GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models

1 Introduction

• The authors assess U.S. occupational tasks with a new exposure rubric applied by human annotators and GPT-4, treating exposure as potential impact rather than a prediction of productivity or automation.The primary occupational data come from O*NET; technical feasibility does not guarantee labor productivity or automation outcomes.Source: §1 · p. 2
• Higher-wage occupations generally show greater exposure; programming and writing skills are positively associated with exposure, while science and critical-thinking skills correlate negatively, and exposure weakly rises with job-preparation barriers.The wage pattern contrasts with similar evaluations of overall machine-learning exposure.Source: §1 · p. 3
• Earlier technology measures and wage controls explain 60–72% of variation in the preferred exposure measure, while manual-routineness and robotics measures correlate negatively and 28–40% remains unexplained.Most other technology exposure measures examined are statistically significantly correlated with the preferred measure.Source: §1 · p. 3
• Information-processing industries show high exposure, while manufacturing, agriculture, and mining show lower exposure; the relationship between past-decade productivity growth and LLM exposure appears weak.The authors suggest this may offer an optimistic case regarding future productivity gains and Baumol’s cost-disease effects.Source: §1 · p. 3
• The authors characterize GPTs as general-purpose technologies, arguing that their potential expands with complementary technologies and may reach broad economically valuable uses, including creating new types of work.Their analysis measures what is technically feasible now and will not capture LLM impact potential as it evolves.Source: §1 · pp. 3–4

2.1 The Advancement of Large Language Models

• LLMs have advanced into broad applications such as translation, classification, writing, and code generation, while emerging tool-control abilities could extend their reach across computer-based tasks.Progress has been associated with larger models and training datasets and improved training configurations (Brown et al., 2020; Radford et al., 2019; Hernandez et al., 2021; Kaplan et al., 2020); fine-tuning and human feedback improve steerability and utility (Ouyang et al., 2022; Bai et al., 2022).Source: §2.1 · p. 4
• Although LLMs are often deployed as task-specific specialists, integrated workflows can help address their reliability, bias, privacy, and disinformation risks while enabling adoption in writing, coding, and legal research.Domain expertise and human review can complement models; examples include Casetext legal research tools and GitHub Copilot (Abid et al., 2021; Schramowski et al., 2022; Goldstein et al., 2023).Source: §2.1 · p. 4
• LLMs may help create tools and support machine-learning development, potentially lowering tooling costs and accelerating adoption, while also presenting serious risks (Chen et al., 2021; Peng et al., 2023; Khlaaf et al., 2022; Solaiman et al., 2019).Potential roles include coding assistance, data labeling, synthetic data generation, and task allocation between people and machines (Singla et al., 2015).Source: §2.1 · p. 5

2.2 The Economic Impacts of Automation Technologies

• Task-based labor research links routine-task automation to displacement and wage inequality, while prior approaches map AI capabilities to worker tasks and often find exposure diversified within occupations (Katz and Murphy, 1992; Autor et al., 2003; Acemoglu and Autor, 2011b; Acemoglu and Restrepo, 2018; Autor et al., 2006; Van Reenen, 2011; Acemoglu and Restrepo, 2022b; Webb, 2020; Meindl et al., 2021; Felten et al., 2018, 2023; Tolan et al., 2021; Frey and Osborne, 2017; Arntz et al., 2017; Grace et al., 2018; Brynjolfsson et al., 2018, 2023; Autor et al., 2022a).Autor et al. (2022a) also finds automation and augmentation exposures tend to be positively correlated.Source: §2.2 · p. 5
• General-purpose technologies depend on widespread adoption, continuous improvement, complementary innovations, and costly, time-consuming organizational co-invention.The paper uses task-level information to assess whether LLMs meet these criteria and connect technology-labor perspectives.Source: §2.2 · pp. 5–6
• The paper focuses on LLM impacts, introduces GPT-4-assisted task exposure and automation scoring, and aggregates results to occupations and industries in the contemporary U.S. labor market.The method is intended to bolster human scoring efforts.Source: §2.2 · p. 6

3.1 Data on Activities and Tasks Performed by Occupation in the US

• The study uses O*NET 27.2 data on 1,016 occupations, comprising 19,265 occupation-linked tasks and 2,087 Detailed Work Activities (DWAs).Tasks and DWAs may map to multiple counterparts, while some tasks have no associated DWA.Source: §3.1 · p. 6

3.2 Data on Wages, Employment, and Demographics

• BLS employment, wage, education, training, and demographic data are linked to O*NET using a recommended crosswalk; these sources primarily cover the formal economy’s non-self-employed, documented workers.The employment and wage data come from the 2020 and 2021 Occupational Employment series, with employment projections for 2031.Source: §3.2 · p. 6

3.3 Exposure

• Exposure requires at least a 50% task-time reduction at equivalent quality; E1 denotes direct LLM exposure, while E2 captures potential gains from complementary software.The rubric also included image-generation capabilities under software exposure, and DWA labels were aggregated to tasks before occupation-level analysis.Source: §3.3 · p. 7
• Human annotators labeled DWAs and a subset of tasks, while an early GPT-4 version labeled all task/occupation pairs using a slightly modified rubric.The authors and experienced annotators contributed human labels; human annotations are the study’s primary results.Source: §3.3 · p. 8, §3.4.2 · p. 9
• The study defines α as E1 exposure, β as E1 plus half of E2, and ζ as E1 plus E2, representing lower, intermediate, and upper exposure measures.The half-weight on E2 accounts for the additional investment potentially needed to deploy complementary tools.Source: §3.3 · pp. 8–9

3.4.1 Subjective human judgments

• Human labels may be biased because annotators lacked broad occupational expertise, so reliable task-level judgments may require workers or experts familiar with each occupation.The authors identify validation of labels with occupational knowledge as important future work.Source: §3.4.1 · p. 9

3.4.2 Measuring LLMs with GPT-4

• GPT-4 classifications are sensitive to prompt and rubric wording, examples, and order, so neither GPT-4 nor human annotations are treated as definitive ground truth.The authors report high occupation-level agreement between the sources, while using human annotations as primary results.Source: §3.4.2 · p. 9, Figure 2 · p. 10

3.4.3 Additional Weaknesses

• The task-based framework may omit tacit skills or misrepresent linked subtasks, while annotators’ limited task context caused aggregation ambiguities and label discrepancies.The authors collected additional labels for task/occupation pairs with substantial disagreement; disagreements also involved adoption norms, required human judgment, or existing automation.Source: §3.4.3 · pp. 9–10

4.1 Summary Statistics

• Under β, 80% of workers occupy jobs with at least 10% exposed tasks, and 19% occupy jobs with over 50% exposed tasks.Mean occupation-level α exposure is 0.14–0.15; β exceeds 30% and ζ exceeds 50%. Using O*NET task-importance weights did not significantly change findings, though equal task weights produce curious rankings such as Barbers. Realizing potential effects may require broader systems and face co-invention and oversight barriers; the assessment also omits total factor productivity and capital input potential.Source: §4.1 · p. 11, §4 · p. 11, Table 3 · p. 12

4.2 Wages and Employment

• Human and GPT-4 ratings trend toward higher exposure in higher-wage occupations, while neither shows significant exposure differences across employment levels.Numerous low-wage occupations also have high exposure.Source: §4.2 · p. 12, Figure 4 · p. 13

4.3 Skill Importance

• Science and critical-thinking skill importance is strongly negatively associated with exposure, while programming and writing skill importance is strongly positively associated with exposure.The regression uses occupation-level exposure measures and O*NET Basic Skills, plus programming skill.Source: §4.3 · p. 14, Table 5 · p. 17

4.4 Barriers to Entry

• Greater-than-50% β exposure rises from 0.00% in Job Zone 1 to 34.5% in Job Zone 4, then falls to 26.45% in Job Zone 5.Across α, β, and ζ, exposure increases from Job Zone 1 to Zone 4 and then remains similar or decreases; median income rises from $30,230 in Zone 1 to $80,980 in Zone 5 as preparation requirements increase.Source: §4.4 · p. 14, Figure 5 · p. 15

4.4.1 Typical Education Needed for Entry

• Bachelor’s, Master’s, and professional degree holders, as well as people with some college but no degree, show greater exposure than those without formal educational credentials.Jobs requiring the most on-the-job training have the least exposure and potentially lower median income, while jobs requiring no training or only internship/residency appear higher-income and more exposed. The education and training summaries exclude 3,504,000 workers lacking those data.Source: §4.4.1 · p. 14, Table 7 · p. 18

5.1 Comparison to Earlier Efforts

• The paper builds primarily on the SML approach, using a rubric to assess overlap between LLM capabilities and O*NET worker tasks and comparing new scores with prior occupation-level measures.Regressions include annualized BLS occupational salaries as a control; core tasks receive full weight and supplemental tasks half weight.Source: §5.1 · p. 19, Table 9 · p. 21
• Webb robot exposure, routine manual task content, and Frey and Osborne’s overall automation measure are negatively correlated with primary GPT-4 and human overall exposure ratings, conditional on other measurements.The paper attributes this pattern to the limited exposure of physical tasks to LLMs, including LLMs with additional systems integration.Source: §5.1 · p. 20
• Regression R^2 ranges from 60.7% in column 3 to 72.8% in column 5, leaving 28–40% of the variation in the new measure unexplained by earlier measurements.The authors note that earlier efforts had limited information about future LLM and LLM-powered software progress, which their rubric also captures imperfectly.Source: §5.1 · p. 20, Table 9 · p. 21
• Software, SML, and routine cognitive scores are positively associated with LLM exposure at 1%; Webb AI scores are positive and significant at 5%, except for the secondary-prompt ratings in columns 3 and 4.These results compare the paper’s GPT-4 and human ratings with earlier occupation-level measures, controlling for the other measurements.Source: §5.1 · p. 20, Table 9 · p. 21

6.1 GPTs as a General-Purpose Technology

• LLMs show two general-purpose-technology traits: pervasive potential impacts across the economy and complementary innovations, especially software and digital tools.The discussion identifies improvement over time as the third criterion and notes that prior AI and machine-learning literature documents it.Source: §6.1 · p. 22
• 0.42 (GPT-4 annotations) and 0.32 (human annotations) are α–ζ mean differences, versus mean ζ exposure of 0.14 for both, suggesting tools may more than double direct impact.The difference between α and ζ represents exposure potential attributable to LLM-powered tools and software over direct exposure to LLMs alone.Source: §6.1 · p. 22
• ChatGPT’s interface launch was followed by skyrocketing usage, and commercial surveys indicated rising firm and worker adoption (Chow, 2023; OpenAI, 2022; Constantz, 2023; ResumeBuilder.com, 2023).The paper notes that it does not systematically analyze adoption.Source: §6.1 · p. 22
• Adoption depends on trust, costs, flexibility, preferences, incentives, ethical and safety risks, and sectoral differences in data, regulation, and power.For example, legal professionals’ use may depend on trusting model outputs without checking original documents or conducting independent research; risks include bias, fabricated facts, and misalignment.Source: §6.1 · p. 22
• One proposed trajectory is augmentation before automation, potentially making jobs more precarious before full automation.The discussion also suggests that time savings and seamless application may matter more than quality improvement for most tasks.Source: §6.1 · p. 22

6.2 Implications for US Public Policy

• The authors say worker exposure highlights the need for US preparedness for potential disruption, while leaving specific policy prescriptions outside the paper’s scope.They point to education, worker training, and safety-net reforms as policy directions discussed in prior work (Autor et al., 2022b).Source: §6.2 · p. 23

6.3 Limitations and Future Work

• The US focus limits generalizability because other nations differ in industrial organization, infrastructure, regulation, linguistic diversity, and culture.The authors hope to extend the study’s scope and share its methods so other researchers can build on them.Source: §6.3 · p. 23
• Future work should study adoption across sectors and occupations and test model capabilities against worker activities beyond exposure scores.The study excluded vision from direct-exposure ratings despite GPT-4’s multimodal advances (OpenAI, 2023b), and notes possible gaps between theoretical and practical performance on complex, open-ended, domain-specific tasks.Source: §6.3 · p. 23

7 Conclusion

• Approximately 19% of jobs have at least 50% of their tasks exposed to LLMs when current capabilities and anticipated LLM-powered software are considered.The conclusion also says higher-wage occupations generally present more tasks with high exposure.Source: §7 · p. 23
• LLMs may affect a wide range of U.S. occupations and economic activities, but social, economic, regulatory, and other factors will influence actual labor productivity outcomes.The authors note that policymakers may face challenges predicting and regulating LLMs’ trajectory as capabilities evolve.Source: §7 · p. 23
• Further research is needed on whether LLMs augment or displace labor and how they affect job quality, inequality, and skill development.The authors say understanding these effects can help policymakers and stakeholders make more informed decisions about AI and work.Source: §7 · pp. 23–24

7.1 LLM Conclusion (GPT-4’s Version)

• The GPT-4 conclusion describes a rubric for gauging GPT exposure across tasks, particularly in the U.S. labor market.It frames the study as probing GPTs’ potential trajectories.Source: §7.1 · p. 24

7.2 LLM Conclusion (Author-Augmented Version)

• The author-augmented version characterizes GPTs as “permeating tasks” and “gutting professional management,” and calls to “generalize past today.”These phrases appear in the version’s concluding text.Source: §7.2 · p. 24

A.1 Exposure

• The assumed LLM handles text-in/text-out tasks whose input context fits within 2,000 words, but cannot retrieve facts from the past year unless supplied in the input.Source: §A.1 · p. 24
• Workers are assumed to have average role expertise, access to the LLM and task-mentioned tools, and commonly available laptop tools—but no other physical tools or materials.Source: §A.1 · p. 24
• E0 tasks cannot be made at least twice as fast by direct LLM access at equivalent quality, while E1 tasks can; high-interaction tasks are E0.E1 includes text or code transformation, translation, document summarization, feedback, and question answering.Source: §A.1 · p. 25
• E2 applies when LLM-powered software could halve task time even if the LLM alone cannot, through capabilities such as long-document processing, current-web retrieval, or organizational search.Examples include tailored internal-data summaries, customer-service response suggestions, legal research, essay grading, and multilingual news summaries.Source: §A.1 · p. 25
• E3 covers tasks substantially sped up by LLMs with image viewing, captioning, or creation, including reading PDF text, scanning images, and editing digital images.The assumed system cannot take video or accurately retrieve fine image details such as dimensions.Source: §A.1 · p. 25
• Examples classify hands-on equipment repair as E0, text-based theoretical technology work as E1, and medication prescribing as E2 with a human making the final decision.Source: §A.1 · pp. 25–26

Basic Skills

• Basic skills are developed capacities that facilitate learning or the more rapid acquisition of knowledge.Source: Basic Skills · p. 26

Content

• Content skills are background structures needed to work with and acquire more specific skills across different domains.Source: Content · p. 26
• Reading comprehension means understanding work-related written material, while active listening involves attending, understanding, questioning appropriately, and avoiding interruptions.Source: Content · p. 26
• Writing communicates effectively for an audience, while speaking conveys information effectively to others.Source: Content · p. 26
• Mathematics uses mathematical methods to solve problems, while science uses scientific rules and methods for problem-solving.Source: Content · p. 26

Process

• Process skills are procedures that accelerate knowledge and skill acquisition across domains; critical thinking uses logic and reasoning to assess alternative solutions, conclusions, or approaches.Source: Process · p. 26
• Active learning considers how new information affects current and future problem-solving and decisions, while learning strategies select suitable methods for learning or teaching.Source: Process · p. 26
• Monitoring assesses one's own, others', or organizations' performance to support improvement or corrective action.Source: Process · p. 26

• Programming was the only cross-functional skill selected, based on the authors’ prior knowledge of the models’ ability to code.Source: Cross-Functional Skills · p. 27

C Industrial and Productivity Exposure

• LLM impact potential appears in nearly all industries, with wide heterogeneity in employment-weighted relative exposure.Industry exposure is assessed using the exposure rubric and shown for human raters and GPT-4.Source: C Industrial and Productivity Exposure · p. 27
• Human raters and GPT-4 generally agree on relative exposure, with data processing, information processing, and hospitals among the highly exposed industries.Source: C Industrial and Productivity Exposure · p. 27
• Recent total-factor and labor productivity growth since 2012 appears unrelated to current LLM exposure, offering little evidence that already fast-growing industries will become more productive.The passage frames this as a concern about Baumol’s cost disease, but reports little to suggest such an outcome.Source: C Industrial and Productivity Exposure · pp. 27, 30

D Occupations Without Any Exposed Tasks

• 34 occupations had no tasks labeled as exposed by any measure.Table 11 lists these occupations.Source: Table 11 · p. 31