journal of Economics, 11 2003, 118 (4), 1279–1333. Bai, Jushan and Pierre Perron, “Computation and analysis of multiple structural change models,”Journalofappliedeconometrics,2003,18(1),1–22. Baily,Martin,DavidByrne,AidanKane,andPaulSoto,“GenerativeAIattheCrossroads: LightBulb,Dynamo,orMicroscope?,”arXivpreprintarXiv:2505.14588,2025. Bessen,James,“ToilandTechnology,”Finance&Development,March2015,52(1),16–19. 38

Bick,Alexander,AdamBlandin,andDavidJDeming,“TheRapidAdoptionofGenerative AI,”WorkingPaper32966,NationalBureauofEconomicResearchSeptember2024. Bonney,Kathryn,CoryBreaux,CathyBuffington,EminDinlersoz,LuciaSFoster,Nathan Goldschlag, John C Haltiwanger, Zachary Kroff, and Keith Savage, “Tracking firm use of AI in real time: A snapshot from the Business Trends and Outlook Survey,” Technical Report,NationalBureauofEconomicResearch2024. Borusyak, Kirill, Peter Hull, and Xavier Jaravel, “A practical guide to shift-share instruments,”JournalofEconomicPerspectives,2025,39(1),181–204. Brynjolfsson,Erik,BharatChandar,andRuyuChen,“Canariesinthecoalmine? sixfacts abouttherecentemploymenteffectsofartificialintelligence,”TechnicalReport2025. , , and , “Canaries, Interest Rates, and Timing: More on the Recent Drivers of EmploymentChangesforYoungWorkers,”February2026. Onlinenews/insightarticle. Chandar,Bharat,“TrackingEmploymentChangesinAI-ExposedJobs,”June2025.Working paper. Chatterji, Aaron, Thomas Cunningham, David J Deming, Zoe Hitzig, Christopher Ong, CarlYanShan,andKevinWadman,“HowPeopleUseChatGPT,”WorkingPaper34255, NationalBureauofEconomicResearchSeptember2025. Coglianese, John M, Seth Murray, and Christopher J Nekarda, “Harmonized Population andLaborForceStatistics,”2025. Cottier, Ben, Tamay Besiroglu, and David Owen, “Challenges in Predicting AI Automation,”112023. EpochAIblog. Cowx,Mary,RebeccaLester,andMichelleLNessa,“Theconsequencesoflimitingthetax deductibilityofR&D,”2025. 39

Crane, Leland, Michael Green, and Paul Soto, “Measuring AI Uptake in the Workplace,” 2025. Davis, Steven J. and John Haltiwanger, “Gross Job Creation, Gross Job Destruction, and EmploymentReallocation,”TheQuarterlyJournalofEconomics,1992,107(3),819–863. Decker,RyanA.,AaronFlaaen,andMariaD.Tito,“UnravelingtheOilConundrum: Productivity Improvements and Cost Declines in the U.S. Shale Oil Industry,” FEDS Notes, BoardofGovernorsoftheFederalReserveSystemMarch2016. , MeaganMcCollum, andGregoryB.Upton, “BoomTownBusinessDynamics,” Journal ofHumanResources,2024,59(2),627–651. Decker,RyanandJohnHaltiwanger,“Hightechbusinessentryinthepandemicera,”2024. Deming, David J., “The Growing Importance of Social Skills in the Labor Market*,” The QuarterlyJournalofEconomics,062017,132(4),1593–1640. Dominski,JacobandYongSukLee,“AdvancingAICapabilitiesandEvolvingLaborOutcomes,”2025. Du, Wendi and Dongmei Li, “Real Effects of External Innovation: Evidence from the Tax CutsandJobsAct,”AvailableatSSRN5773322,2025. Eckhardt, Sarah and Nathan Goldschlag, “AI and Jobs: The Final Word (Until the Next One),”August2025. Onlineanalysis/report. Eisfeldt, Andrea L, Gregor Schubert, and Miao Ben Zhang, “Generative AI and firm values,”TechnicalReport,NationalBureauofEconomicResearch2023. Eloundou, Tyna, Sam Manning, Pamela Mishkin, and Daniel Rock, “GPTs are GPTs: LabormarketimpactpotentialofLLMs,”Science,2024,384(6702),1306–1308. Felten,EdwardW,ManavRaj,andRobertSeamans,“Occupationalheterogeneityinexposuretogenerativeai,”AvailableatSSRN4414065,2023. 40

Gimbel,Martha,MollyKinder,JoshuaKendall,andMaddieLee,“Evaluatingtheimpact ofAIonthelabormarket: Currentstateofaffairs,”TheBudgetLabatYaleUniversity,2025. Goldsmith-Pinkham, Paul, Isaac Sorkin, and Henry Swift, “Bartik instruments: What, when,why,andhow,”AmericanEconomicReview,2020,110(8),2586–2624. Handa, Kunal, Alex Tamkin, Miles McCain, Saffron Huang, Esin Durmus, Sarah Heck, Jared Mueller, Jerry Hong, Stuart Ritchie, Tim Belonax, Kevin K. Troy, Dario Amodei, Jared Kaplan, Jack Clark, and Deep Ganguli, “Which Economic Tasks are Performed withAI?EvidencefromMillionsofClaudeConversations,”2025. Humlum, Anders, “Robot Adoption and Labor Market Dynamics,” November 2021. Manuscript. andEmilieVestergaard,“LargeLanguageModels,SmallLaborMarketEffects,”TechnicalReport,NationalBureauofEconomicResearch2025. Iscenko,ZannaandFabienCurtoMillet,“TheTroublewithTiming: WhytheAIDisplacement Narrative is Premature,” Technical Report, Economic Innovation Group January 2026. Jaimovich, Nir and Henry E. Siu, “Job Polarization and Jobless Recoveries,” The Review of EconomicsandStatistics,032020,102(1),129–147. Katz, LawrenceF.andKevinM.Murphy, “ChangesinRelativeWages, 1963-1987: Supply andDemandFactors,”TheQuarterlyJournalofEconomics,1992,107(1),35–78. Korinek, Anton and Donghyun Suh, “Scenarios for the Transition to AGI,” Technical Report,NationalBureauofEconomicResearch2024. Lichtinger,GuyandSeyedMahdiHosseiniMaasoum,“GenerativeAIasSeniority-Biased Technological Change: Evidence from US Resume and Job Posting Data,” Available at SSRN,2025. 41

Massenkoff, Maxim and Peter McCrory, “Labor Market Impacts of AI: A New Measure andEarlyEvidence,”Report,AnthropicMarch2026. Soto,PaulE,“ResearchinCommotion: MeasuringAIResearchandDevelopmentthrough ConferenceCallTranscripts,”2025. Tomlinson, Kiran, Sonia Jaffe, Will Wang, Scott Counts, and Siddharth Suri, “Working withAI:MeasuringtheOccupationalImplicationsofGenerativeAI,”2025. Wiles, Emma and John J Horton, “Generative ai and labor market matching efficiency,” AvailableatSSRN5187344,2025. 42

A Imputation of the Computer Programming Skill For about 40 occupations we impute the skill variable on the basis of the occupation’s tasks.21 Inparticular,weconcatenatethetasksintoastringandgettheembeddingfromthe e5-large-v2 embedding model. Then we run a random forest to predict skill from the task embeddings using the sample of occupations with both task data and skill data. The rank correlationbetweenactualandpredictedskillisabout0.9onthissample,indicatingthatthe randomforestdoesagoodjobtranslatingtasksintoprogrammingskillrequirements. About 80 occupations are residual groups, like “machinists, all others.” These occupationscaptureremaindersanddon’thavetaskorskilldata. Weimputetheseusingtherelated occupations based on the O*NET/SOC classification structure, where the occupations that are related share the first few digits in their O*NET/SOC code. For the remainder occupationswetruncatetheendoftheoccupationcodeuntilwefindsomeoccupationswithdata andthenimputeusingtheaverage.22 A.1 SoftwareDevelopers Thetask-basedskillimputationassignssoftwaredevelopersacomputerprogrammingskill importance of 2.28, less than that of business intelligence analysts and slightly above computerusersupportspecialists. ThisisbecausetheO*NETtasksforsoftwaredevelopersonly mention coding in passing and instead emphasize planning, requirements gathering, writingreport,etc. Oursenseisthatactuallymanypeoplethatwouldcallthemselvessoftware developersorsoftwareengineers(anoccupationtitlethatismappedtosoftwaredevelopers, see below) are doing work O*NET would associate with computer programmers: writing codeandattendingmeetingstodesignandtroubleshootcode. IntheCPSdataoccupations 21Skills and importance ratings are based on surveys of incumbent workers in the occupation as well as assessmentsbyoccupationalexpertsandanalysts.Forfurtherinformationontheskillsratingprocedure,please seehttps://www.onetcenter.org/dl_files/AOSkills_Proc.pdf 22The exception is occupation 43-9199, “all other office workers.” If we impute using truncated codes it matchesthemwith“statisticalassistants”,ahighlyspecializedandtechnicaloccupation.Inthiscasewetruncate furthersotheremaindermatcheswithabroadersetofofficeworkers. 43

aresupposedtobeassignedbasedonthetasksdone,notwhattheworkercallsthemselves, butwesuspectmeasurementerrorcouldplayarolehere. InCensus-fieldedsurveysliketheCPSrespondentsareaskedfortheirjobtitleandtheir usual duties. On the basis of this information analysts assign the respondent’s occupation using the Census occupation codes (which can be crosswalked to SOC codes and O*NET occupations). Remarkably, Census provides a public-use microdata sample of written-in job titles and duties, as well as the Census occupation that was assigned.23 These data are from the 2018 American Communities Survey (ACS), but should be comparable to the CPS data. Table A1 lists the written-in job duties for people coded as software developers (upper panel) and computer programmers (lower panel). There are many more examples forsoftwaredevelopersbecauseitisamorecommonoccupationlabel. Softwaredevelopers frequently mention writing code, which suggests they may be fairly similar to computer programmers. On the other hand, they more frequently use terms like “develop software” which(iftheO*NETdescriptioniscorrect)couldmeanthingsotherthancoding. To examine this in more depth we train a random forest on embeddings of the selfreportedjobdutiestoclassifytheduty(againusinge5-large-v2forembeddings). Wemanuallylabelthedutiesforcomputerprogrammers,softwaredevelopers,andasampleofother occupations. We label three classes: computer programming, software development, and non-programming/non-development. Forthisexerciseweintentionallytreat“softwaredevelopment”asanambiguousclassdistinctfromcoding. Nearlyeverywrite-inforcomputer programming(fromTableA1)ishandlabeledasprogramming,whileonlyabout1/3ofthe softwaredevelopmentobservationsare. The random forest has a good fit on the hand labeled data and in the complete dataset the write-ins it classifies as programming are sensible. For each write-in observed we take themostlikelyclass asthelabelandcalculatethefraction ofanoccupation’swrite-insthat are classified as computer programming. Table A2 shows the occupations that are most 23Seehere. 44

programming-intensive according to this metric. Software developer is near the top of the ranking, only behind a few small occupations with very few observations and noisy estimates. Thisshowsthatsoftwaredevelopersareamongtheclosestoccupationstocomputer programmingintermsofintensityofcomputerprogramming,evenifwedon’tconsider“software development” and “developing software” to include coding. This makes us more confident that software developer duties are very close to those of programmers. With this in mind we override the task-based imputation and hard-code a 4 for the computer programming skillvariableforsoftwaredevelopers. SoftwareDeveloper WORKEDONRESPONSIVEDESIGNANDDEVELOPEDASINGLERESPONSIVEWEBSITE,USGOVERNMENT,DESIGNDATACENTERSFORSTATELO- CALGOVERNMENTANDHIGHEREDUCATIONINSTITUTIONS,C4ISRSYSTEMDESIGN,COMPUTERWORK,DESIGNANDDEVELOPSOFTWAREFOR DEFENSESYSTEMS,SYSTEMSENGINEERING,VALIDATINGPROGRAMREQUIREMENTS,SYSTEMARCHITECTURE,TECHNOLOGY,DEVELOPMET- RICSANDANALYTICSSOLUTIONSEGDASHBOARDSREPORTSETC,DEMOPRODUCTSASSISTTROUBLESHOOTINGFORPRODUCTTRIALSPRO- VIDECUSTOMIZATIONSFORTHECUSTOMER,TECHNICALSUPPORTFORHRSYSTEMSANDINTERFACES,AUTOMATEDREGRESSIONTESTING FORSOFTWAREDEVELOPMENT,BUILDEDUCATIONALWEBSOFTWARE,COMPUTERPROGRAMMER,CONFIGUREANDSUPPORTCLIENTSOFT- WAREENVIRONMENTS,CREATEWEBSITESANDAPPSONLINE,DESIGNANDDEVELOPSOFTWARE,DESIGNANDWRITECOMPUTERSOFTWARE, DESIGNARCHITECTWRITESOFTWAREANDLEADTEAM,DESIGNDEVELOPANDTESTSOFTWAREAPPLICATIONS,DESIGNSOFTWAREARCHI- TECTUREANDIMPLEMENTSERVICES,DESIGNINGCODINGTESTINGMAINTAININGCOMPUTERSOFTWARE,DEVELOPANDMAINTAINTECH- NOLOGYMONITORINGSOFTWARE,DEVELOPCOMPUTERPROGRAMSTOMANAGEWEBSERVICEPLATFORM,DEVELOPSOFTWARE,DEVELOP SOFTWARE,DEVELOPINGSOFTWARE,LEADGRAPHICSENGINEER,MAINTAINGOVERNMENTSATELLITEPROGRAMS,MAINTAINSYSTEMSAND SOFTWAREAPPLICATIONSASSOCIATEDWITHVEHICLERENTAL,PROGRAMMING,SOFTWAREENGINEERING,SOFTWAREENGINEERING,SOFT- WAREPROGRAMMINGINPL/SQLJAVAETC,SOFTWARESERVICE,TESTINGSOFTWARE,WORKONCOMPANYWEBSITE,WRITECOMPUTERCODE, WRITINGCODEDESIGNINGSOFTWARESYSTEMSREVIEWINGCODE,WRITINGSOFTWARE,ENGINEERSOFTWARE,FOLLOWBEHAVIORDRIVENDE- VELOPMENTMETHODOLOGYTODEVELOPSOFTWAREUSEJAVAJBEHAVECD/CICONCEPTS,DEVELOPESOFTWARE,PROGRAMMING,SOFTWARE DEVELOPMENT,CREATEANDMAINTAINSOFTWAREUSERINTERFACESREPORTSAUTOMATEDPROCESSES,DEVELOPSOFTWARE,DEVELOPSOFT- WARE,DEVELOPSOFTWAREFORCLIENTS,EVALUATEBUSINESSNEEDSTHENDESIGNANDDEVELOPSOFTWARESOLUTIONS,SOFTWAREDEVEL- OPMENT,SOFTWAREDEVELOPMENTANDTRAINNEWHIRES,VERIFYTRADINGAPPLICATION,WEBDEVELOPMENT,WRITECODE,WRITECODE, WRITECODEFORMOBILEAPPS,WRITETESTANDMAINTAINSOFTWARE,DEVELOPANDSUPPORTMARKETINGCAMPAIGNSANDREPORTING SOFTWARE,HELPTOBUILDANDINSTALLSOFTWAREPRODUCTFORCLIENT,DOCUMENTSOFTWAREREQUIREMENTSFORCLIENTSANDVAL- IDATEDELIVEREDPRODUCTS,DESIGNANDDEVELOPSOFTWAREDATABASESYSTEMS,DEVELOPSOFTWARE,IMPLEMENTATIONOFSOFTWARE TOPROVIDEELECTRONICCREDENTIALSTOLEARNERSANDSCHOOLS,CREATESOFTWARE,SOFTWAREONSATELLITE,TESTINGSOFTWARE,PRO- GRAMMINGENHANCEMENTS,WRITECOMPUTERCODE&MONITORCOMPUTERSERVICES,TRACKCOMPANYPROGRESSFORPROJECTS,HELP CUSTOMERSAGENCIESWITHTHEIRPAYROLLQUESTIONSANDPROJECTS,WRITEREQUIRMENTDOCUMENTSTHATTELLCOMPUTERPROGRAM- MERSWHATTOPROGRAM,ITSTRATEGYANDSOFTWAREIMPLEMENTATION,MAINTAININGSYSTEMSINTHECASINORESORT,INFRASTRUCTURE DESIGNIMPLEMENTANDMAINTAIN,SPECPROVISIONMAINTAINUNIXANDLINUXSERVERS,WRITEWEBAPPLICATIONS,SALES,ENGINEERIT SYSTEMS,DEVELOPINGIAMTOOLS,DEVELOPCODETHATALLOWSROBOTSTOMOVE,QUALIFICATIONOFPRODUCTS,CODESOFTWARE,DE- VELOPSOFTWAREFORNETWORKINGDEVICES,WRITESOFTWARE,CODING,WRITEFULLSTACKCODEANDPLANSOFTWARESTRATEGYFOR SERVICES-,DEVELOPSOFTWARES,CODING,DESIGNANDANALYZECOMPUTERPROGRAMS,ANALYZEPROBLEMSANDCREATECOMPUTERPRO- GRAMS,COMPUTERPROGRAMMING,PUBLICANDCONGRESSIONALAFAIRS,DEVELOPCLINICALAPPLICATIONSFORMASSSPECTROMETRY PRODUCTSANDMARKETTOAPPROPRIATECLIENTS,PLANANDBUILDSOFTWARE,CODEDEVELOPMENTFORINFORMATIONSYSTEMS,INFOR- MATIONDEVELOPING,SUPPORTANDENHANCESOFTWAREDEVELOPMENT,ANALYSIS ComputerProgrammer SAME,PROGRAMMER,MANAGESTUDENTSTUTORSANDWRITECURRICULUM, DESIGN AND IMPLEMENT SOFTWARE, DEVELOPING SOFTWARE/ONLINE FOR COMPANIES, PROGRAM, SUPPORT, IT, PROGRAMMER, PROGRAMMING, SOFT- WARE PROGRAMMER, COMPUTER PROGRAMMER, AUTOMATION PROGRAM- MING, COMPUTER PROGRAMMER, CODING AND BUSINESS ANALYTICS, COM- PUTERPROGRAMMING,COMPUTERPROGRAMMING,CREATEPROGRAMSAND REPORTS, DESIGN AND CONSTRUCT COMPUTER PROGRAMS, PROGRAMMING, WRITE AND MAINTAIN CLIENT SOFTWARE SYSTEM, DOING CODING FOR A CONSTRUCTIONCOMPANY TableA1: Self-reportedjobduties 45

Code Title FractionProgramming NumberofObs. 1010 Computerprogrammers 0.727 22 5920 Statisticalassistants 0.500 2 1240 Othermathematicalscienceoccupations 0.364 11 1031 Webdevelopers 0.333 6 1021 Softwaredevelopers 0.277 101 1800 Economists 0.167 6 7905 Computernumericallycontrolledtooloperatorsandprogrammers 0.091 11 1220 Operationsresearchanalysts 0.077 13 Note:TopCensusoccupationsbyfractionofACSdutywrite-insclassifiedas“computerprogramming”.“FractionProgramming”givensthefractionofthatoccupation’sdutywrite-insclassifiedasprogrammingbyarandomforestrunonthetextembeddings. NumberofObs. isthenumberofrespondentsinthatoccupationwith write-indatainthepublic-usefile. Source: ACS,authors’calculations TableA2: Fractionsofdutywrite-insclassifiedascomputerprogrammingbyoccupation Table A3 shows the most coding-intensive occupations in O*NET. The “Importance” variableistheprogrammingskillvariableweuse,whichisdesignedtocapturehowcritical having the skill is for the occupation. “Level” is a separate measure that captures the degree of computer programming sophistication needed for the occupation, it comoves very closely with Importance. O*NET explains the difference using the the example of lawyers and paralegals with respect to the speaking skill. Speaking is an important skill for both occupations, but level required is higher for lawyers, who often need to argue cases in formal settings. The ordering appears sensible, with the top ranks dominated by jobs that are obviouslycoding-intensive. Lowerdownintherankswestartseeingslightlylesscomputerfocusedoccupations,e.g. PhysicistsandBiostatisticians. Wechooseathresholdvalueof2.76todefineourprogramming-intensive,orcoder,occupationgroup. Thisislowenoughtoincludecomputerscientistsandwebdevelopers(clearly coding-heavy jobs) but excludes, e.g., automotive engineers and information system managers. 46

Code Title Importance Level ImputationType ComputerProgrammers 15-1251.00 4.750 4.880 - WebDevelopers 15-1254.00 4.120 4.250 - ComputerNumericallyControlledToolProgrammers 51-9162.00 4.120 4.120 - VideoGameDesigners 15-1255.01 4.000 3.880 - SoftwareDevelopers 15-1252.00 4.000 4.000 C ComputerNetworkArchitects 15-1241.00 3.880 3.620 - DataWarehousingSpecialists 15-1243.01 3.750 3.750 - SoftwareQualityAssuranceAnalystsandTesters 15-1253.00 3.620 3.750 - NetworkandComputerSystemsAdministrators 15-1244.00 3.620 3.880 - ComputerandInformationResearchScientists 15-1221.00 3.620 4.500 - ComputerSystemsEngineers/Architects 15-1299.08 3.500 4.250 - Biostatisticians 15-2041.01 3.380 4.000 - DatabaseArchitects 15-1243.00 3.380 4.120 - Physicists 19-2012.00 3.380 3.880 - DatabaseAdministrators 15-1242.00 3.380 3.880 - ClinicalDataManagers 15-2051.02 3.250 3.620 - ComputerSystemsAnalysts 15-1211.00 3.250 4.000 - BioengineersandBiomedicalEngineers 17-2031.00 3.120 3.380 - RoboticsEngineers 17-2199.08 3.120 3.880 - BioinformaticsTechnicians 15-2099.01 3.120 3.000 - WebAdministrators 15-1299.01 3.120 2.880 - MathematicalScienceOccupations,AllOther 15-2099.00 3.120 3.000 B Statisticians 15-2041.00 3.000 3.120 - ComputerScienceTeachers,Postsecondary 25-1021.00 3.000 3.120 - StatisticalAssistants 43-9111.00 3.000 3.500 - PhysicalScientists,AllOther 19-2099.00 2.880 2.750 B RemoteSensingScientistsandTechnologists 19-2099.01 2.880 2.750 - RoboticsTechnicians 17-3024.01 2.880 2.750 - GeographicInformationSystemsTechnologistsandTechnicians 15-1299.02 2.880 3.120 - HealthInformaticsSpecialists 15-1211.01 2.880 3.120 - Industrial-OrganizationalPsychologists 19-3032.00 2.750 2.880 - AutomotiveEngineers 17-2141.02 2.750 2.880 - SearchMarketingStrategists 13-1161.01 2.750 2.120 - Biologists 19-1029.04 2.750 3.620 - ComputerandInformationSystemsManagers 11-3021.00 2.750 2.620 - ElectronicsEngineers,ExceptComputer 17-2072.00 2.750 2.880 - RemoteSensingTechnicians 19-4099.03 2.750 3.250 - LogisticsEngineers 13-1081.01 2.620 2.250 - DesktopPublishers 43-9031.00 2.620 2.620 - Note: Top O*NET occupations for coding importance sorted in descending order. Horizontal line shows our thresholdforanoccupationbeingcoding-intensive.Imputationtype“A”meansrandomforest-basedusingtask embeddings(notshown),imputationtype“B”meansimputationusingthetruncatedSOCcodeoccupational groupings,imputationtype“C”meansthesoftwaredevelopershardcodebasedonwritteninjobduties. Source: CPS,O*NET,authors’calculations TableA3: Coding-intensiveONETOccupations 47

B Link to CPS BLS provides crosswalks between O*NET and the SOC codes used in the “National Employment Matrix” (NEM), the industry-occupation employment totals they publish from the OES. The NEM codes are a modification of the SOC codes. Crosswalking from O*NET to NEM only results in the loss of 19 O*NET occupations, mostly military-related. When multipleO*NEToccupationsmatchtoasingleNEMoccupationwetakethesimpleaverage oftheskillvariable. Weendupwith832usableNEMoccupations. BLS also has a crosswalk between NEM codes and CPS occupation codes. Again we averagetheskillvariablewhenmultipleNEMoccupationsmatchoneCPSoccupation. We endupwith525useableCPSoccupations. The O*NET and CPS codes we match to are based on the 2018 SOC, which the CPS data use from 2020. To get consistency before 2020 we need to use the standardized 2010 occupation codes that IPUMS has developed, occ2010. All workers in all CPS data post- 2010haveocc2010codesonaconsistentbasis. Inthepost-2020sample,employedindividualshavebothocc2010and2018-vintageoccupation codes. We average to get a weighted crosswalk between the 2018 vintage and occ2010, which lets us assign programming skill to occ2010 codes as the employmentweightedaverageoftheoccupationsmatchingtothempost-2020. There are a small number of occ2010 codes that were in use prior to 2020 but not afterwards. These codes don’t appear in the weighted crosswalk above but they account for a (small)portionofpre-2020employment. Togetaskillvariableforabout50occupationsfittingthisdescriptionweusetheofficial2010-2018Censusoccupationcodebridge,andseeif wecanmatchtheocc2010valuestotheNEMcodes. Afterthissteptherearestillahandful ofoccupationswecan’tpopulate,wedroptheseholdouts. 48

C Industries CensusIndustry NAICStitle NAICScodes NAICSpercentofgroup Computersystemsdesignandrelatedservices Computersystemsdesignandrelatedservices 5415 100 Softwarepublishers Softwarepublishers 5112 100 Dataprocessing,hosting,andrelatedservices Dataprocessing,hosting,andrelatedservices 5182 100 Scientificresearchanddevelopmentservices Scientificresearchanddevelopmentservices 5417 100 Computerandperipheralequipmentmanufacturing Computerandperipheralequipmentmanufacturing 3341 100 Othertelecommunicationsservices Telecommunications,exceptwiredtelecommunicationscarriers 517exc.517311 100 Newspaperpublishers Newspaperpublishers 51111 13.79 Newspaperpublishers Periodical,book,anddirectorypublishers 5111exc.51111 19.34 Newspaperpublishers Broadcasting(exceptinternet) 515 34.4 Newspaperpublishers Internetpublishingandbroadcastingandwebsearchportals 51913 28.47 Newspaperpublishers Otherinformationservices,exceptlibrariesandarchives,andinternetp... 5191exc.51912,51913 4.01 Pharmaceuticalandmedicinemanufacturing Pharmaceuticalandmedicinemanufacturing 3254 100 Electricandgas,andothercombinations Electricandgas,andothercombinations Pts.2211,2212 100 Othergeneralgovernmentandsupport Othergeneralgovernmentandsupport 92119 100 Nationalsecurityandinternationalaffairs Nationalsecurityandinternationalaffairs 928 100 Aircraftandpartsmanufacturing Aircraftandpartsmanufacturing 336411,336412,336413 100 TableA4: Industriessortedbyintensityofcoders,withNAICScodes D Robustness Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend 0.0359 -0.233 -0.0529 -1.847*** -2.254*** -1.578*** 1.777*** 1.632** 1.455** (0.356) (0.470) (0.478) (0.376) (0.307) (0.448) (0.495) (0.658) (0.580) Post-GPT*Trend -2.827*** -2.557*** -1.253* -1.055 -0.648 -3.246** -4.057*** -3.911*** 0.0741 (0.503) (0.607) (0.747) (0.903) (0.867) (1.438) (1.054) (1.218) (0.881) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A5: Regressions Controlling for Counterfactual, 25 percent stricter definition of “coder” 49

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -1.800*** -1.972*** -1.661*** -2.722*** -2.870*** -2.779*** 0.317 0.0968 0.850* (0.151) (0.129) (0.200) (0.230) (0.229) (0.324) (0.390) (0.457) (0.454) Post-GPT*Trend -1.897 -1.725 -5.453*** -3.017** -2.869** -7.463*** 0.121 0.341 -1.709 (1.148) (1.179) (1.446) (1.305) (1.324) (0.912) (1.201) (1.304) (2.985) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A6: Regressions Controlling for Counterfactual, 25 percent looser definition of “coder” Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -1.655*** -1.971*** -1.574*** -2.942*** -3.408*** -2.536*** -0.170 -0.325 -0.340 (0.260) (0.282) (0.353) (0.423) (0.298) (0.459) (0.271) (0.370) (0.343) Post-GPT*Trend -1.240* -0.924 1.670** 0.381 0.846 -0.605 -2.714* -2.559* 3.395*** (0.652) (0.757) (0.646) (0.773) (0.705) (1.438) (1.374) (1.534) (0.687) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A7: Regressions Controlling for Counterfactual, 50 percent stricter definition of “coder” 50

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -2.162*** -2.288*** -2.040*** -2.976*** -3.088*** -2.936*** 0.0797 -0.0737 0.392 (0.125) (0.156) (0.178) (0.182) (0.210) (0.246) (0.277) (0.329) (0.316) Post-GPT*Trend -1.795* -1.669 -5.116*** -3.154*** -3.041*** -6.996*** 1.318 1.471 -0.738 (1.041) (1.085) (1.372) (1.110) (1.143) (0.961) (1.137) (1.227) (2.821) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A8: Regressions Controlling for Counterfactual, 50 percent looser definition of “coder” Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -1.360*** -1.480*** -1.296** -3.027*** -3.368*** -2.805*** 0.00567 0.00994 0.0280 (0.392) (0.414) (0.525) (0.756) (0.779) (0.943) (0.233) (0.256) (0.349) Post-GPT*Trend -4.231*** -4.111*** -6.952*** -4.622*** -4.281*** -7.750*** -4.033*** -4.037*** -6.644*** (0.866) (0.896) (1.003) (1.148) (1.135) (1.071) (0.867) (0.930) (1.353) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. TableA9: RegressionsControllingforCounterfactual,excludingsoftwaredevelopers 51

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend 1.792*** 1.666*** 1.794*** 0.128 -0.107 0.316 3.130*** 3.106*** 2.987*** (0.202) (0.200) (0.280) (0.375) (0.347) (0.446) (0.217) (0.237) (0.267) Post-GPT*Trend -2.246*** -2.120*** -3.014*** -2.185*** -1.951*** -3.646*** -2.339*** -2.315*** -2.665*** (0.437) (0.421) (0.683) (0.650) (0.615) (0.688) (0.398) (0.410) (0.826) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. TableA10: RegressionsControllingforCounterfactual,usingallworkerstocomputecounterfactualindustrygrowth Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -0.962*** -1.184*** -0.849* -2.359*** -2.797*** -1.967*** 0.616*** 0.579*** 0.533* (0.331) (0.300) (0.437) (0.642) (0.568) (0.743) (0.206) (0.211) (0.287) Post-GPT*Trend -3.211*** -2.989*** -4.405*** -3.590*** -3.152*** -6.000*** -3.002*** -2.965*** -3.382*** (0.632) (0.594) (0.845) (1.080) (1.006) (1.172) (0.410) (0.416) (0.811) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. TableA11: RegressionsControllingforCounterfactual,usingemploymentsharesinsteadof levels 52

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -1.325*** -1.542*** -1.202*** -2.395*** -2.815*** -2.016*** -0.0149 -0.0569 -0.0667 (0.344) (0.315) (0.456) (0.635) (0.573) (0.738) (0.208) (0.202) (0.312) Post-GPT*Trend -3.683*** -3.467*** -4.673*** -3.696*** -3.276*** -5.848*** -3.723*** -3.680*** -3.925*** (0.605) (0.568) (0.794) (1.039) (0.974) (1.183) (0.374) (0.373) (0.686) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A12: Regressions Controlling for Counterfactual, monthly industry respondent thresholdofzero Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend 0.0200 -0.181 0.0181 -1.429** -1.817*** -1.226* 1.665*** 1.627*** 1.522*** (0.296) (0.301) (0.397) (0.547) (0.515) (0.665) (0.234) (0.255) (0.291) Post-GPT*Trend -3.130*** -2.929*** -4.236*** -3.465*** -3.077*** -5.352*** -2.952*** -2.915*** -3.540*** (0.586) (0.554) (0.809) (0.959) (0.888) (1.055) (0.429) (0.440) (0.827) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A13: Regressions Controlling for Counterfactual, monthly industry respondent thresholdof50 53

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -1.800*** -1.999*** -1.662*** -3.365*** -3.778*** -2.891*** -0.156 -0.174 -0.256 (0.346) (0.333) (0.453) (0.681) (0.631) (0.752) (0.203) (0.205) (0.305) Post-GPT*Trend -2.306*** -2.108*** -4.136*** -1.831 -1.418 -5.475*** -2.564*** -2.546*** -2.644*** (0.762) (0.752) (0.753) (1.237) (1.201) (1.169) (0.427) (0.439) (0.796) Observations 120 120 120 120 120 120 120 120 120 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A14: Regressions Controlling for Counterfactual, sample ending in 2024 with pre- NAICS2022codes Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend 0.425* 0.341* 0.491 -0.644 -0.949 0.120 0.616*** 0.579*** 0.533* (0.216) (0.185) (0.309) (0.805) (0.700) (0.802) (0.206) (0.211) (0.287) Post-GPT*Trend -4.301*** -4.216*** -5.299*** -9.133*** -8.828*** -13.09*** -3.016*** -2.978*** -3.410*** (0.583) (0.580) (0.805) (1.773) (1.783) (1.430) (0.414) (0.420) (0.818) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A15: Regressions Controlling for Counterfactual, dropping NAICS 5415 (computer systemsdesignandrelatedservices) 54

Allindustries Codingintensiveindustries Non-codingintensiveindustries (1) (2) (3) (4) (5) (6) (7) (8) (9) Trend -0.922*** -1.139*** -0.833* -2.310*** -2.743*** -1.957*** 0.645*** 0.613*** 0.548* (0.318) (0.293) (0.423) (0.619) (0.550) (0.726) (0.209) (0.217) (0.283) Post-GPT*Trend -3.415*** -3.197*** -4.894*** -3.769*** -3.336*** -6.504*** -3.186*** -3.153*** -3.797*** (0.657) (0.623) (0.831) (1.101) (1.030) (1.145) (0.434) (0.442) (0.809) Observations 133 133 133 133 133 133 133 133 133 Sample All NoCovid No2022/3 All NoCovid No2022/3 All NoCovid No2022/3 Standarderrorsinparentheses *p<.10,**p<.05,***p<.01 Note:Coefficientsareannualizedlogpoints,standarderrorsareNewey-West.Dependentvariableisdifference between monthly (log) coder employment and the counterfactual (log) employment series. Post-GPT period startsinNovember2022. Table A16: Regressions Controlling for Counterfactual, using smoothed Coglianese et al. (2025)weights 55

7000 6000 5000 4000 3000 sboj fo sdnasuohT All Industries Coders Counterfactual 2014m1 2016m1 2018m1 2020m1 2022m1 2024m1 2026m1 Note:EmploymentlevelofcoderoccupationsbasedonO*NETprogrammingskill.Verticallinesshowstructural breakdatesestimatedviaBai-Perrontests.ThedashedlineshowsChatGPTreleasedate(November2022). Source: O*NET,CPS,authors’calculations FigureA1: Bai-PerronBreakpoints 56