Source-linked AI summary
How will Language Modelers like ChatGPT Affect Occupations and Industries?
Ed Felten, Manav Raj, Robert Seamans
TL;DR
The paper asks how advancing AI language modeling may affect occupations and industries, an important question amid uncertainty about AI’s economic effects. It adapts the Felten et al. exposure methodology to language modeling and finds especially high exposure among telemarketers, post-secondary teachers, legal services, and investment-related industries, with higher-wage occupations more exposed. The measure captures exposure but does not determine whether effects involve substitution or augmentation.
Problem
The paper examines how advancing AI language modeling may affect occupations, industries, and geographies amid uncertainty about its economic effects.
Method
The paper adapts Felten et al.’s AI Occupational Exposure methodology to assess exposure to advances in AI language modeling.
Results
Telemarketers, several post-secondary teaching occupations, legal services, and securities, commodities, and investments are among the most exposed; higher-wage occupations are more likely to be exposed.
Takeaways & Limitations
The methodology provides a systematic way to assess language-modeling exposure across occupations and industries as AI capabilities advance.
Takeaways & Limitations
AIOE measures exposure to AI, but whether exposure leads to augmentation or substitution depends on the specifics of the occupation.
Abstract
from arXiv · showhide
Recent dramatic increases in AI language modeling capabilities has led to many questions about the effect of these technologies on the economy. In this paper we present a methodology to systematically assess the extent to which occupations, industries and geographies are exposed to advances in AI language modeling capabilities. We find that the top occupations exposed to language modeling include telemarketers and a variety of post-secondary teachers such as English language and literature, foreign language and literature, and history teachers. We find the top industries exposed to advances in language modeling are legal services and securities, commodities, and investments. We also find a positive correlation between wages and exposure to AI language modeling.
1. Introduction
The paper addresses uncertainty about how advancing AI language modeling may affect work by adapting a systematic exposure methodology across occupations, industries, and geographies. It identifies exposed occupations and industries and examines how exposure relates to wages.
- Motivation: Advancing AI capabilities make understanding effects on work difficult, while language modeling raises both job-displacement concerns and practical or commercial expectations.The paper frames AI’s effects as potentially involving substitution or complementarity.
- Approach: The paper adapts Felten et al.’s AI Occupational Exposure measure to study language-modeling exposure across occupations, industries, and geographies.The approach is presented as a systematic examination of language modeling’s economic effects.
- Findings: Telemarketers and several post-secondary teaching occupations, including English, foreign-language, and history teachers, rank among the occupations most exposed.These findings extend the emerging literature on ChatGPT and language modelers’ effects on the economy.
- Findings: Legal services and securities, commodities, and investments are among the industries most exposed to advances in language modeling.The paper presents this as part of its systematic cross-industry assessment.
- Findings: The paper finds a positive and statistically significant relationship between occupational mean or median wages and language-modeling exposure.The introduction motivates examining occupational wages alongside exposure.
- Contribution: The study demonstrates that Felten et al.’s methodology can be adjusted dynamically as AI capabilities change.This positions the method as flexible for studying evolving AI technologies.
2. AI Occupational Exposure Methodology
The AI Occupational Exposure measure links AI applications to occupational abilities and weights those abilities by their prevalence and importance within occupations. It can then be aggregated to industry and geographic levels.
- Measure construction: AIOE measures each occupation’s exposure to AI by linking 10 AI applications with 52 occupational abilities.The links come from a crowd-sourced matrix of application–ability relatedness scores.
- Interpretation: The measure uses “exposure” without determining whether AI will substitute for or augment work in a given occupation.Those effects depend on the specifics of the occupation.
- Measure construction: The 52 abilities describe more than 800 occupations, each represented as a weighted combination using prevalence and importance weights from O*NET.O*NET is developed by the United States Department of Labor.
- Measure construction: Ability-level exposure is calculated by summing application–ability relatedness scores before constructing an occupation-level AIOE score.The equation indexes AI applications by i and occupational abilities by j.
- Measure construction: Occupation-level AIOE weights ability-level exposure by each ability’s prevalence and importance within the occupation, with equal scaling of those weights.The occupation is indexed by k.
- Extensions: AIOE scores can be weighted to construct AI Industry Exposure and AI Geographic Exposure measures.Felten et al. also describe validation exercises and potential scholarly and practitioner uses.
3. Language Modeling AI Occupational Exposure
The paper modifies AIOE so that only language modeling contributes to ability-level exposure, then computes occupation-level language-modeling exposure. The resulting scores closely track the original AIOE scores.
- Adjustment: The language-modeling adjustment modifies the original AIOE calculation by assigning zero weight to every AI application except language modeling.Language modeling retains a weight of 1.
- Adjustment: The adjusted ability-level exposure therefore counts only application–ability relatedness associated with language modeling.The relatedness scores are constructed from crowd-sourced survey data.
- Occupation scores: The paper computes the resulting language-modeling AIOE score for each occupation using the modified ability-level exposure.The score captures each occupation’s exposure to advances in language modeling due to AI.
- Validation: 0.979 is the correlation coefficient between the resulting language-modeling-adjusted scores and the original AIOE scores.Figure 1 plots both scores for each occupation.
4. Results
The section ranks occupations and industries by exposure to advances in AI language modeling, highlighting telemarketing, post-secondary education, finance, legal services, and higher education-related industries.
- Overview: The language-modeling-focused exposure analysis compares top-20 occupations and industries with rankings based on the original Felten et al. measures.The section presents separate lists for occupations and industries exposed to AI-enabled advances in language-modeling capabilities.
- Top occupations: Language modeling exposure includes many post-secondary education occupations, including English, foreign-language, history, law, philosophy, sociology, and political science teachers.The language-modeling list contains more education-related occupations than the original comparison list.
- Top occupations: Telemarketers rank first among occupations exposed to language modeling.The AIOE ranking places telemarketers above English language and literature teachers, foreign language and literature teachers, and history teachers.
- Top occupations: Telemarketer exposure could support either human augmentation or substitution by language-modeling-enabled bots.The section gives real-time customer-specific prompts as an augmentation example and bots replacing human telemarketers as a substitution example.
- Top industries: Securities, commodity contracts, and other financial investments rank highly in both exposure measures, while legal services and insurance-related activities also appear among the most exposed industries.The language-modeling-focused ranking places the securities, commodities, and investments industry second, and identifies legal services among the top industries.
- Top industries: Junior colleges, grantmaking and giving services, business schools, and computer and management training appear among the top twenty language-modeling-exposed industries.These results indicate comparatively higher exposure within higher education and adjacent industries.
5. Relationship to Wages
The paper examines whether occupational wages vary with language-modeling exposure by matching AIOE data to 2021 BLS wages and grouping occupations by exposure score. Mean- and median-wage figures show a strong positive correlation: higher-wage occupations are more likely to be exposed.
- Data and grouping: The analysis matches AIOE occupations to 2021 BLS wage data using occupational codes and then titles, yielding 708 occupations in both datasets.The two-stage matching addresses changes in occupational definitions over time.
- Data and grouping: Occupations are grouped into 20 equal-sized bins by language-modeling AIOE score, and each bin’s average exposure is compared with mean and median wages.The resulting comparisons are plotted in Figure 2 for mean wages and Figure 3 for median wages.
- Results: A strong positive correlation appears between language-modeling AIOE scores and both mean and median occupational wages.The results are consistent across the mean-wage and median-wage figures.
- Results: Higher-wage occupations are more likely to be exposed to rapid advances in language modeling from products such as ChatGPT.The paper notes that the most exposed occupations appear to be white-collar occupations that may be classified as high-skilled labor.
6. Conclusion
The paper adapts an existing occupational-exposure approach to assess how advances in AI language modeling may affect occupations, industries, and geographies. It identifies highly exposed occupations and industries and reports that higher-wage occupations are more likely to be exposed.
- Contribution: The paper presents a systematic methodology for assessing exposure of occupations, industries, and geographies to advances in AI language modeling.The methodology is intended to help understand how ChatGPT and other language modelers affect these economic units.
- Contribution: The methodology relies on Felten et al. (2021) but adapts that approach to account for recent advances in language modeling.The paper describes this adaptation as the basis for its exposure assessment.
- Results: Telemarketers and several post-secondary teacher occupations are among the occupations most exposed to language modeling.Examples include English language and literature, foreign language and literature, and history teachers.
- Results: Legal services and securities, commodities, and investments are among the industries most exposed to advances in language modeling.These are the paper’s highlighted examples of highly exposed industries.
- Results: Occupations with higher wages are more likely to be exposed to rapid advances in language modeling.The conclusion reports this relationship as a finding of the paper.
- Implications: The results are intended to be useful to scholars, practitioners, and policymakers.The paper frames this usefulness as a consequence of its reported results and methodology.
SORTED BY LANGUAGE MODELING EXPOSURE SCORE
The table orders occupations by their language-modeling AIOE exposure score, with telemarketers and several post-secondary teaching occupations at the top and lower-scoring occupations listed later.
- Highest exposure: Telemarketers have the highest listed language-modeling exposure score, with an AIOE score of 1.926.English language and literature teachers, postsecondary follow with 1.857; foreign language and literature teachers, postsecondary with 1.814; history teachers, postsecondary with 1.813.
- Highest exposure: The highest-ranked occupations include telemarketers and numerous post-secondary teachers in language, history, law, philosophy, sociology, and political science.The table’s first entries are dominated by telemarketing and post-secondary teaching occupations.
- Middle ranks: The table includes mid-ranked occupations such as public relations specialists, judicial law clerks, purchasing agents, historians, and financial analysts.These occupations appear across the listed ranks from 34 through 102 with exposure scores ranging from 1.518 to 1.273.
- Lower exposure: Lower-ranked entries include computer network architects, network and systems administrators, microbiologists, anesthesiologists, and retail-sales supervisors.The listed scores for these occupations range from 0.495 to 0.201.