《普林斯頓大學：ChatGPT 類大語言模型將如何影響職業和行業（英文版）（12頁）.pdf》由會員分享，可在線閱讀，更多相關《普林斯頓大學：ChatGPT 類大語言模型將如何影響職業和行業（英文版）（12頁）.pdf（12頁珍藏版）》請在三個皮匠報告上搜索。

1、1 How will Language Modelers like ChatGPT Affect Occupations and Industries?Ed Felten(Princeton)Manav Raj(University of Pennsylvania)Robert Seamans(New York University)1 March 2023 Abstract:Recent dramatic increases in AI language modeling capabilities has led to many questions about the effect of t

2、hese technologies on the economy.In this paper we present a methodology to systematically assess the extent to which occupations,industries and geographies are exposed to advances in AI language modeling capabilities.We find that the top occupations exposed to language modeling include telemarketers

3、 and a variety of post-secondary teachers such as English language and literature,foreign language and literature,and history teachers.We find the top industries exposed to advances in language modeling are legal services and securities,commodities,and investments.Keywords:artificial intelligence,Ch

4、atGPT,language modeling,occupation,technology 2 1.Introduction Artificial Intelligence(AI)will likely affect the economy in many ways,potentially boosting economic growth and changing the way people work and play.The effect of AI on work will likely be multi-faceted.In some cases,AI may substitute f

5、or work previously done by humans,and in other cases AI may complement work done by humans.The effect on work will likely also vary across industries.Recent research by Goldfarb et al(2020)document that adoption of AI is relatively high in some industries such as information technology and finance,b

6、ut low in others such as health care and construction.Moreover,trying to understand how AI will affect work is like trying to hit a moving target because the capabilities of AI are still advancing.A prominent example of how AI capabilities continue to advance are the recent improvements in AI langua

7、ge modeling.In particular,ChatGPT,a language modeler released by Open AI in late 2022,has garnered a huge amount of attention and controversy.Some worry about the negative effects of tools like ChatGPT on jobs,as in the New York Post article headlined“ChatGPT could make these jobs obsolete:The wolf

8、is at the door.”1 Others see practical and commercial promise from language modeling.For example,Microsoft announced a$10 billion partnership with Open AI and has linked ChatGPT with its Bing search engine.2 Google felt compelled to demonstrate its own language modeler,Bard,but mistakes during the d

9、emonstration led Googles stock price to drop 7%.3 ChatGPT has been banned by J.P.Morgan.4 However,at present,most of this is speculation.In order to better understand how language modelers such as ChatGPT will affect occupations,industries and geographies,we use a methodology developed by Felten et

10、al(2018,2021).Felten et al created the AI Occupational Exposure(AIOE)measure and used this measure to identify which occupations,industries and geographies are most exposed to AI.In this paper,we describe how the AIOE approach can be adapted to account for the recent advancement of language modeling

11、.1 https:/ https:/ 3 https:/ 4 https:/ We find that the top occupations affected include telemarketers and a variety of post-secondary teachers such as English language and literature,foreign language and literature,and history teachers.We also find the top industries exposed to advances in language

12、 modeling are legal services and securities,commodities,and investments.This article contributes to several literatures.First,by providing a systematic examination of the effect of language modeling across occupations,industries and geographies,it contributes to a nascent literature on the effects o

13、f ChatGPT and other language modelers on the economy(e.g.Agarwal et al.,2022;Zarifhonarvar,2023).More generally,the article builds on a broader set of literature studying the effect of AI on the economy(Furman and Seamans,2019;Goldfarb et al.,2019).Second,the article builds on and extends a set of p

14、apers that provide systematic methodologies for studying how AI affects occupations(e.g.,Brynjolfsson et al,2018;Frey&Osborne,2017;Tolan et al.,2021;Webb,2020).The article specifically builds off and extends the methodology described in Felten et al.(2018,2021).In so doing,the article demonstrates t

15、he flexibility of the original Felten et al methodology;it can be adjusted dynamically to assess the impact of changes in AI capabilities.Finally,the article adds to a large literature on the effect of automating technologies on labor(e.g.,Acemoglu et al.,2022;Autor,2015;Frank et al.,2019;Genz et al

16、.,2021).The article proceeds as follows.Section 2 describes the AI Occupational Exposure(AIOE)measure developed by Felten et al(2018,2021).Section 3 extends the AIOE to account for recent advances in language modeling.Section 4 provides results,including listing the top 20 most affected occupations

17、and industries.Section 5 concludes.2.AI Occupational Exposure Methodology According to Felten et al(2021),the AI Occupational Exposure(AIOE)is a measure of each occupations“exposure”to AI.The term“exposure”is used so as to be agnostic as to the effects of AI on the occupation,which could involve sub

18、stitution or augmentation depending on various factors associated with the occupation itself.The AIOE measure was constructed by linking 10 AI applications(abstract strategy games,real-time video games,image recognition,visual question answering,image generation,reading comprehension,language modeli

19、ng,translation,speech recognition,and instrumental track 4 recognition)to 52 human abilities(e.g.,oral comprehension,oral expression,inductive reasoning,arm-hand steadiness,etc)using a crowd-sourced matrix that indicates the level of relatedness between each AI application and human ability.Data on

20、the AI applications come from the Electronic Frontier Foundation(EFF)which collects and maintains statistics about the progress of AI across multiple applications.Data on human abilities comes from the Occupational Information Network(O*NET)database developed by the United States Department of Labor

21、.O*NET uses these 52 human abilities to describe the occupational makeup of each of 800+occupations that it tracks.Each of 800+occupations can be thought of as a weighted combination of the 52 human abilities.O*NET uses two sets of weights:prevalence and importance.Once the 10 AI categories and 52 h

22、uman abilities are linked through the matrix,the AIOE can then be calculated for each occupation.To do this,first we calculate an ability-level exposure as follows:=10=1 (1)Where i indexes the AI application and j indexes the occupational ability.The ability-level exposure,A,is calculated as the sum

23、 of the 10 application-ability relatedness scores,x,as constructed using the matrix of crowd-sourced survey data.We then calculate the AIOE for each occupation k as follows:=52=152=1 (2)In this equation,i indexes the AI application,j indexes the occupational ability,and k indexes the occupation.Aij

24、represents the ability-level exposure score.We weight the ability-level AI exposure by the abilitys prevalence(Ljk)and importance(Ijk)within each occupation as measured by O*NET by multiplying the ability-level AI exposure by the prevalence and importance scores for that ability within each occupati

25、on,scaled so that they are equally weighted.Felten et al(2021)explain the construction of the AIOE scores in more detail,describe how they can be weighted at the industry level to construct an AI Industry Exposure score,or weighted at the geographic level to construct an AI Geographic Exposure score

26、.They also provide results 5 from a number of validation exercises and describe a number of ways in which the scores can be used by scholars and practitioners.5 3.Language Modeling AI Occupational Exposure The original AIOE described in Felten et al(2021)explicitly weighted each of the AI applicatio

27、ns the same.In order to update the AI Occupational Exposure score to account for advances in Language Modeling we modify equation(1)as follows.=10=1 (3)Where i indexes the AI application and j indexes the occupational ability.The ability-level exposure,A,is calculated as the weighted sum of the 10 a

28、pplication-ability relatedness scores,x,as constructed using the matrix of crowd-sourced survey data.is the weight placed on each application i.The weights used in Felten et al(2021)set equal to 1 for each application i.Next,we set equal to 0 for every AI application except for language modeling,whi

29、ch retains a weight of 1.This then constructs an ability-level exposure measure that only“counts”the value of abilities that are related to language modeling.We then proceed to calculate the for each occupation k using this new“language modeling”weighted.The resulting therefore captures the extent t

30、o which each occupation is exposed to advances in language modeling due to AI.A complete list of the occupations and their resulting AIOE language modeling score are listed in an appendix.The resulting scores are highly correlated with the original AIOE scores(correlation coefficient:0.979).This can

31、 be seen in Figure 1 which plots the original AIOE score and the new language modeling adjusted AIOE score for each occupation.4.Results 5 The Felten et al(2021)paper is open access and available here:https:/ The data and code used to create the AIOE scores described in Felten et al(2021)is availabl

32、e on GitHub:https:/ 6 In this section we present and briefly discuss tables of“top 20”occupations and industries exposed to language modeling.4.1.Top 20 Occupations Exposed to Language Modeling Table 1 provides the list of top 20 occupations exposed to AI based on the original Felten et al(2021)AI O

33、ccupational Exposure(AIOE)measure as well as the top 20 occupations exposed to AI enabled advances in language modeling capabilities.Some occupations occur in both lists,including“clinical,counseling,and school psychologists”,and“history teachers,postsecondary”.Notably,the language modeling list inc

34、ludes more education-related occupations,indicating that occupations in the field of education are likely to be relatively more impacted by advances in language modeling than other occupations.This accords well with the recent spate of articles around how ChatGPT and other language modeling tools af

35、fect the way teachers assign work and detect cheating or could use language modeling tools to develop teaching materials.Also of interest,the top occupation in the language modeling list is“telemarketer.”One might imagine that human telemarketers could benefit from language modeling being used to au

36、gment their work.For example,customer responses can be fed into a language modeling engine in real time and relevant,customer-specific prompts quickly fed to the telemarketer.Or,one might imagine that human telemarketers are substituted with language modeling enabled bots.The potential for language

37、modeling to augment or substitute for human telemarketers work highlights one aspect of the AIOE measure:it measures“exposure”to AI,but whether that exposure leads to augmentation or substitution will depend on specifics of any given occupation.4.2.Top 20 Industries Exposed to Language Modeling Tabl

38、e 2 provides the list of 20 industries most exposed to AI based on the original Felten et al.(2021)AI Industry Exposure(AIIE)measure as well as the top 20 industries exposed to AI enabled advances in language modeling capabilities.7 As before,we see some similarities in the industries categorized as

39、 most exposed to AI based on the original AIOE as well as the version that focuses on advances in language modeling capabilities.For example,“Securities,Commodity Contracts,and Other Financial Investments and Related Activities”is categorized as the most exposed industry using the original AIOE and

40、is the second most exposed industry using the language modeling-focused version of the AIOE.Legal services,insurance and employee benefit funds,and agencies,brokerages,and other insurance related activities are among the top five most exposed industries across both lists.However,some differences eme

41、rge.One salient difference is that the language modeling-focused AIOE suggests a higher exposure to advances in AI within higher education and higher education-adjacent industries.Junior colleges,grantmaking and giving services,and business schools and computer and management training all appear wit

42、hin the top twenty exposed industries.5.Conclusion In this paper we present a methodology to systematically assess the extent to which occupations and industries are exposed to advances in AI language modeling capabilities.This methodology relies on the approach described in Felten et al(2021)but ad

43、apts it to account for recent advances in language modeling.We find that the top occupations exposed to language modeling include telemarketers and a variety of post-secondary teachers such as English language and literature,foreign language and literature,and history teachers.We also find the top i

44、ndustries exposed to advances in language modeling are legal services and securities,commodities,and investments.At a broad level,this paper adds to a growing literature studying the effects of AI on labor and work.More specifically,the paper provides a systematic approach for understanding how Chat

45、GPT and other language modelers will affect occupations,industries and geographies.We believe these results will be useful for other scholars as well as practitioners and policymakers.8 References Acemoglu,D.,Autor,D.,Hazell,J.,&Restrepo,P.(2022).Artificial intelligence and jobs:Evidence from online

46、 vacancies.Journal of Labor Economics,40(S1),S293-S340.Agrawal,A.,Gans,J.,Goldfarb,A.2022.“ChatGPT and How AI Disrupts Industries”Harvard Business Review.Available:https:/hbr.org/2022/12/chatgpt-and-how-ai-disrupts-industries Autor,D.(2015).Why Are There Still So Many Jobs?The History and Future of

47、Workplace Automation.Journal of Economic Perspectives,29(3),330.Brynjolfsson,E.,Mitchell,T.,&Rock,D.(2018).What can machines learn,and what does it mean for occupations and the economy?AEA Papers and Proceedings,108,4347.https:/doi.org/10.1257/pandp.20181019 Felten,E.W.,Raj,M.,&Seamans,R.(2018).A me

48、thod to link advances in artificial intelligence to occupational abilities.In AEA Papers and Proceedings(Vol.108,pp.54-57).Felten,E.,Raj,M.,&Seamans,R.(2021).Occupational,industry,and geographic exposure to artificial intelligence:A novel dataset and its potential uses.Strategic Management Journal,4

49、2(12),2195-2217.Frank,M.R.,Autor,D.,Bessen,J.E.,Brynjolfsson,E.,Cebrian,M.,Deming,D.J.,.&Rahwan,I.(2019).Toward understanding the impact of artificial intelligence on labor.Proceedings of the National Academy of Sciences,116(14),6531-6539.Frey,C.B.,&Osborne,M.A.(2017).The future of employment:How su

50、sceptible are jobs to computerisation?Technological Forecasting and Social Change,114,254280.https:/doi.org/10.1016/j.techfore.2016.08.019 Furman,J.,&Seamans,R.(2019).AI and the Economy.Innovation policy and the economy,19(1),161-191.9 Genz,S.,Gregory,T.,Janser,M.,Lehmer,F.,&Matthes,B.(2021).How do

51、workers adjust when firms adopt new technologies?.ZEW-Centre for European Economic Research Discussion Paper,(21-073).Goldfarb,A.,Gans,J.,&Agrawal,A.(2019).The economics of artificial intelligence:An agenda.Chicago,IL:University of Chicago Press.Goldfarb,A.,Taska,B.Teodoridis,F.(2020).“Artificial In

52、telligence in Healthcare?Evidence from Online Job Postings”,AEA Papers and Proceedings,110(5):400-404 Tolan,S.,Pesole,A.,Martnez-Plumed,F.,Fernndez-Macas,E.,Hernndez-Orallo,J.,&Gmez,E.(2021).Measuring the occupational impact of AI:tasks,cognitive abilities and AI benchmarks.Journal of Artificial Int

53、elligence Research,71,191-236.Webb,M.(2020).The impact of artificial intelligence on the labor market.Stanford University working paper.Zarifhonarvar,A.2023.“Economics of ChatGPT:A Labor Market View on the Occupational Impact of Artificial Intelligence”Indiana University working paper.Available:http

54、:/dx.doi.org/10.2139/ssrn.4350925 10 Figure 1:Comparison between Original AIOE and Language Modeling Adjusted AIOE Notes:This figure plots the original AIOE score(x-axis)and the new language modeling adjusted AIOE score(y-axis)for each occupation.11 Table 1:Top 20 Occupations Exposed to AI,Original

55、and with Language Modeling Adjustment Notes:This table lists the top 20 occupations most exposed to AI from the original AIOE(Felten et al.,2021)and the top 20 occupations most exposed to language modeling.12 Table 2:Top 20 Industries Exposed to AI,Original and with Language Modeling Adjustment Notes:This table lists the top 20 industries most exposed to AI from the original AIOE(Felten et al.,2021)and the top 20 industries most exposed to language modeling.