《世界經濟論壇：2023未來工作白皮書-大型語言模型對就業的影響（英文版）（34頁）.pdf》由會員分享，可在線閱讀，更多相關《世界經濟論壇：2023未來工作白皮書-大型語言模型對就業的影響（英文版）（34頁）.pdf（34頁珍藏版）》請在三個皮匠報告上搜索。

1、Jobs of Tomorrow:Large Language Models and JobsW H I T E P A P E RS E P T E M B E R 2 0 2 3In collaboration with Accenture Images:Getty Images 2023 World Economic Forum.All rights reserved.No part of this publication may be reproduced or transmitted in any form or by any means,including photocopying

2、 and recording,or by any information storage and retrieval system.Disclaimer This document is published by the World Economic Forum as a contribution to a project,insight area or interaction.The findings,interpretations and conclusions expressed herein are a result of a collaborative process facilit

3、ated and endorsed by the World Economic Forum but whose results do not necessarily represent the views of the World Economic Forum,nor the entirety of its Members,Partners or other stakeholders.ContentsForeword 3Executive summary 4Introduction:How will large language models impact 5 the jobs of tomo

4、rrow?1 Identifying exposure potential of tasks and jobs 71.1 Exposed tasks 71.2 Detailed examples of exposed jobs 91.3 Analysis by occupation 101.4 Analysis by industry 141.5 Analysis by function 162 LLMs and the growth and decline of jobs and tasks 172.1 Expected growth and decline of tasks 172.2 E

5、xpected growth and decline of jobs 17Conclusion:Ensuring that large language models work for workers 19Appendices 20A1 Exposure potential by industry groups 20A2 Exposure potential by function groups 28A3 Methodology 30Contributors 32Endnotes 33Jobs of Tomorrow:Large Language Models and Jobs2Forewor

6、dGenerative AI and,in particular,large language models(LLMs),underpinned by advancements in machine learning and natural language processing,represent a paradigm shift in how we interact with information and,by extension,how we work.These technologies can create original content,generate insights fr

7、om large amounts of data,translate languages with near-human accuracy,and potentially even make complex decisions.The versatility and efficiency of these technologies could have profound implications for jobs and the future of work.While the application of LLMs could lead to significant productivity

8、 gains and the creation of new types of jobs,there is also a risk that they could displace existing roles,exacerbating socioeconomic disparities and creating a sense of job insecurity among the global workforce.As such,integrating AI into our workplaces is a balancing act between seizing opportuniti

9、es and managing potential disruptions.Generative artificial intelligence(AI)encompasses a broad set of technologies that can perform a variety of tasks.As a result,the public debate on its potential impact on workers is often polarized and uncertain across timeframes.In this report,we focus on large

10、 language models and the activities they can perform.This paper takes a structured approach to understanding the direct impact of LLMs on specific jobs.This analysis will enable stakeholders business leaders,policy-makers,workers and the broader public to make more informed decisions regarding skill

11、ing,workforce planning and other strategic investments.Generative AI will reshape industries and business in profound ways through new operating models,and new products and services.Yet,by proactively understanding and addressing the direct disruptions,organizations can use LLMs to enhance productiv

12、ity and unlock new opportunities while ensuring a smooth transition for their workforce.In addition,the structured approach proposed in this paper to analysing the direct impact on jobs also provides a case study for future waves of technological advancement across sectors.This white paper continues

13、 our Jobs of Tomorrow series,which has previously analysed Green and Social Jobs,and now seeks to analyse the impact of LLMs on jobs.At the end of 2023,this series will be concluded with a toolkit that serves as a call to action for businesses.The series complements the results of the Future of Jobs

14、 Report 2023,which delves deeper into the expectations of global business leaders on the direction of the workforce transition across all key geographies and industries.We are deeply grateful to the Centre for the New Economy and Society partners and constituents for their leadership of the jobs age

15、nda,as well as for the partnership of the Accenture team,whose members served as core collaborators on this report.The findings of this paper will serve as a key tool for the Jobs Consortium,which is a global coalition of ministers and chief executive officers that promotes a better future of work v

16、ia job creation and job transitions,as well as for jobs accelerators,which are country-specific platforms that facilitate public-private collaboration.Structured analysis,planning and proactive preparation by business,government and workers can ensure that generative AI and other technological advan

17、cements lead to an improved future of work and new opportunities for workers.Kathleen OReilly Communications,Media and Technology Industry Practices Chair,AccentureSaadia Zahidi Managing Director,World Economic ForumJobs of Tomorrow:Large Language Models and Jobs September 2023Jobs of Tomorrow:Large

18、 Language Models and Jobs3Executive summaryAs advances in generative artificial intelligence(AI)continue at an unprecedented pace,large language models(LLMs)are emerging as transformative tools with the potential to redefine the job landscape.The recent advancements in these tools,like GitHubs Copil

19、ot,Midjourney and ChatGPT,are expected to cause significant shifts in global economies and labour markets.These particular technological advancements coincide with a period of considerable labour market upheaval from economic,geopolitical,green transition and technological forces.The World Economic

20、Forums Future of Jobs Report 2023 predicts that 23%of global jobs will change in the next five years due to industry transformation,including through artificial intelligence and other text,image and voice processing technologies.This white paper provides a structured analysis of the potential direct

21、,near-term impacts of LLMs on jobs.With 62%of total work time involving language-based tasks,1 the widespread adoption of LLMs,such as ChatGPT,could significantly impact a broad spectrum of job roles.To assess the impact of LLMs on jobs,this paper provides an analysis of over 19,000 individual tasks

22、 across 867 occupations,assessing the potential exposure of each task to LLM adoption,classifying them as tasks that have high potential for automation,high potential for augmentation,low potential for either or are unaffected(non-language tasks).The paper also provides an overview of new roles that

23、 are emerging due to the adoption of LLMs.The longer-term impacts of these technologies in reshaping industries and business models are beyond the scope of this paper,but the structured approach proposed here can be applied to other areas of technological change and their impact on tasks and jobs.Th

24、e analysis reveals that tasks with the highest potential for automation by LLMs tend to be routine and repetitive,while those with the highest potential for augmentation require abstract reasoning and problem-solving skills.Tasks with lower potential for exposure require a high degree of personal in

25、teraction and collaboration.The jobs ranking highest for potential automationare Credit Authorizers,Checkers and Clerks(81%of work time could be automated),ManagementAnalysts(70%),Telemarketers(68%),StatisticalAssistants(61%),and Tellers(60%).Jobs with the highest potential for taskaugmentation emph

26、asize mathematicaland scientific analysis,such as InsuranceUnderwriters(100%of work time potentiallyaugmented),Bioengineers and BiomedicalEngineers(84%),Mathematicians(80%),andEditors(72%).Jobs with lower potential for automation oraugmentation are jobs that are expected toremain largely unchanged,s

27、uch as Educational,Guidance,and Career Counsellors and Advisers(84%of time spent on low exposure tasks),Clergy(84%),Paralegals and Legal Assistants(83%),and Home Health Aides(75%).In addition to reshaping existing jobs,the adoption of LLMs is likely to create newroles within the categories of AI Dev

28、elopers,Interface and Interaction Designers,AI ContentCreators,Data Curators,and AI Ethics andGovernance Specialists.An industry analysis is done by aggregatingpotential exposure levels of jobs to the industrylevel,noting that jobs may exist in more thanone industry.Results reveal that the industrie

29、swith the highest estimates of total potentialexposure(automation plus augmentationmeasures)are both segments of financialservices:financial services and capital marketsand insurance and pension management.Thisis followed by information technology and digitalcommunications,and then media,entertainme

30、ntand sports.Additional lists of jobs ranked byhighest exposure potential for each majorindustry category are compiled in the appendix.Similarly,a function group analysis revealsthat the two thematic areas with the greatesttotal potential exposure to LLMs are informationtechnology,with 73%of working

31、 hoursexposed,and finance,with 70%of workinghours exposed.As with the industry groups,additional lists of jobs ranked by highestexposure potential for each function groupare compiled in the Appendices.These new findings connect directly to earlierwork done by the Centre for the New Economyand Societ

32、y in the Future of Jobs Report 2023.Many of the jobs found to have high potentialfor automation by LLMs were also expectedby business leaders to undergo employmentdecline within the next five years,such as banktellers and related clerks,data entry clerks,and administrative and executive secretaries.

33、Meanwhile,jobs with high potential foraugmentation are expected to grow,such asAI and Machine Learning Specialists,DataAnalysts and Scientists,and Database andNetwork Professionals.Together,these twopublications identify and reaffirm salient themesin the connection between technological changeand la

34、bour market transformation.The findings of this report shed light on how implementing LLMs could alter the landscape of jobs,providing valuable insights for policy-makers,educators and business leaders.Rather than leading to job displacement,LLMs may usher in a period of task-based transformation of

35、 occupations,requiring proactive strategies to prepare the workforce for these jobs of tomorrow.Jobs of Tomorrow:Large Language Models and Jobs4Introduction:How will large language models impact the jobs of tomorrow?Labour markets are undergoing rapid transformation from the trajectory of growth,geo

36、economics,sustainability and technology.The Future of Jobs Report 2023 found that business leaders expect 23%of global jobs to change in the next five years.2 In particular,generative artificial intelligence(AI)has undergone a profound leap in capabilities,embodied in products such as GitHubs Copilo

37、t for programming,Midjourney for image generation and ChatGPT as a universal language assistant.The Future of Jobs Report 2023 also found that AI and text,image and voice processing technologies more generally are top of mind for businesses.The report found that 75%of survey respondents report havin

38、g plans to adopt AI in their organizations operations,and 62%report having plans to adopt text,image and voice processing technologies.3 This has raised questions about how this new technology will affect organizations and labour markets around the world.This white paper examines the potential near-

39、term,direct impact on jobs of a particular type of generative AI,large language models(LLMs),which have been highly visible in public debate over the past year due to their human-like ability to create and understand language.As LLM services have exploded in popularity,with free services such as Cha

40、tGPT reaching as many as 100 million active users within the first two months of its debut,4 the capabilities of these models,paired with their accessibility and rapid adoption rate,suggest that many work tasks and jobs that emphasize them could be impacted by the use of LLMs in the years to come.By

41、 some estimates,up to 62%of work time involves language-based tasks.5 Yet,artificial intelligence and text,image and voice processing technologies also have the potential to augment work and create new jobs.In addition,many roles remain wholly unaffected by these developments.Rapid technological cha

42、nge often generates anticipation regarding its effects on daily life,particularly jobs.In the aggregate,previous innovations have led to more employment opportunities,better-quality jobs and a higher quality of life,but they also create disruption and displacement.6 This paper aims to support the de

43、tailed analysis required to take a clear-eyed view around impact,opportunity and preparation.Generative AI,LLMs and language tasksThe newest forms of groundbreaking generative AI models are created via deep learning,which is the process of training foundation models on very large sets of data.These

44、foundation models are typically created in the form of a neural network,whose structure is inspired by the arrangement of neurons in the human brain.Large foundation models are trained on vast amounts of data and have seemingly super-human levels of predictive capacity,which can be harnessed by prod

45、ucing text or images in response to a written prompt.7So far,generative AI models have been configured into a variety of different tools to serve different contexts,such as image,audio or video creation,identifying financial fraud and other security risks,and a host of general language capabilities,

46、including the ability to generate natural,mathematical and computational language.While there is a broad range of implementations of generative AI,this study will focus on LLMs and their unique language-generating capabilities,as these models have the greatest potential to impact the largest number

47、of jobs in the near term.LLMs can perform a broad spectrum of language tasks,usually in response to a simple user prompt,on nearly any topic:LLMs can reformulate and provide detailed feedback on a provided set of text,including summarizing it,translating it to another language,proofreading it,discus

48、sing its style or tone of voice and even rewriting it in a different style or tone of voice.LLMs can also generate new text and provide some degree of expertise on topics present in the LLMs training data,such as in the form of a literature review or completing a task typically done by a research as

49、sistant.8 As programming languages are text,LLMs can serve as programming assistants.Implementations like GitHubs Copilot have been shown to increase programmer productivity by 56%.9 Language capabilities overlap substantially with tasks performed on the job,with estimates suggesting that up to 62%o

50、f work time involves language-based tasks.Jobs of Tomorrow:Large Language Models and Jobs5By integrating LLMs with other systems,these capabilities can be extended to a greater range of abstract tasks,such as scheduling meetings,placing orders,responding to emails,or providing research on a particul

51、ar topic.Given the large overlap between LLM capabilities and current job tasks,how will introducing LLMs into the workplace change jobs?Which parts of a job will be impacted the most,and which jobs will be impacted most?Finally,with the introduction of these new technologies,which new jobs can be e

52、xpected to arise?A task-based approach to job exposureTo answer these questions,the methods deployed in this white paper assess the potential exposure of language-based job tasks to the ability of LLMs to perform these tasks.The approach is to first think of a job as consisting of many different tas

53、ks and then assess how each task may be affected by LLMs.The magnitude of impact on a job ultimately depends on the degree of language-based skills required for specific tasks in that job and the time spent on those tasks.Language-dependent,standardized,routine and process-oriented tasks are prime c

54、andidates for automation and replacement by LLMs.At the same time,those requiring a greater degree of human interaction are more likely to be augmented and performed in collaboration with LLMs.For example,some job tasks are routine and predictable and are performed by people working individually,suc

55、h as clerks and administrators,which involves reading and entering data,cross-referencing records between different databases and reviewing transactions.These tasks are more likely to be exposed to and ultimately automated by the introduction of LLMs,implying that they will no longer be performed by

56、 humans.The outcome is that jobs emphasizing these tasks will either transform to take on non-automatable tasks or go into decline.Other job tasks require a great deal of abstract reasoning,creativity and problem-solving.While language tasks may not be their primary product,they may rely heavily on

57、language and communication.For example,Mathematicians and Editors rely heavily upon language,yet need to incorporate creative insights from their fields of expertise.Similarly,Software Developers work a lot with computer languages but also need to grasp complex systems at various levels of abstracti

58、on to create a finished software product.Workers in these jobs would not have their tasks replaced by LLMs;rather,LLMs would supercharge their ability to complete these tasks.Teachers,for example,could rely on LLMs for assistance in lesson planning and correcting student work.According to one study

59、in the US,three in ten teachers have already used ChatGPT for lesson planning(30%),generating creative ideas for classes(30%)and building background knowledge for lessons and classes(27%).10 Software Developers could turn to LLMs to generate standardized blocks of code with clear functional paramete

60、rs,speeding up the development process and allowing for more time to be spent on high-level architectural tasks.Software Developers also perform many tasks with high potential for automation,suggesting that many jobs will be transformed rather than automated or augmented.The research methods employe

61、d in this paper aim to identify which tasks will be exposed to LLMs and how they will be impacted:whether they have the potential to be automated and replaced by LLMs or augmented and enhanced by LLMs.Data for analysis comes from O*NET and the United States Bureau of Labor Statistics(BLS),which char

62、acterizes 867 jobs with respect to over 19,000 individual tasks.Using both machine learning and manual methods,job tasks are individually rated with respect to their potential exposure to the adoption of LLMs,thereby classifying them into one of four categories:1.High potential for automation:Going

63、forward,the task will be performed by LLMs,not humans.2.High potential for augmentation:Humans will continue to perform the task,and LLMs will increase human productivity.3.Low potential for automation or augmentation:Humans will continue to perform the task with no significant impact from LLMs.4.Un

64、affected(i.e.non-language tasks).Job tasks are then mapped to the occupations in which they are deployed along with a share of time spent on each task,and with both of these metrics,a measure of potential exposure to LLMs is created at the occupation level.Chapter 1 of this paper presents these resu

65、lts for tasks and jobs in detail,using the detailed occupations list from the Standard Occupation System(SOC)from the US Bureau of Labor Statistics,the highest resolution list available,featuring 867 occupation titles.11 This chapter also provides analysis of the differences between industries and f

66、unctions in terms of the expected impact of LLMs on jobs.In chapter 2,these detailed occupations are aggregated and mapped to the occupation classification system used in the Future of Jobs Survey 2022 to directly connect results on exposure of jobs to LLMs to survey results of global business leade

67、rs on the potential for growth or decline of specific jobs,and the forces underlying these trends,as covered in greater detail in the Future of Jobs Report 2023.Three in ten teachers have already used ChatGPT for lesson planning(30%),generating creative ideas for classes(30%)and building background

68、knowledge for lessons and classes(27%).Jobs of Tomorrow:Large Language Models and Jobs6Identifying exposure potential of tasks and jobs1LLMs hold transformative potential and are set to significantly reshape the future employment landscape.A preliminary analysis reveals which tasks have the highest

69、potential for automation or augmentation and which have lower or no potential(see Table 1).The tasks with the highest potential for automation by an LLM tend to be more routine,such as performing administrative or clerical activities,and some tasks that relate to elementary analysis,such as designin

70、g databases or analysing data.The tasks with the highest potential for augmentation require more abstract reasoning skills,especially those that combine interaction with people.At the top of the list is evaluating personnel capabilities or performance,such as in the context of the responsibilities o

71、f a human resources professional,followed by collecting data about consumer needs or opinions.For the latter,while running a survey,for example,could be a highly automated process via email and the internet,the crafting and wording of survey questions still require a high degree of attention and app

72、roval by the person collecting the data.Tasks with lower potential for exposure require a high degree of personal interaction and collaboration,such as negotiation of contracts,development of educational programmes,and other scientific and technical work,the latter of which already employ a strong d

73、egree of technical augmentation.Finally,non-language tasks are,as expected,those that emphasize physical movement,such as loading products,materials or equipment for transport,assembly activities,agricultural activities,and grooming and hairstyling.The tasks with the highest potential for augmentati

74、on require more abstract reasoning skills,especially those that combine interaction with people.1.1 Exposed tasksThis chapter outlines task exposure to LLMs,providing deeper analysis on two jobs highly likely to be impacted,and ranks jobs by their automation and augmentation potential.It also identi

75、fies emerging jobs due to LLM adoption and summarizes exposure risks by industry and job function.Jobs of Tomorrow:Large Language Models and Jobs7Key tasks impactedTABLE 1LevelTaskHigher potential for automationPerform administrative or clerical activitiesDesign databasesAnalyse data to improve oper

76、ationsMonitor external affairs,trends or eventsObtain information about goods or servicesDocument technical designs,procedures or activitiesHigher potential for augmentationEvaluate personnel capabilities or performanceCollect data about consumer needs or opinionsRead documents or materials to infor

77、m work processesEvaluate patient or client condition or treatment optionsPrepare informational or instructional materialsTest performance of computer or information systemsLower potential for exposure(automation or augmentation)Negotiate contracts or agreementsAdvocate for individual or community ne

78、edsCollaborate in the development of educational programmesDirect scientific or technical activitiesCoordinate with others to resolve problemsEvaluate designs,specifications or other technical dataNon-language tasksLoad products,materials or equipment for transport or further processingAssemble equi

79、pment or componentsPrepare mixtures or solutionsPerform agricultural activitiesGroom or style hairInstall energy or heating equipmentJobs of Tomorrow:Large Language Models and Jobs8Example of an exposed and non-exposed jobFIGURE 1Software Developers(more exposed)Human Resource Managers(less exposed)

80、AutomationAugmentationLower potentialNon-language tasksHigher potential for automation:Higher potential for augmentation:Lower potential for automation or augmentation:Analyse data to improve operations Prepare informational or instructional materials Coordinate with others to resolve problems Commu

81、nicate with others about business strategies Evaluate the characteristics,usefulness or performance of products or technologies Analyse performance of systems or equipmentHigher potential for automation:Higher potential for augmentation:Lower potential for automation or augmentation:Determine resour

82、ce needs of projects or operations Explain regulations,policies or procedures Interview people to obtain information Coordinate group,community or public activities Train others on operational or work procedures Manage budgets or finances28.7%16.1%22.2%61.7%43.2%28%1.2 Detailed examples of exposed j

83、obsTo provide a detailed example of how the tasks involved in a job determine how LLMs will impact the job,Figure 1 presents an analysis of an exposed and non-exposed job and the task exposure of each.The left panel of the figure provides an overview of Software Developers,a highly exposed job in th

84、e analysis,showing high potential for both augmentation and automation of tasks.A total of 28.7%of time spent in the occupation has high potential for automation by LLMs,including“analyse data to improve operations”and“analyse the performance of systems of equipment”.In contrast,up to 43.2%of time s

85、pent on tasks in the occupation has high potential for augmentation,including“preparing informational or instructional materials and evaluating the characteristics,usefulness or performance of products or technologies”.The right panel of Figure 1 provides an overview of Human Resource Managers,which

86、 is a less exposed job.Only 16.1%of time has potential for automation,including“determine resource needs of projects or operations and manage budgets or finances”,and 22.2%of time has potential for augmentation,including“explain regulations,policies or procedures and train others on operational or w

87、ork procedures”.The majority of tasks involved,totalling 61.7%of time spent,have lower potential for exposure,as these tasks involve working directly with individuals and coordinating and communicating with large groups.Jobs of Tomorrow:Large Language Models and Jobs9Jobs with the highest potential

88、for automationFIGURE 20102030405060708090100Management AnalystsTelemarketersStatistical AssistantsTellersForensic Science TechniciansCredit Authorizers,Checkers and ClerksOccupationsReceptionists and Information ClerksBrokerage ClerksProduction,Planning and Expediting Clerks76%87%74%93%60%88%69%74%7

89、2%63%63%78%77%80%63%File ClerksWord Processors and TypistsBookkeeping,Accounting and Auditing ClerksLegal Secretaries and Administrative AssistantsLoan Interviewers and ClerksBill and Account Collectors81%70%68%61%60%58%58%58%57%56%7%11%26%15%18%16%17%10%10%34%18%13%13%26%7%7%24%7%12%1%4%11%37%31%Ex

90、posure55%55%54%54%53%5%23%23%27%9%21%17%13%7%12%11%22%40%AutomationAugmentationLower potentialNon-language tasks1.3 Analysis by occupationJobs with potential for automationResults from the task-based analysis reveal that jobs with the highest potential for automation of tasks by LLMs emphasize routi

91、ne and repetitive procedures and do not require a high degree of interpersonal communication.Roles with the highest amount of potentially automatable work time are Credit Authorizers,Checkers and Clerks(81%of time),Management Analysts(70%),Telemarketers(68%),Statistical Assistants(61%)and Tellers(60

92、%).Jobs with high potential for automation often include various kinds of office clerks,particularly those focused on record-keeping and managing information tasks where LLMs have demonstrated a strong degree of competency.For example,Legal Secretaries and Administrative Assistants spend approximate

93、ly 54%of their time on tasks with high automation potential.Jobs of Tomorrow:Large Language Models and Jobs10Jobs with the highest potential for augmentationFIGURE 30102030405060708090100Bioengineers and Biomedical EngineersMathematiciansEditorsDatabase ArchitectsStatisticiansInsurance UnderwritersO

94、ccupationsExposureTraining and Development Specialists ClerksDatabase AdministratorsInsurance Appraisers,Auto DamageGraphic Designers12%72%15%15%6%16%34%14%65%18%3%66%17%66%72%80%7%28%14%84%68%68%16%26%100%16%AutomationAugmentationLower potentialNon-language tasksProperty Appraisers and AssessorsApp

95、raisers and Assessors of Real EstateOperations Research AnalystsMedical Transcriptionists100%93%72%86%84%100%74%83%100%79%88%88%78%100%76%Interpreters and Translators26%26%18%40%16%60%24%60%60%22%62%12%62%12%Jobs with potential for augmentationThe same analysis methods demonstrate that the jobs with

96、 the highest potential for augmentation by LLMs emphasize critical thinking and complex problem-solving skills,especially those in science,technology,engineering and mathematics(STEM)fields(see Figure 3).Topping the list is Insurance Underwriters,with analysis suggesting that they spend 100%of their

97、 time on tasks that have the potential to be augmented by generative AI systems.This is followed by Bioengineers and Biomedical Engineers(84%of time augmentable),Mathematicians(80%)and Editors(72%).The remaining top 15 jobs are likewise technical or highly specialized,often requiring advanced degree

98、s or training,such as Database Architects and Statisticians.Note that many jobs with the highest potential for augmentation also have some potential for automation,resulting in very high total exposure for these jobs,such as Medical Transcriptionists,Insurance Appraisers and Assessors of Real Estate

99、.Jobs of Tomorrow:Large Language Models and Jobs11Jobs with lower potential for transformation and non-language tasksJobs emphasizing non-language tasks are expected to be less exposed,or not exposed at all,to the potential impacts of LLMs.Results of the task analysis suggest this,indicating that jo

100、bs with the lowest potential of exposure(either automation or augmentation)are those that require a high degree of personal interaction,such as Healthcare Professionals or Teachers,or physical movement,such as Athletes or Manual Labourers (see Figure 4).The occupation with the highest proportion of

101、tasks rated low potential for transformation,at 84%of total time,is Educational,Guidance,and Career Counsellors and Advisers.This is followed by Clergy(84%of time),Paralegals and Legal Assistants(83%),Home Health Aides(75%)and then Anaesthesiologists(74%).Community,social service and healthcare occu

102、pations feature prominently among those with low potential for automation or augmentation,making up 10 of the top 15 occupations with the least exposure potential.Jobs with the lowest potential for exposureFIGURE 40102030405060708090100ClergyParalegals and Legal AssistantsHome Health AidesAnaesthesi

103、ologistsHealthcare Social WorkersEducational,Guidance,and Career Counsellors and AdvisersOccupationsMarriage and Family TherapistsTitle Examiners,Abstractorsand SearchersOral and Maxillofacial SurgeonsExposure16%17%16%26%26%16%27%27%27%0%10%32%30%34%35%Athletes and Sports Competitors3%73%73%27%27%73

104、%27%73%21%5%74%17%83%75%16%9%26%74%84%6%10%84%8%8%25%73%1%AutomationAugmentationLower potentialNon-language tasksFarm Labour ContractorsMental Health CounsellorsSubstitute Teachers,short-termMental Health and Substance Abuse Social WorkersPaediatricians,general9%6%26%68%3%67%16%14%34%34%66%66%72%18%

105、Beyond jobs with low potential exposure,a number of jobs feature no language tasks at all and have no potential to be impacted by the adoption of LLMs in the workplace.These jobs include:Dishwashers Highway Maintenance Workers Meat,Poultry and Fish Cutters and Trimmers Rail-Track Laying and Maintena

106、nce Equipment Operators Helpers,Carpenters Paper Goods Machine Setters,Operators and Tenders Slaughterers and Meat Packers Roustabouts,Oil and Gas Pressers,Textile,Garment and Related Materials Fibreglass Laminators and FabricatorsJobs of Tomorrow:Large Language Models and Jobs12Emerging jobsAs gene

107、rative AI introduces a new paradigm of collaboration between humans and AI,it will redefine how work is conducted and reshape the nature of various job roles.No predictions can be 100%certain regarding which new roles may appear with the widespread adoption of LLMs.Still,it is apparent that there is

108、 room for job development in several key areas.The following illustrative groupings of emerging jobs can help unlock the value of generative AI and mitigate associated consequences.AI Model and Prompt Engineers:Engineers and scientists will continue developing and fine-tuning LLMs at the most detail

109、ed level of AI systems innovation.Some of the skill sets in these jobs may already exist,but they will continue to evolve simultaneously with AI systems progress.These jobs cover the range of programmers designing more efficient algorithms,electrical engineers designing custom chips to train and run

110、 the models,systems administrators building server infrastructure,and infrastructure and power systems engineers ensuring these systems have the stable energy sources needed for extended runs.In addition,Prompt Engineers will be critical to developing,refining and reframing prompts or inputs for LLM

111、s to reach more optimal results.Interface and Interaction Designers:Completed and trained LLMs are still highly technical and will require well-crafted interfaces to be accessible to the public.In some ways,Interface and Interaction Designers can be considered user experience(UX)designers for LLMs.T

112、his family of jobs will be responsible for crafting LLMs to adapt to a particular kind of user input(for example,typing or spoken voice)or to perform particular tasks,such as in the development of personalized AI assistants,tutors or coaches.These jobs could include the important stage of reinforcem

113、ent learning with human feedback(RHLF),in which models are trained on favoured responses,and other performance evaluators.AI Content Creators:Building off the infrastructure of Technologists and Interface Designers,AI Content Creators will harness the knowledge and understanding of LLMs to rapidly p

114、roduce in-depth content on a topic in any field or domain.The type of content produced could vary from articles and books to teaching and training material to entire storylines for movies and television series,potentially automatically generating any accompanying visual and audio media.Data Curators

115、 and Trainers:Massive training data sets are integral to maintaining the performance of LLMs.Ensuring high-quality data is a priority in LLM development,as the quality of an LLMs output directly reflects the quality of its training data.As most training data are curated from text posted to the inter

116、net,data quality and integrity checks are critical and will lead to the development of a dedicated,specialized workforce.Ethics and Governance Specialists:The presence of prejudiced or other unsavoury language in training data can lead LLMs to produce biased,harmful or unethical content.Not only wil

117、l training data need to be checked for quality,but trained LLM systems will need to be rigorously tested before being released to the public.This will fall into the purview of specially trained AI Safety Officers and Ethicists at the company level and even domain-specific regulators and lawyers at t

118、he government level.AI Content Creators will harness the knowledge and understanding of LLMs to rapidly produce in depth content on a topic in any field or domain.Jobs of Tomorrow:Large Language Models and Jobs13While the emergence of new job categories can be expected,the reinvention of existing ro

119、les should also be anticipated.Analysis of the impact of LLMs on customer service jobs found that,of the 13 core customer service tasks,four tasks remained unchanged and within human capabilities,four tasks could be fully automated using generative AI,five tasks could be augmented to enhance human p

120、erformance,and five new high-value tasks emerge.12 With generative AI,Customer Service Representatives(CSRs)can engage in new tasks like providing feedback for system improvement,aligning with customer needs,testing for biases and ensuring ethical machine behaviour,and monitoring data privacy.These

121、responsibilities empower CSRs to shape AI deployment,optimize customer experiences and uphold ethical standards in customer service operations.Additionally,another study of CSRs found that the implementation of generative AI was associated with lower employee turnover.13 These findings demonstrate h

122、ow organizations could use generative AI alongside human expertise to rethink job design,enhance productivity and improve employee experience.1.4 Analysis by industryTo take the analysis a step further,the job exposure ratings may be aggregated to generate estimates for potential automation or augme

123、ntation at the industry level(see Figure 5).To generate these estimates,the exposure measures for all occupations within a particular industry are averaged,weighted by total employment,and taking into account that occupations may belong to more than one industry.The two industries with the highest e

124、stimates of total potential exposure(automation plus augmentation measures)are both segments of financial services:financial services and capital markets,and insurance and pension management.This is followed by information technology and digital communications,and then media entertainment and sports

125、.A trend to note with the industry estimates is that industries with high potential for exposure to LLMs have high potential for both automation and augmentation.This suggests that introducing these new technologies will change the nature of the labour market but will not necessarily reduce the tota

126、l number of jobs.The industries rated with high potential exposure also plan to adopt AI technologies,as the business leaders surveyed in the Future of Jobs Report 2023 reported.Three of the top five industries planning to adopt AI technologies are those with the greatest exposure to LLMs namely,ins

127、urance and pension management,information technology services,and media,entertainment and sports.14 Introducing these new technologies will change the nature of the labour market but will not necessarily reduce the total number of jobs.14Jobs of Tomorrow:Large Language Models and JobsIndustries with

128、 the highest exposure(automation and augmentation)FIGURE 50102030405060708090100Financial services and capital markets Insurance and pensions managementInformation and technology servicesTelecommunicationsMedia and publishingResearch design and business management servicesRental,reservation and leas

129、ing servicesRetail wholesale of consumer goodsNon-profit organizations,professional bodies and unionsReal estateArts,entertainment and recreationGovernment and public sectorOil and gasElectronics manufacturingPersonal care,well-being and repair servicesEducation and trainingEnergy technology and uti

130、litiesAutomotive and aerospaceMedical and healthcare servicesAdvanced manufacturingAccommodation,food and leisure servicesWater and waste managementCare and social work servicesEngineering and constructionEmployment servicesSupply chain and transportBusiness support and premises maintenance services

131、Chemical and advanced materialsMining and metalsProduction of consumer goodsAgriculture,forestry and fishingAutomationAugmentationLower potentialNon-language tasksJobs of Tomorrow:Large Language Models and Jobs15The appendix to this report contains additional lists of jobs ranked by exposure for eac

132、h of the 31 primary industry groups.15 For these tables,jobs are ranked by total exposure potential,which is the sum of automation and augmentation measures.Similar to the job exposure rankings by industry,the appendix to this report also includes additional lists of jobs ranked by exposure for each

133、 of the eight function groups.For these tables,just as with industry groups,jobs are ranked by total exposure potential,which is the sum of automation and augmentation measures.1.5 Analysis by functionThe exposure ratings for occupations may also be aggregated into functional groups,which reveals si

134、milar themes for potential automation and augmentation(see Figure 6).As was found with the industry analysis and much of the analysis in the Future of Jobs Report 2023,the two thematic areas with the greatest total potential exposure to LLMs are information technology,with a total of 73%of work time

135、 exposed,and finance,with a total of 70%of work time exposed.These functions are followed by customer sales(67%total exposure),operations(65%),and human resources(57%).Job functions likely to be automated also tend to have a high chance of being enhanced or augmented by technology,and vice versa.Thi

136、s is an important counterpoint to the notion that technological innovation displaces jobs:technological innovation transforms jobs,with some tasks being eliminated and others becoming more important.Job function groups with the highest exposure(automation and augmentation)FIGURE 60102030405060708090

137、100FinanceCustomer SalesOperationsHuman Resources(HR)MarketingIT/TechnologyOccupationsLegalSupply Chain41%26%32%28%21%10%1%42%34%16%17%33%21%18%17%44%35%41%3%3%22%34%41%22%27%50%4%39%19%14%18%29%AutomationAugmentationLower potentialNon-language tasksJobs of Tomorrow:Large Language Models and Jobs16L

138、LMs and the growth and decline of jobs and tasks2Growth expectations underscore workforce shifts in job tasks and titles,suggesting readiness for transformation.This section identifies common themes between the results presented in this white paper and the findings on expected growth and decline in

139、jobs collected from global business leaders in the Future of Jobs Survey,presented in the Future of Jobs Report 2023.Themes emerge in job tasks and specific roles,especially those susceptible to automation,augmentation and those with lower exposure potential.2.1 Expected growth and decline of tasks2

140、.2 Expected growth and decline of jobsThe analysis presented in the Future of Jobs Report 2023 indicates some striking parallels between the results in Table 1 and the predictions of global business leaders.16 The report found that the number one task predicted to be automated now and in the next fi

141、ve years is information and data processing.The present task-based analysis also indicates that this task has high potential for automation,namely in designing databases,analysing data to improve operations and obtaining information about goods or services.Similarly,results from the Future of Jobs R

142、eport 2023 found that the lowest potential for automation will be in reasoning and decision-making tasks.This papers task-based analysis denotes decision-making tasks as having high potential for augmentation(especially for evaluating personal capabilities or performance,reading documents or materia

143、ls to inform work processes,and evaluating patient or client condition or treatment options),or low potential for exposure(such as direct scientific or technical activities).The fact that the survey results from the Future of Jobs Report 2023 and the task-based analysis in this paper identify common

144、 threads underlying the workforce transition suggests that these changes are indeed fundamental shifts.Additionally,it implies that business leaders are already identifying these trends and are expected to be in a good position to prepare their workforces for future changes.Figure 7 combines the key

145、 findings on job exposure to LLMs in this white paper with growth expectations for the same jobs,as identified in the Future of Jobs Report 2023.In this figure,the vertical axis indicates the exposure potential of jobs,with augmentation potential scores in the top box,low potential scores in the mid

146、dle box and automation scores in the bottom box.The horizontal axis indicates the net expected growth for jobs within the next five years,measured as the expected percentage change in workforce employment.Note that to produce this chart,the O*NET detailed occupations used earlier in the report have

147、been mapped to the job classification used by the Future of Jobs Report 2023.The most immediate takeaway from the figure is the positive association between job augmentation and growth and the negative association between job automation and growth.In contrast,jobs with lower potential for exposure h

148、ave much lower expected growth.Jobs of Tomorrow:Large Language Models and Jobs17Job exposure potential vs growth potentialFIGURE 7100755025100755025Exposure%(matched with Future of Jobs and re-calculated)Net growth%StableGrowingHighest augmentationLower potentialHighest automationDeclining100755025-

149、40-30-20-10-3-2-10110203040Business and financial operationsEducation and trainingHealthcare practitioners and techniciansManagementSalesArts,design,entertainment,sports and mediaFuture of Jobs job familiesComputer and mathematicalInsurance UnderwritersFarming,fishing and forestryLegalOffice and adm

150、inistrativeWeb DevelopersSales and Purchasing Agents and BrokersAuthors and JournalistsData Warehousing SpecialistsManagement and Organization AnalystsTraining and Development SpecialistsMathematicians,Actuaries and StatisticiansGraphic DesignersDatabase and Network ProfessionalsDatabase Analysts an

151、d ScientistsAI and Machine Learning SpecialistsRepresentatives,Wholesale and Manufacturing,Technical and Scientific Products Secondary Education TeachersParalegals and Legal AssistantsBank Tellers and Related ClerksTelemarketersSecurities and Finance Dealers and BrokersAccounting,Bookeeping and Payr

152、oll ClerksLegal SecretariesStatistical,Finance and Insurance ClerksAdministrative and Executive SecretariesMaterial-recording and Stock-keeping ClerksData Entry ClerksHuman Resources SpecialistsLawyersManufacturing,Mining,Construction and Distribution ManagersFarmworkers and LabourersUniversity and

153、Higher Education TeachersVocational Education TeachersPersonal Care Workers in Health ServicesRegulatory andGovernment AssociateProfessionalsThe jobs with the highest potential for automation are featured in the lower box of Figure 7(mapped from the list in Figure 2)and overlap significantly with th

154、ose that are expected to decline over the next five years,as indicated by survey results from business leaders in the Future of Jobs Report 2023.17 The three with the greatest expected declines are Bank Tellers and Related Clerks(87%of time spent on tasks with high potential for automation,41%expect

155、ed decline),Data Entry Clerks(58%potential automation,36%expected decline),and Administrative and Executive Secretaries(69%potential automation,34%expected decline).All jobs with high potential for automation have negative expected growth,with the exception of Securities and Finance Dealers and Brok

156、ers.The jobs with high potential for augmentation are featured in the upper box(mapped from the list in Figure 3)and mostly align with jobs with strong expected growth over the next five years,as indicated in the Future of Jobs Report 2023.18 Jobs largely in the families of technologists,engineers a

157、nd analysts have potential for augmentation and growth.AI and machine learning specialists have the highest expected growth(75%of time spent on tasks with high potential for augmentation,and 39%expected growth),followed by Data Analysts and Scientists(84%potential augmentation,and 31%expected growth

158、)and Database and Network Professionals(83%potential augmentation,and 14%expected growth).However,jobs with high potential for augmentation feature significant variation in expected growth,and several jobs in this category have low growth expectations,such as Data Entry Clerks and Web Developers.Nev

159、ertheless,the general congruence between augmentation potential and expected growth suggests that adopting generative AI,with its potential to increase individual worker productivity,could fuel significant job growth.Of the jobs with low potential for exposure,indicated in the middle box(mapped from

160、 the list in Figure 4),many are found in the field of education,especially University and Higher Education Teachers,which were also identified in the Future of Jobs Report 2023 as having modest expected growth within the next five years(41%of time spent on tasks with low potential for exposure to LL

161、Ms,and 10%expected growth).19 Jobs with low potential for exposure also vary in their expected growth numbers but tend to average close to zero expected growth.The prominence of care jobs among those with lower potential for automation by LLMs aligns with earlier findings from the previous edition o

162、f the Jobs of Tomorrow report.That report found high levels of unmet demand for care and education jobs across many countries from all income groups.20 This could be in part because generative AI cannot easily automate these jobs,and yet demand for their services will only continue to grow in the co

163、ming years.Adopting generative AI,with its potential to increase individual worker productivity,could fuel significant job growth.Jobs of Tomorrow:Large Language Models and Jobs18Conclusion:Ensuring that large language models work for workersAdopting LLMs will transform business and the nature of wo

164、rk,displacing some existing jobs,enhancing others and ultimately creating many new roles.Yet,it will be incumbent upon businesses and governments to take proactive steps in preparing the workforce for the imminent transformation to ensure that all members of society benefit from the potential of gen

165、erative AI.This paper takes a structured approach to assessing the impact positive and negative of LLMs on jobs,allowing stakeholders to responsibly address both challenges and opportunities.Policy-makers will need to adapt strategicworkforce planning capabilities,lifelong learningsystems and social

166、 safety nets to manage theupcoming period of disruption.Similar analysisto that shared in this paper can help providemore precise views of the situation in specificgeographies.Governments can also partnerwith and support employers and educationalinstitutions to provide training programmes thatprepar

167、e workers for the jobs that will grow andbenefit the most from LLMs.Additionally,socialsafety nets and assistance in transitioning tonew roles will need to be reimagined and be more precisely targeted for those most likely to be affected.Business leaders can use insights on the directimpact of LLMs

168、on jobs to understand whichroles will be most affected and responsiblysupport the transition of workers to newroles and ways of working.Internal workforceplanning,learning and development,andtalent management practices should alsobe strengthened to support the adoptionof generative AI in the workpla

169、ce,recruitnew talent in growing jobs or invest heavilyin reskilling and upskilling workers towardsgrowing roles.Large language models present an opportunity to extend human potential,grow industries and strengthen global economies.Yet their rapid adoption contains both risks and opportunities for th

170、e workforce.The approach presented in this white paper helps plan for the direct impact on tasks and jobs and informs government,business and workers on the actions they can take now to prepare for the future.Jobs of Tomorrow:Large Language Models and Jobs19AppendicesA1 Exposure potential by industr

171、y groupsJob exposure by industry:ranked by exposure(augmentation and automation potential)FIGURE 80102030405060708090100Hosts,Restaurant,Lounge and Coffee ShopWaiting StaffMaids and Housekeeping CleanersFast Food and Counter WorkersBartendersFirst-Line Supervisors of Food Preperation and Serving Wor

172、kersAccommodation,Food and Leisure:Accommodation,food and leisure servicesCooks,RestaurantCooks,Fast FoodFood Preperation WorkersDishwashers0102030405060708090100Counter and Rental ClerksSales Representatives of Services,expect Advertising,Insurance,Financial Services and TravelCustomer Service Repr

173、esentativesGeneral and Operations ManagersLight Truck DriversTravel AgentsAccommodation,Food and Leisure:Rental,reservation and leasing servicesHeavy Tractor-Trailer Truck DriversCleaners of Vehicles and EquipmentLabourers and Freight,Stock and Material Movers,HandBus and Truck Mechanics and Diesel

174、Engine Specialists0102030405060708090100First-Line Supervisors of Production and Operating WorkersMaintenance and Repair Workers,GeneralPackers and Packagers,HandFarmworkers and Labourers,Crop,Nursery and GreenhousePackaging and Filling Machine Operators and TendersHeavy and Tractor-Trailer Truck Dr

175、iversAgriculture and Natural Resources:Agriculture,forestry and fishingLabourers and Freight,Stock and Material Movers,HandMeat,Poultry and Fish Cutters and TrimmersSlaughterers and Meat PackersIndustrial Truck and Tractor Operators0102030405060708090100Customer Service RepresentativesGeneral and Op

176、erations ManagersFirst-Line Supervisors of Retail Sales WorkersParts SalespersonsMiscellaneous Assemblers and FabricatorsRetail SalespersonsAutomotive and Aerospace:Automotive and aerospaceAutomotive Service Technicians and MechanicsInspectors,Testers,Sorters,Samplers and WeighersCleaners of Vehicle

177、s and EquipmentLabourers and Freight,Stock and Material Movers,HandAutomationAugmentationLower potentialNon-language tasksJobs of Tomorrow:Large Language Models and Jobs20010203040506070809010001020304050607080901000102030405060708090100Social and Human Service AssistantsSubstance Abuse,Behavioural

178、Disorder and Mental Health CounsellorsChild,Family and School Social WorkersLicensed Practical and Licensed Vocational NursesRegistered NursesTeaching Assistants,except Post-SecondaryCare,Personal Services and Well-being:Care and social work servicesCare,Personal Services and Well-being:Personal car

179、e,well-being and repair servicesHome Health and Personal Care AidesChildcare WorkersPreschool Teachers,except Special EducationNursing Assistants0102030405060708090100Retail SalespersonsGeneral and Operations ManagersPharmacy TechniciansAnimal CaretakersPharmacistsCashiersHairdressers,Hairstylists a

180、nd CosmetologistsManicurists and PedicuristsAutomotive Service Technicians and MechanicsCleaners of Vehicles and EquipmentSecretaries and Administrative Assistants,except Legal,Medical and ExecutiveTeaching Assistants,except Post-SecondaryEducation Administrators,Kindergarten through SecondarySecond

181、ary School Teachers,except Special and Career/Technical EducationElementary School Teachers,except Special EducationOffice Clerks,GeneralEducation and Training:Education and trainingEnergy and Materials:Chemical and advanced materialsMiddle School Teachers,except Special and Career/Technical Educati

182、onSubstitute Teachers,short-termEducational,Guidance and Career Counsellors and AdvisersJanitors and Cleaners,except Maids and Housekeeping CleanersHeavy and Tractor-Trailer Truck DriversFirst-Line Supervisors of Production and Operating WorkersMiscellaneous Assemblers and FabricatorsChemical Equipm

183、ent Operators and TendersInspectors,Testers,Sorters,Samplers and WeighersGeneral and Operations ManagersPackaging and Filling Machine Operators and TendersLabourers and Freight,Stock and Material Movers,HandCutting,Punching and Press Machine Setters,Operators,and Tenders,Metal and PlasticWelders,Cut

184、ters,Solderers and BrazersAutomationAugmentationLower potentialNon-language tasksJobs of Tomorrow:Large Language Models and Jobs21AutomationAugmentationLower potentialNon-language tasksGeneral and Operations ManagersElectrical and Electronics Repairers,Powerhouse,Substation and RelayElectrical Engin

185、eersIndustrial Machinery MechanicsFirst-Line Supervisors of Mechanics,Installers and RepairersCustomer Service RepresentativesEnergy and Materials:Energy technology and utilitiesEnergy and Materials:Mining and metalsFirst-Line Supervisors of Production and Operating WorkersControl and Valve Installe

186、rs and Repairers,except Mechanical DoorPower Plant OperatorsElectrical Power-Line Installers and RepairersIndustrial Machinery MechanicsFirst-Line Supervisors of Construction Trades and Extraction WorkersHeavy and Tractor-Trailer Truck DriversOperating Engineers and Other Construction Equipment Oper

187、atorsFirst-Line Supervisors of Production and Operating WorkersGeneral and Operations ManagersService Unit Operators,Oil and GasPlating Machine Setters,Operators and Tenders,Metal and PlasticInspectors,Testers,Sorters,Samplers and WeighersRoustabouts,Oil and Gas01020304050607080901000102030405060708

188、090100General and Operations ManagersSecretaries and Administrative Assistants,except Legal,Medical and ExecutiveAccountants and AuditorsIndustrial Machinery MechanicsFirst-Line Supervisors of Construction Trades and Extraction WorkersBookkeeping,Accounting and Auditing ClerksEnergy and Materials:Oi

189、l and gasFinancial Services:Financial services and capital marketsPetroleum EngineersPetroleum Pump System Operators,Refinery Operators and GaugersRotary Drill Operators,Oil and GasWellhead PumpersLoan Interviewers and ClerksPersonal Financial AdvisersSecurities,Commodities and Financial Services Sa

190、les AgentsFirst-Line Supervisors of Office and Administrative Support WorkersLoan OfficersTellersFinancial ManagersSales Representatives,Wholesale and Manufacturing,except Technical and Scientific ProductsCustomer Service RepresentativesGeneral and Operations Managers01020304050607080901000102030405

191、060708090100Jobs of Tomorrow:Large Language Models and Jobs2201020304050607080901000102030405060708090100Financial Services:Insurance and pensions management01020304050607080901000102030405060708090100Insurance Claims and Policy Processing ClerksInsurance Sales AgentsManagement AnalystsFirst-Line Su

192、pervisors of Office and Administrative Support WorkersSoftware DevelopersInsurance UnderwritersGovernment and Public Sector:Government and public sectorCustomer Service RepresentativesGeneral and Operations ManagersOffice Clerks,GeneralClaims Adjusters,Examiners and InvestigatorsCourt,Municipal and

193、Licencte ClerksOffice Clerks,GeneralSecretaries and Administrative Assistants,except Legal,Medical and ExecutivePostal Service Mail CarriersPolice and Sheriffs Patrol OfficersManagement AnalystsRegistered NursesMaintenance and Repair Workers,GeneralFirefightersCorrectional Officers and JailersMedica

194、l Secretaries and Administrative AssistantsMedical and Health Services ManagersClinical Laboratory Technologists and TechniciansNurse PractitionersRegistered NursesReceptionists and Information ClerksHealth and Healthcare:Medical and healthcare servicesInformation Technology and Digital Communicatio

195、ns:Information and technology servicesMedical AssistantsNursing AssistantsDental AssistantsDental HygienistsSoftware Quality Assurance Analysts and TestersComputer User Support SpecialistsSales Representatives of Services,except Advertising,Insurance,Financial Services and TravelSoftware DevelopersM

196、arket Research Analysts and Marketing SpecialistsComputer Systems AnalystsCustomer Service RepresentativesGeneral and Operations ManagersProject Management SpecialistsComputer and Information Systems ManagersAutomationAugmentationLower potentialNon-language tasksJobs of Tomorrow:Large Language Model

197、s and Jobs230102030405060708090100010203040506070809010001020304050607080901000102030405060708090100Computer User Support SpecialistsSales Representatives of Services,except Advertising,Insurance,Financial Services and TravelSoftware DevelopersComputer Network ArchitectsCustomer Service Representati

198、vesComputer Network Support SpecialistsInformation Technology and Digital Communications:TelecommunicationsInfrastructure:Engineering and constructionElectronics Engineers,except ComputerTelecommunications Equipment Installers and Repairers,except Line InstallersFirst-Line Supervisors of Mechanics,I

199、nstallers and RepairersTelecommunications Line Installers and RepairersGeneral and Operations ManagersOffice Clerks,GeneralFirst-Line Supervisors of Construction Trades and Extraction WorkersOperating Engineers and Other Construction Equipment OperatorsElectriciansRetail SalespersonsHeating,Air Cond

200、itioning and Refrigeration Mechanics and InstallersPlumbers,Pipefitters and SteamfittersCarpentersOffice Clerks,GeneralFirst-Line Supervisors of Construction Trades and Extraction WorkersHeavy and Tractor-Trailer Truck DriversRefuse and Recyclable Material CollectorsSeptic Tank Servicers and Sewer P

201、ipe CleanersGeneral and Operations ManagersInfrastructure:Water and waste managementManufacturing:Advanced manufacturingHazardous Materials Removal WorkersWater and Wastewater Treatment Plant and System OperatorsLabourers and Freight,Stock and Material Movers,HandBus and Truck Mechanics and Diesel E

202、ngine SpecialistsIndustrial EngineersMechanical EngineersIndustrial Machinery MechanicsFirst-Line Supervisors of Production and Operating WorkersMiscellaneous Assemblers and FabricatorsGeneral and Operations ManagersMachinistsElectrical,Electronic,and Electromechanical Assemblers,except Coil Winders

203、,Tapers and FinishersInspectors,Testers,Sorters,Samplers and WeighersWelders,Cutters,Solderers and BrazersAutomationAugmentationLower potentialNon-language tasksConstruction LabourersJobs of Tomorrow:Large Language Models and Jobs2401020304050607080901000102030405060708090100010203040506070809010001

204、02030405060708090100General and Operations ManagersIndustrial EngineersElectrical and Electronic Engineering Technologists and TechniciansElectrical EngineersFirst-Line Supervisors of Production and Operating WorkersSoftware DevelopersInfrastructure:Water and waste managementManufacturing:Advanced m

205、anufacturingMiscellaneous Assemblers and FabricatorsSemiconductor Processing TechniciansElectrical,Electronic,and Electromechanical Assemblers,except Coil Winders,Tapers and FinishersInspectors,Testers,Sorters,Samplers and WeighersGeneral and Operations ManagersFirst-Line Supervisors of Production a

206、nd Operating WorkersMiscellaneous Assemblers and FabricatorsFood BatchmakersInspectors,Testers,Sorters,Samplers and WeighersSales Representatives,Wholesale and Manufacturing,except Technical and Scientific ProductsPackaging and Filling Machine Operators and TendersLabourers and Freight,Stock and Mat

207、erial Movers,HandMoulding,Coremaking and Casting Machine Setters,Operators and Tenders,Metal and PlasticIndustrial Truck and Tractor OperatorsAutomationAugmentationLower potentialNon-language tasksCashiersGeneral and Operations ManagersAmusement and Recreation AttendantsProducers and DirectorsUshers

208、,Lobby Attendants and Ticket TakersReceptionists and Information ClerksMedia,Entertainment and Sports:Arts,entertainment,and recreationMedia,Entertainment and Sports:Media and publishingWaiting StaffCoaches and ScoutsExercise Trainers and Group Fitness InstructorsLandscaping and Groundskeeping Worke

209、rsAdvertising Sales AgentsSales Representatives of Services,except Advertising,Insurance,Financial Services and TravelEditorsSoftware DevelopersMarket Research Analysts and Marketing SpecialistsGraphic DesignersCustomer Service RepresentativesGeneral and Operations ManagersProducers and DirectorsPri

210、nting Press OperatorsJobs of Tomorrow:Large Language Models and Jobs250102030405060708090100010203040506070809010001020304050607080901000102030405060708090100AutomationAugmentationLower potentialNon-language tasksGeneral and Operations ManagersOffice Clerks,GeneralSecretaries and Administrative Assi

211、stants,except Legal,Medical and ExecutivePublic Relations SpecialistsRecreation WorkersBookkeeping,Accounting and Auditing ClerksNon-governmental and Membership Organizations:Non-profit organizations,professional bodies and unionsProfessional Services:Business support and premises maintenance servic

212、esLabour Relations SpecialistsBartendersMaintenance and Repair Workers,GeneralClergyCustomer Service RepresentativesGeneral and Operations ManagersOffice Clerks,GeneralPest Control WorkersMaids and Housekeeping CleanersSales Representatives of Services,except Advertising,Insurance,Financial Services

213、 and TravelFirst-Line Supervisors of Landscaping,Lawn Service and Groundskeeping WorkersSecurity GuardsLandscaping and Groundskeeping WorkersJanitors and Cleaners,except Maids and Housekeeping CleanersCustomer Service RepresentativesOffice Clerks,GeneralStockers and Order FillersRegistered NursesMis

214、cellaneous Assemblers and FabricatorsHuman Resources SpecialistsProfessional Services:Employment servicesProfessional Services:Research,design and business management servicesPackers and Packagers,HandLabourers and Freight,Stock and Material Movers,HandIndustrial Truck and Tractor OperatorsJanitors

215、and Cleaners,except Maids and Housekeeping CleanersManagement AnalystsSoftware DevelopersMarket Research Analysts and Marketing SpecialistsCustomer Service RepresentativesGeneral and Operations ManagersBookkeeping,Accounting and Auditing ClerksProject Management SpecialistsLawyersAccountants and Aud

216、itorsParalegals and Legal AssistantsJobs of Tomorrow:Large Language Models and Jobs26010203040506070809010001020304050607080901000102030405060708090100AutomationAugmentationLower potentialNon-language tasksBookkeeping,Accounting and Auditing ClerksReal Estate Sales AgentsGeneral and Operations Manag

217、ersOffice Clerks,GeneralSecretaries and Administrative Assistants,except Legal,Medical and ExecutiveCounter and Rental ClerksReal Estate:Real estateRetail and Wholesale of Consumer Goods:Retail and wholesale of consumer goodsReal Estate BrokersProperty,Real Estate and Community Association ManagersM

218、aintenance and Repair Workers,GeneralJanitors and Cleaners,except Maids and Housekeeping CleanersCashiersRetail SalespersonsCustomer Service RepresentativesGeneral and Operations ManagersFirst-Line Supervisors of Retail Sales WorkersSales Representatives,Wholesale and Manufacturing,except Technical

219、and Scientific ProductsShipping,Receiving and Inventory ClerksHeavy and Tractor-Trailer Truck DriversStockers and Order FillersLabourers and Freight,Stock and Material Movers,HandLight Truck DriversBus Drivers,SchoolHeavy and Tractor-Trailer Truck DriversFirst-Line Supervisors of Transport and Mater

220、ial Moving Workers,except Aircraft Cargo Handling SupervisorsFlight AttendantsGeneral and Operations ManagersSupply Chain and Transport:Supply chain and transportStockers and Order FillersPackers and Packagers,HandLabourers and Freight,Stock and Material Movers,HandIndustrial Truck and Tractor Opera

221、torsJobs of Tomorrow:Large Language Models and Jobs27A2 Exposure potential by function groupsJob exposure by functional area:ranked by exposure(augmentation and automation potential)FIGURE 901020304050607080901000102030405060708090100AutomationAugmentationLower potentialNon-language tasksPayroll and

222、 Timekeeping ClerksHuman ResourcesIT/TechnologyCustomer SalesFinanceTraining and Development SpecialistsHuman Resources SpecialistsHuman Resources Assistants,except Payroll and TimekeepingCompliance OfficersCompensation,Benefits and Job Analysis SpecialistsTraining and Development ManagersCompensati

223、on and Benefits ManagersLabour Relations SpecialistsHuman Resources ManagersTravel AgentsTelemarketersCounter and Rental ClerksInsurance Sales AgentsSales EngineersSales Representatives,Wholesale and Manufacturing,Technical and Scientific ProductsSecurities,Commodities and Financial Services Sales A

224、gentsSales Representatives of Services,except Advertising,Insurance,Financial Services and TravelAdvertising Sales AgentsDemonstrators and Product PromotersTellersPersonal Financial AdvisersBookkeeping,Accounting and Auditing ClerksBilling and Posting ClerksFinancial ManagersFinancial and Investment

225、 AnalystsFinancial ExaminersFinancial Risk SpecialistsTax PreparersBill and Account CollectorsDatabase ArchitectsComputer Network Support SpecialistsDatabase AdministratorsWeb DevelopersComputer Systems AnalystsSoftware Quality Assurance Analysts and TestersComputer User Support SpecialistsNetwork a

226、nd Computer Systems AdministratorsData ScientistsComputer Programmers01020304050607080901000102030405060708090100Jobs of Tomorrow:Large Language Models and Jobs280102030405060708090100010203040506070809010001020304050607080901000102030405060708090100AutomationAugmentationLower potentialNon-language

227、tasksProduction,Planning and Expediting ClerksLogisticiansCargo and Freight AgentsShipping,Receiving and Inventory ClerksPurchasing ManagersProcurement ClerksSupply ChainLegalBuyers and Purchasing AgentsIndustrial Production ManagersTransport,Storage and Distribution ManagersDriver/Sales WorkersJudi

228、cial Law ClerksAdministrative Law Judges,Adjudicators and Hearing OfficersCompliance OfficersJudges,Magistrate Judges and MagistratesLawyersLegal Secretaries and Administrative AssistantsArbitrators,Mediators and ConciliatorsTitle Examiners,Abstractors and SearchersParalegals and Legal AssistantsFun

229、draising ManagersMarket Research Analysts and Marketing SpecialistsPublic Relations SpecialistsPublic Relations ManagersAdvertising and Promotions ManagersAdvertising Sales AgentsMarketingOperationsSales ManagersMarketing ManagersProperty,Real Estate and Community Association ManagersFirst-Line Supe

230、rvisors of Office and Administrative Support WorkersProcurement ClerksExecutive Secretaries and Executive Administrative AssistantsReceptionists and Information ClerksNew Accounts ClerksCorrespondence ClerksTelephone OperatorsGeneral and Operations ManagersOffice Clerks,GeneralFile ClerksJobs of Tom

231、orrow:Large Language Models and Jobs29A3 MethodologyOccupational analysis data comes from O*NET and the United States Bureau of Labor Statistics(BLS).21 O*NET provides a taxonomy of over 19,000 tasks,a matching of which tasks are performed in which occupations,and a breakdown of the time spent on ea

232、ch task within an occupation.Task frequency was converted to the distribution of time spent on each task,by occupation,and calibrated to 100%.Both machine learning and manual methods were used to classify each of the 19,000 tasks according to their potential for automation,represented by a numerical

233、 score.As a first step,researchers assess which tasks require intensive use of language(natural,mathematical,computational)that would be relevant to large language models(LLMs).A binary value of 1 for yes and 0 for no is assigned to each task.As a second step,if the task was language-based,three rem

234、aining criteria determine the level of human involvement needed:Human-to-human interaction(vs human-to-computer):If the task requires real-time human exchange in the same physical space or virtually,a binary value of 1 for yes and 0 for no is assigned to the task.The task is non-routine and/or non-w

235、ell-defined:If the task requires a proactive effort to process complex,unstructured and unrelated information to solve for the task,i.e.non-routine or unstructured,a binary value of 1 for yes and 0 for no is assigned to the task.Human involvement enforced by law,ethics or social conventions:If the t

236、ask requires some level of commitment or compliance,e.g.regarding legal,financial,health or other types of risks,a binary value of 1 for yes and 0 for no is assigned to the task.A binary value is assigned for each of the four criteria,once by humans and once by a machine learning model powered by GP

237、T-4.These two scores are then added together for each criterion and task.If both humans and AI tag a task as non-language-based,then the task is considered non-language.If the task is tagged as language-based by either method,then the remaining three criteria scores are summed to determine the class

238、ification of exposure as follows:0:Higher potential for automation 1-2:Higher potential for augmentation 3-6:Lower potential for automation and augmentationTherefore,each task is classified into one of four final groups:high potential for automation,high potential for augmentation,low potential for

239、automation or augmentation and non-language tasks.Researchers initialized GPT-4 with representative examples of manually tagged tasks and re-trained the model on initially incorrect tags.Any remaining misalignment between human and GPT tags was resolved by favouring lower impact scores to err on the

240、 conservative side of estimates.Combining the task score,BLS data on occupational employment and hours worked,and data from O*NET on task frequency by occupation,the proportion of work time that LLMs could transform can then be determined.Occupations with a high potential for transformation are defi

241、ned as those with greater than 50%of work time in tasks with a high potential for automation and/or augmentation.Estimates are then aggregated to industry sectors and function groups weighed by the employed population in the US.Jobs of Tomorrow:Large Language Models and Jobs30Methodology overviewFIG

242、URE 10Modelling analysis is conducted in four stepsStep 1Step 2Step 3Step 4Data scope19,265 tasksTask coverage900 jobs(varying granularity for other countries in progress)Jobs coverageUS(2022)(and 10+geographies with latest dates in progress)Geographic coverageIdentify language taskstasks19,265langu

243、age tasks9,934high transformation potential tasks4,646augmentable and automatable tasks3,2111,435Assess knowledge use on taskLabel task transformation potentialRoll up tasks to job and industry level Task tagging overviewFIGURE 11Three steps to identify the automation/augmentation potentialStep 1Ste

244、p 2Step 3Language tasks identificationAssess knowledge use on taskLabel task transformation potentialPrompt:Do the following tasksLanguage Require a certain level of language ability(natural,mathematical,computational)?Knowledge Does the task require a certain level of knowledge skill(1)to solve cre

245、atively ambiguous problems?(2)to work together with others in real time?(3)to validate outputs with small and medium-sized enterprises?NoYesNoAutomation Generative AI transformation potential labelsAugmentationLower PotentialNon-languageYesTo just oneTo at least twoJobs of Tomorrow:Large Language Mo

246、dels and Jobs31ContributorsProductionWorld Economic Forum Attilio Di BattistaHead,Insights CollaborationElselot HasselaarHead,Mission for Work,Wages and Job CreationAndrew SilvaInsights Lead,Work,Wages and Job CreationSaadia ZahidiManaging DirectorAccentureTomas CastagninoPrincipal Director,Accentur

247、e ResearchNicole DAgostinoResearch Manager,Accenture ResearchNathan DecetyConsultant,Accenture StrategyHernan EspinosaSenior Manager,Accenture StrategyAllison HornTalent and Organization/Human Potential Managing DirectorMary Kate Morley RyanTalent and Organization/Human Potential Managing DirectorCh

248、ristine NananSenior Principal,Accenture StrategyKathleen OReillyAccenture Industry Practices Chair,Communications,Media and Technology Leila YosefSenior Principal,Accenture ResearchWe are further grateful to our colleagues at the Centre for the New Economy and Society for helpful suggestions and com

249、ments,in particular the authors of the Future of Jobs Report 2023,Sam Grayling,Ricky Li and Mark Rayner.Accenture extends thanks to many of its colleagues for providing insights and guidance that assisted in shaping this report,including Francis Hintermann,Keegan Moore,Macarena Ortt,Deeksha Khare Pa

250、tnaik,Rahul Raichura,Michael J.Robinson,Tal Roded,Luciano Rossi,Padampreet Singh,Paul Walsh,and H.James Wilson.For more information,or to get involved,please contact cnesweforum.org.Studio MikoLaurence Denmark Creative DirectorMartha Howlett EditorJames Turner DesignerOliver Turner DesignerJobs of T

251、omorrow:Large Language Models and Jobs32Endnotes1.Daugherty,P.,B.Ghosh,K.Narain,L.Guan and J.Wilson,A new era of generative AI for everyone,Accenture,2023,https:/ Economic Forum,The Future of Jobs Report 2023,2023,https:/www.weforum.org/reports/the-future-of-jobs-report-2023/.3.World Economic Forum,

252、The Future of Jobs Report 2023,2023,pp.25,https:/www.weforum.org/reports/the-future-of-jobs-report-2023/.4.Hu,Krystal,“ChatGPT sets record for fastest-growing user base-analyst note”,Reuters,2 February 2023 https:/ Faces a Serious Threat From ChatGPT”,The Washington Post,7 December 2022,https:/ and

253、J.Wilson,A new era of generative AI for everyone,Accenture,2023,https:/ and N.L.Ziebarth,“The history of technological anxiety and the future of economic growth:Is this time different?”,Journal of economic perspectives,vol.29,no.3,2015,pp.31-50.7.For example,see blog entry:Larsen,Benjamin and Jayant

254、 Narayan,“Generative AI:a game-changer that society and industry need to be ready for”,World Economic Forum,9 January 2023,https:/www.weforum.org/agenda/2023/01/davos23-generative-ai-a-game-changer-industries-and-society-code-developers/.8.However,to be used with caution,making sure to verify source

255、s.9.Peng,S.,E.Kalliamvakou,P.Cihon and M.Demirer,The Impact of AI on Developer Productivity:Evidence from Github Copilot,arXiv,2023.10.“Teachers and Students Embrace ChatGPT for Education”,The Walton Family Foundation,1 March 2023,https:/www.waltonfamilyfoundation.org/learning/teachers-and-students-

256、embrace-chatgpt-for-education.11.“Occupation Employment and Wage Statistics”,Bureau of Labor Statistics,25 April 2023,https:/www.bls.gov/oes/current/oes_stru.htm.12.Daugherty,Wilson,Narain,“Generative AI Will Enhance Not Erase Customer Service Jobs”,Harvard Business Review,30 March 2023,https:/hbr.o

257、rg/2023/03/generative-ai-will-enhance-not-erase-customer-service-jobs.13.Brynjolfsson,Erik,Danielle Li and Lindsey R.Raymond,Generative AI at Work,National Bureau of Economic Research,2023,https:/www.nber.org/papers/w31161.14.World Economic Forum,The Future of Jobs Report 2023,2023,pp.46,https:/www.

258、weforum.org/reports/the-future-of-jobs-report-2023/.15.Using the World Economic Forums industry classification.16.World Economic Forum,The Future of Jobs Report 2023,2023,pp.27,https:/www.weforum.org/reports/the-future-of-jobs-report-2023/.17.World Economic Forum,The Future of Jobs Report 2023,2023,

259、pp.30,https:/www.weforum.org/reports/the-future-of-jobs-report-2023/.18.Ibid.19.Ibid.20.World Economic Forum,Jobs of Tomorrow:The Triple Returns of Socials Jobs in the Economic Recovery,2022,pp.8-13.21.“Task Statements”,O*NET,n.d.,https:/www.onetcenter.org/dictionary/20.1/excel/task_statements.html;

260、“Standard Occupational Classification”,Bureau of Labor Statistics,2018,https:/www.bls.gov/SOC/.Jobs of Tomorrow:Large Language Models and Jobs33World Economic Forum9193 route de la CapiteCH-1223 Cologny/GenevaSwitzerland Tel.:+41(0)22 869 1212Fax:+41(0)22 786 2744contactweforum.orgwww.weforum.orgThe World Economic Forum,committed to improving the state of the world,is the International Organization for Public-Private Cooperation.The Forum engages the foremost political,business and other leaders of society to shape global,regional and industry agendas.