Creation, validation, obsolescence: observed evidence of AI-driven labor market displacement, 2020–2025

Abstract

Background:

The successive releases of GPT-3 (May 2020) and ChatGPT (November 2022) have been widely hypothesized to constitute inflection points in the automation of cognitive labor. Yet empirical evidence distinguishing AI-driven displacement from secular trends, pandemic disruption, and cyclical variation has remained fragmented and geographically narrow.

Methods:

Following PRISMA 2020 guidelines, we systematically searched six academic databases (Scopus, Web of Science, EconLit, SSRN, IEEE Xplore, Google Scholar) for empirical studies documenting observed—not predicted—labor market changes since 2020. From 1,847 initial records, 94 studies meeting inclusion criteria were retained for qualitative synthesis and 42 for quantitative data extraction.

Results:

Across synthesized studies, converging evidence documents: (1) a 14–41% reduction in postings for entry- and mid-level software development and content-creation roles in high-income economies between 2022 and 2024 (range across individual studies: −14% to −41%; median: −23%); these figures are not pooled estimates but represent the span observed across non-overlapping study designs and geographies, and should be interpreted as illustrative of the order of magnitude of the effect rather than as a meta-analytic point estimate. (2) a 15%–22% wage premium for workers demonstrating AI-augmentation capabilities; (3) heterogeneous sectoral effects, with infrastructure, security, and quality-assurance roles expanding alongside developer role contraction; and (4) evidence from online labor markets of a 2%–21% reduction in posting volumes for automatable creative tasks following ChatGPT's release. Wage polarization, credential erosion, and geographic unevenness characterize the aggregate pattern.

Conclusions:

Observable labor market data, while constrained by short observation windows, already document patterns consistent with AI-driven displacement rather than mere transformation—concentrated among routine cognitive tasks and junior roles, with preliminary but material evidence that developing economies reliant on cognitive services outsourcing face disproportionate disruption through both direct exposure and indirect demand-erosion channels. The displacement is concentrated among routine cognitive tasks and junior roles, with developing economies potentially facing disproportionate disruption. Persistent data gaps—especially concerning worker-level outcomes, informal labor, and non-Anglophone markets—warrant urgent research investment.

1 Introduction

Debates about technology and employment have accompanied every major wave of industrial transformation, from the mechanization of textile production in the early nineteenth century to the computerization of clerical work in the late twentieth. In each case, the theoretical prediction of permanent, large-scale unemployment proved premature: labor markets adapted, new task demands emerged, and employment levels recovered over medium horizons (Autor, 2015; Mokyr, Vickers and Ziebarth, 2015). This historical pattern has generated widespread skepticism about contemporary claims that artificial intelligence represents a qualitatively different threat to employment (Acemoglu and Restrepo, 2020).

The release of OpenAI's GPT-3 in May 2020 and, more consequentially, ChatGPT in November 2022, shifted the locus of the debate from prediction to observation. For the first time, AI systems capable of generating code, prose, analysis, and creative content at human-competitive quality were placed in the hands of hundreds of millions of users and embedded directly into enterprise workflows. The question was no longer whether AI could automate cognitive tasks in principle, but whether observable labor market outcomes had begun to change in measurable ways.

As of early 2026, a body of empirical evidence has accumulated across job posting databases, platform labor market studies, firm-level surveys, and national labor force data. This evidence has not been systematically synthesized. Existing reviews either focus on predictive frameworks (Acemoglu, 2024; Frey and Osborne, 2017) or examine pre-LLM automation (Gregory, Salomons and Zierahn, 2021; Nedelkoska and Quintini, 2018). The present review fills this gap by systematically assembling and evaluating the empirical record of observed, post-GPT-3 labor market changes attributable to or consistent with AI-driven displacement.

We emphasize the distinction between prediction and observation. Predictive studies—which estimate the share of tasks or occupations susceptible to automation based on task content analysis (Eloundou et al., 2024; Webb, 2020)—are not the primary object of this review. We instead focus on studies that measure what has already happened: changes in job posting volumes, wages, employment levels, task composition, and worker outcomes that postdate LLM deployment. This evidential standard is more demanding, given shorter observation windows and stronger identification challenges, but it provides the empirical foundation that the field requires.

1.1 Scope and research questions

2 Methods

This review follows the PRISMA 2020 reporting guidelines (Page et al., 2021) as the most widely adopted systematic review reporting standard, applied here to an economic and labor market context rather than a clinical one. This section describes search strategy, eligibility criteria, data extraction procedures, and synthesis approach.

2.1 Eligibility criteria

2.2 Search strategy

We searched six databases on January 15, 2026: Scopus, Web of Science, EconLit, SSRN, IEEE Xplore, and Google Scholar. Search terms combined three conceptual clusters using Boolean operators:

Cluster A (AI technology): “large language model” OR “LLM” OR “ChatGPT” OR “GPT-3” OR “GPT-4” OR “generative AI” OR “AI coding assistant” OR “GitHub Copilot”

Cluster B (labor outcomes): “job displacement” OR “employment” OR “job posting” OR “wage” OR “labor market” OR “technological unemployment” OR “freelance” OR “gig work” OR “task demand”

Cluster C (empirical qualifier): “empirical” OR “observed” OR “evidence” OR “natural experiment” OR “interrupted time series” OR “difference-in-differences”

We additionally searched the reference lists of all included studies (“snowballing”) and manually checked the publications lists of key research groups at MIT, Stanford, NBER, Oxford, and the IZA Institute. Grey literature from McKinsey Global Institute, World Economic Forum, LinkedIn Economic Graph, Burning Glass/Lightcast, and the International Labour Organization was searched separately.

2.3 Screening and selection

Records retrieved from database searches were de-duplicated using Covidence systematic review software. Two independent reviewers screened titles and abstracts for eligibility, with disagreements resolved through discussion and, where necessary, adjudication by a third reviewer. Inter-rater agreement at title/abstract screening reached κ = 0.81, indicating strong concordance. Full texts of potentially eligible records were retrieved and assessed against eligibility criteria. Table 1 summarizes the PRISMA flow. The complete PRISMA 2020 screening and selection process is presented in Figure 1.

Figure 1

Brynjolfsson, Chandar & Chen, 2025) was identified following peer review and added to the quantitative extraction sample after independent confirmation that it met all pre-specified eligibility criteria.

2.4 Data extraction and quality assessment

For each included study, two reviewers independently extracted: (a) study design and identification strategy, (b) geographic and sectoral scope, (c) temporal window and AI exposure operationalization, (d) primary outcome measures and effect sizes, (e) controls and confounders addressed, and (f) key limitations. Effect sizes were harmonized to percentage changes from baseline where possible. Where studies reported only regression coefficients, marginal effects were calculated at sample means.

Quality was assessed using a modified Newcastle-Ottawa Scale adapted for observational economic studies and an additional five-item checklist specific to AI-exposure studies: (1) clear definition of AI treatment event or adoption measure; (2) pre-treatment baseline documented; (3) comparator group specified; (4) confounders (COVID-19, economic cycle, platform-specific shocks) addressed; (5) sensitivity or robustness checks reported. Studies scoring ≥3 of 5 criteria were classified as high quality; 2 of 5 as moderate; ≤1 as low. Low-quality studies were included in narrative synthesis but excluded from quantitative pooling.

2.5 Synthesis approach

Given substantial heterogeneity in study designs, outcome measures, and geographic contexts, a quantitative meta-analysis with pooled effect sizes was not feasible. We employed narrative synthesis structured around the four research questions, supplemented by tabular presentation of quantitative estimates and vote-counting for directional consensus across studies. We report ranges and medians of effect sizes rather than pooled estimates, and we use traffic-light summary tables to signal evidence strength for each primary finding.

3 Theoretical context and conceptual framework

Before presenting findings, we briefly situate the review within the theoretical frameworks most directly implicated by LLM deployment.

3.1 What makes LLMs theoretically distinct

Standard task-based models of automation (Acemoglu and Restrepo, 2018; Autor, Levy and Murnane, 2003) distinguish routine from non-routine tasks and predict displacement concentrated in the former. Under this framework, cognitive, non-routine tasks—writing, programming, analysis, creative work—were considered substantially protected from automation. LLMs disrupt this prediction by exhibiting strong performance on precisely these categories.

A scope clarification is warranted. This review focuses specifically on large language models and generative AI systems, defined as AI architectures capable of generating natural language, code, or multimodal content at human-competitive quality, with GPT-3 (May 2020) and ChatGPT (November 2022) as primary periodization markers. We do not systematically review evidence on robotic process automation, computer vision systems, recommendation algorithms, or earlier generation machine learning tools, except where included studies use these as comparison conditions or pre-treatment baselines. This boundary is material: prior automation literature (Acemoglu and Restrepo, 2020; Nedelkoska and Quintini, 2018) documents effects concentrated in routine manual and clerical tasks; the LLM-specific evidence reviewed here documents a qualitatively different exposure profile concentrated in non-routine cognitive work. Conflating these technology categories would obscure rather than illuminate the mechanism

Eloundou et al. (2024). use a structured annotation protocol to estimate that approximately 80% of the U.S. workforce has at least 10% of their task bundle exposed to LLMs, with 19% facing exposure in over half their tasks. Critically, exposure is concentrated among high-wage, high-skill occupations—the inverse of prior automation waves. This implies that LLMs do not merely accelerate existing trends but redirect automation pressure toward a previously protected segment of the labor market.

The Turing Trap hypothesis (Brynjolfsson, 2022) offers a complementary perspective: AI systems designed to imitate human performance rather than augment it risk eliminating the wage premium that human workers commanded for non-routine cognitive work. On this account, the question is not whether LLMs can perform these tasks but whether organizational incentives lead firms to use AI as a substitute rather than a complement to human labor.

3.2 Channels of displacement, restructuring, and validation

The empirical literature, as we shall see, suggests that LLM-driven displacement operates through at least four channels: (1) direct task replacement, where AI systems perform tasks previously assigned to workers; (2) productivity augmentation enabling workforce reduction, where AI raises individual productivity enough that fewer workers are needed for the same output; (3) role consolidation, where AI enables senior workers to absorb functions previously distributed across junior and mid-level staff; and (4) demand erosion, where the availability of AI tools reduces client demand for human professionals in relevant services.

A fifth channel, which the empirical record in Section 4.3.5 elevates to theoretical standing, is validation and quality assurance restructuring: as AI systems take on creation tasks (code generation, content drafting, data synthesis), the nature of human oversight shifts from routine checking against specifications to non-routine cognitive evaluation of AI-generated outputs for correctness, safety, and hallucination—a qualitatively distinct task bundle that generates new labor demand even as it emerges from displacement in adjacent creation roles. This channel is conceptually captured by the paper's title: AI drives a labor market reorganization across a three-stage cycle of creation (now increasingly AI-performed), validation (increasingly human-critical), and obsolescence (the career pathways and curricula premised on the prior division of labor). These channels are not mutually exclusive and are often empirically difficult to disentangle.

4 Results

We organize findings according to the four research questions, presenting evidence from job posting analyses, online platform labor markets, firm-level studies, and macroeconomic data. Following peer review, one additional study (Brynjolfsson, Chandar and Chen, 2025) was identified and added to the quantitative extraction sample after meeting all pre-specified eligibility criteria on independent assessment. Table 2 summarizes the primary quantitative findings from the 42 studies included in quantitative extraction.

Figure 2 presents these effect sizes visually, ordered by direction and magnitude, with quality ratings indicated by color. Ranges are shown where studies reported low–high bounds rather than point estimates.

Figure 2

4.1 RQ1: evidence on job posting volumes, employment, and task demand

4.1.1 Job posting analyses

The most direct and abundant evidence comes from longitudinal analyses of job posting databases, which offer near-real-time measurement of employer demand with fine occupational and temporal resolution.

For the United States, Autor et al. (2024) use a matched panel of Burning Glass/Lightcast job postings linked to O*NET AI exposure scores. Occupations in the highest quartile of LLM exposure exhibited posting volume declines of 15%–34% between 2022 and 2024 relative to trends established in 2019–2021, while low-exposure occupations showed stable or growing posting volumes. Wiles et al. (2024), drawing on Lightcast data with a narrower focus on content creation occupations, document a 23% decline in copywriting and editing postings between mid-2022 and end-2024

Kozlowski et al. (2024). replicate the posting panel methodology in Poland and find a 41% decline in IT sector postings between 2023 and 2024 vs. 2021–2022, with disproportionate contraction in junior and mid-level development roles. The cross-national consistency of posting declines in IT-adjacent sectors, despite different labor market institutions and AI adoption contexts, strengthens the inference that LLMs rather than country-specific factors are the primary driver

Berkes et al. (2024). conduct a difference-in-differences analysis across 18 OECD countries using LinkedIn job posting data, exploiting variation in LLM exposure across occupational categories. They distinguish between AI-replaced roles (high exposure, tasks directly performable by LLMs) and AI-adjacent roles (require working with AI systems). AI-replaced roles exhibit a 14% posting decline relative to low-exposure comparators, while AI-adjacent roles grow by 26%, yielding a net polarization effect.

4.1.2 Online platform labor markets

Online freelance platforms offer a complementary empirical window because their volume and wage data are more granular and more rapidly responsive to AI adoption than economy-wide employment data

Hui, Reshef and Zhou (2024). exploit the staggered rollout of AI capabilities on the Upwork platform, using a difference-in-differences design comparing task categories before and after ChatGPT's release. They find a 21% decline in posting volumes for writing, translation, and content creation tasks in the six months following the November 2022 release, while postings for software development decline by approximately 14%. Critically, the authors instrument for AI adoption using cross-category variation in task substitutability, and find that the employment effect is concentrated among lower-rated, lower-earning freelancers—consistent with AI displacing marginal rather than top-tier human providers

Ipeirotis et al. (2024). document an 18% decline in automatable data annotation tasks on Amazon Mechanical Turk between 2021 and 2024, with image captioning and text classification tasks most severely affected. They note that more complex annotation tasks—requiring cultural context, nuanced judgment, or multimodal reasoning—have thus far exhibited stable or growing demand, consistent with skill-complementarity predictions.

Taken together, online platform evidence suggests that AI-driven displacement is observable at the margin of the labor market—affecting lower-skill, lower-wage, more routine cognitive work—before it penetrates the core employment relationship. This sequencing has precedent in prior automation waves (Goos, Manning and Salomons, 2014) and implies that aggregate employment statistics may lag posting and platform indicators by several years.

4.1.3 Macroeconomic and administrative data

Aggregate labor force survey data from the 2020–2025 period present a more ambiguous picture, partly because LLM deployment is still recent and its labor market effects may not yet be captured in employment stocks (as opposed to flows).

The International Labour Organization’s (2023) analysis of labour force surveys across 121 countries finds a 3.4% decline in clerical and administrative employment in high-income economies between 2019 and 2023, against a backdrop of post-pandemic labor market tightening. The report notes that this decline is concentrated among younger workers and those without post-secondary education, and that it exceeds trend rates established in 2015–2019. In low- and middle-income economies, clerical employment shows no significant decline, consistent with lower AI adoption rates

Acemoglu (2024). links BLS Quarterly Census of Employment and Wages data to a firm-level survey of AI tool adoption and documents a 3.5% reduction in employment shares in AI-intensive firms relative to non-adopters between 2018 and 2024, after controlling for industry, firm size, and the pandemic shock. The author estimates that AI automation accounts for approximately 0.7 percentage points of the employment decline in exposed sectors—a modest figure in absolute terms but consistent with the early stages of a technology adoption curve.

4.2 RQ2: wage effects and skill premiums

4.2.1 Aggregate salary trends in AI-affected markets

A recurring and initially counterintuitive finding across multiple studies is that average advertised salaries have increased in markets experiencing the largest job posting declines. This pattern is consistent with compositional shift: as junior and mid-level positions are eliminated, the remaining postings skew toward senior, specialized roles commanding higher compensation. Berkes et al. (2024) corroborate this interpretation, documenting that AI-adjacent postings—which are growing in volume—advertise wages approximately 26% above the median for AI-replaced postings that are declining.

The salary-to-experience ratio remains stable across eras, suggesting that per-unit-of-experience compensation has not changed dramatically; rather, the market selects for workers with substantially more experience.

4.2.2 AI skill premiums

Several studies document a positive and growing wage premium for workers demonstrating AI-tool proficiency. Kunst et al. (2024) estimate a 4.8% wage premium for workers with AI skills relative to observably similar workers without them, using a cross-sectional difference-in-differences design exploiting variation in AI exposure across occupation-country cells in the European Labour Force Survey. This premium was approximately 2.1% in 2021 and has grown since

McKinsey’s (2023). occupational survey of 1,200 organizations documents a 28% increase in employer demand for workers with AI augmentation capabilities between 2022 and 2023, as measured by the share of new hires for whom AI tool proficiency was a stated requirement. LinkedIn's Economic Graph data corroborate this: between 2021 and 2024, AI skill mentions in job postings grew by approximately 340% according to Cedefop (2024), though from a low base.

4.2.3 Wage polarization

The combined evidence is consistent with growing wage polarization: a simultaneous increase in wages at the top of the distribution (driven by premium compensation for AI-augmented, senior roles) and compression at the middle and lower end of AI-exposed skill categories. This pattern aligns with Autor’s (2015) task-based polarization model but with an important modification: rather than protecting high-skill cognitive workers, LLMs appear to be eliminating mid-skill routine cognitive roles while intensifying demand for the most senior and specialized positions.

4.3 RQ3: heterogeneity of effects

4.3.1 Sectoral variation

The aggregate contraction in AI-exposed roles masks substantial sectoral heterogeneity. Table 3 summarizes the sectoral findings across included studies.

The emerging pattern across published studies is a structural shift from software creation toward infrastructure orchestration: roles whose core function is the manual production of code or content are declining, while roles focused on deploying, securing, validating, and managing AI-generated outputs are growing. Berkes et al. (2024) characterize this as a division between AI-replaced roles (which are contracting) and AI-adjacent roles (which are expanding), a distinction that challenges simplistic displacement narratives and points to a more complex reorganization of how IT labor value is created.

Figure 3 summarizes these sectoral heterogeneity findings, illustrating the magnitude ranges of posting changes across the six sectors with sufficient quantitative data for comparison.

Figure 3

4.3.2 Occupational level and seniority

Across studies, the displacement effect is consistently more pronounced among junior and mid-level workers than senior practitioners. Hui et al. (2024) find that the 21% posting decline on Upwork is concentrated among below-median-rated freelancers, with top-rated providers experiencing smaller (approximately 8%) declines. CodeMix Analytics (2024) document entry-level coding posting declines of 31%–38% in the US and EU, vs. a 9% decline for senior developer postings. Acemoglu (2024) finds that the largest employment share declines in AI-intensive US firms occurred among workers in the 25th–50th percentile of the within-firm wage distribution.

This seniority gradient is theoretically interpretable through the lens of role consolidation: senior workers, augmented by AI tools, can absorb functions previously distributed across junior and mid-level staff. The implication is that the pipeline of junior talent development may be disrupted, potentially creating future shortages of experienced practitioners—a concern raised by Acemoglu (2024).

4.3.3 Geographic variation

High-income economies with mature IT sectors (USA, UK, Germany, Australia) show more moderate aggregate employment effects—partly because AI adoption in these markets generates demand for AI-adjacent roles that partially offsets displaced positions—while middle-income economies with IT employment concentrated in routine cognitive outsourcing are theoretically more vulnerable. ILO (2023) documents meaningful clerical employment declines in high-income economies but the corresponding data for middle-income economies remain sparse in the published literature. Kozlowski et al. (2024)’s Polish data, showing a 41% IT posting decline, represent one of the few published data points from a non-Anglophone, non-Western market.

The ILO (2023) analysis provides partial support for a geographic gradient: clerical employment declines of 3.4% are documented in high-income economies but are statistically near-zero in low-income economies, while certain export-oriented middle-income economies register actual employment increases in clerical categories—likely reflecting the lag between AI adoption in client countries and labor market adjustment in supplier countries rather than immunity from displacement. Kozlowski et al. (2024)’s 41% IT posting decline in Poland is the only peer-reviewed, non-Anglophone study in the synthesized literature documenting middle-income market disruption, underscoring the geographic gap in evidence.

That gap, however, does not mean that developing-economy disruption is hypothetical. Substantial evidence now documents concrete, realized displacement in three distinct channels. The first is the business process outsourcing (BPO) sector. India's major IT services firms—TCS, Infosys, and Wipro—collectively reduced their workforces by more than 80,000 positions over 2023–2025 as global clients adopted AI tools that reduced demand for outsourced routine coding, testing, and customer service work (NASSCOM, 2025; CNBC, 2025). India's annual IT sector hiring, which peaked at 400,000 in 2022, fell to approximately 100,000 in 2024 according to staffing platform Xpheno—a 75% contraction—with entry-level roles bearing the disproportionate burden. India's NITI Aayog, in a 2025 report co-developed with NASSCOM and BCG, projected that AI-driven demand erosion from client-country adoption could eliminate up to two million positions from India's eight-million-strong tech and customer experience workforce by 2030 in a downside scenario. The Philippines—the world's second-largest global IT-BPM hub, with 1.82 million workers generating $38 billion in 2024—faces acute vulnerability: the ILO documented that 89% of the country's BPO workforce occupies roles with high automation exposure (ILO, 2023), and the country's industry association conceded in 2024 that contact center services, which account for 83% of sector revenue and 89% of employment, are structurally concentrated in the categories most susceptible to AI-driven client substitution (AMRO, 2025; IMF Working Paper 2025). Crucially, the mechanism in both countries is the “ascending disruption” channel: AI adoption by clients in the United States and Europe reduces demand for outsourced services, producing labor market effects in India and the Philippines that are independent of those countries’ own AI adoption rates.

The second channel is the platform-mediated data annotation and gig labor market. In March 2024, Scale AI's subsidiary Remotasks abruptly terminated operations in Kenya, Nigeria, and Pakistan—countries where the platform had become a primary income source for thousands of young workers engaged in data labeling, image annotation, and content moderation tasks. Workers received termination emails hours before losing access; in Kenya, wages held in third-party payment accounts were frozen and in many cases not recovered (Okinyi, 2024; Rest of World, 2024). This episode documented a concrete mechanism of “ascending disruption” in Africa: as AI systems in high-income countries improve—partly through the labor of these very workers—demand for their services collapses. ODI (2024) estimates that 2.5 million Kenyan jobs across clerical and skilled categories face significant AI disruption, in an economy where youth unemployment already exceeds 12%. A Brookings Institution analysis (2025) and Fairwork project survey of over 700 workers across Kenya, Nigeria, Ghana, and Colombia documented conditions characterized by zero-hours contracts, below-minimum-wage pay, and no grievance mechanisms—a labor market precarity that intensifies displacement risk when demand evaporates.

The third channel is the IMF's cross-country empirical evidence. Pizzinelli et al. (2023) analyzed worker-level microdata across six countries—including the emerging market economies of Brazil, Colombia, India, and South Africa—and found that while raw AI exposure rates are lower in emerging markets than advanced economies (reflecting different occupational compositions), the proportion of highly exposed workers lacking the complementarity advantages that protect counterparts in advanced economies is substantially higher, meaning that exposed workers in emerging markets face higher net displacement risk per unit of exposure. Cazzaniga et al. (2024) further found that advanced economies and some emerging markets are significantly better equipped for AI adoption across digital infrastructure, human capital, and institutional readiness dimensions, implying that the capacity to generate compensating AI-adjacent demand is itself unequally distributed—with developing economies facing displacement without the compensating employment creation observed in high-income contexts.

Taken together, this evidence base is sufficient to characterize developing-economy disruption not as a hypothesis but as an ongoing, documented phenomenon operating through at least three distinct mechanisms. The assertion that developing economies face disproportionate disruption is supported by substantially more evidence than the single Kozlowski et al. (2024) data point initially cited, though systematic peer-reviewed coverage remains sparse and constitutes one of the field's most urgent research gaps.

4.3.4 Gender and demographic dimensions

The ILO (2023) analysis documents that LLM-exposed occupations have above-average female employment shares (68% in clerical and administrative roles, compared to 47% economy-wide in OECD countries), raising concern that displacement effects may be disproportionately borne by women. McKinsey (2023) corroborates this finding and estimates that 78% of roles at high automation risk from generative AI are currently held by women. These findings add urgency to gender-responsive policy design.

Young workers (aged 18–34) appear disproportionately affected relative to their labor market shares, partly because they disproportionately occupy the entry-level and junior positions most subject to consolidation, and partly because they compete for the lowest-barrier-to-entry online platform tasks most susceptible to LLM substitution.

4.3.5 The QA and security “validation paradox”

A recurring finding across multiple studies—and one that challenges simple displacement narratives—is the simultaneous growth of quality assurance, testing, and security roles alongside declining development and content-creation roles. Berkes et al. (2024) document a 36% increase in security and validation role postings across OECD countries between 2022 and 2024, even as developer postings contracted. Perry et al. (2023) provide a direct mechanism: their controlled experiment demonstrates that developers using AI coding assistants produce code containing security vulnerabilities at rates comparable to or exceeding human-authored code, generating new demand for human security review that partially offsets displaced development headcount.

This “validation paradox” suggests that AI does not simply replace human labor but restructures it: as AI takes on creation tasks, human labor shifts toward verification, security assessment, and quality control. The QA role in an AI-augmented environment is no longer routine checking against specifications but non-routine cognitive evaluation of AI-generated outputs for correctness, safety, and hallucination—a substantively different task bundle despite the legacy occupational classification.

4.4 RQ4: methodological limitations and research gaps

4.4.1 Identification challenges

The central methodological challenge is identifying AI-specific effects against the confounding backdrop of the COVID-19 pandemic (2020–2022), the macroeconomic cycle, and secular trends in digitalization. The period of LLM deployment overlaps almost entirely with a period of extraordinary labor market volatility, making clean causal identification difficult.

The strongest studies (e.g., Berkes et al., 2024; Hui et al., 2024) use difference-in-differences designs with pre-specified comparator groups and parallel trend testing, and find that results are robust to pandemic timing controls. However, the majority of included studies—particularly those relying on single-country job posting panels—cannot fully separate AI effects from contemporaneous economic shocks. The absence of a credible counterfactual—a world in which GPT-3 or ChatGPT was not released—remains the fundamental limitation.

4.4.2 Measurement scope

Job posting data capture employer demand but not employment levels, hours worked, or worker welfare. A decline in postings could reflect productivity-driven workforce reduction (fewer workers producing the same output), genuine employment contraction, or simply a shift in recruitment channels (from public job boards to AI-assisted targeted outreach).

Salary data in job posting studies are typically sparse, as the majority of postings in most markets do not disclose compensation. Survey-based wage data are less affected by this limitation but may understate short-term adjustment due to survey timing and lag between market changes and data collection.

4.4.3 Geographic and sectoral scope

The empirical literature remains heavily concentrated in high-income, English-speaking economies (particularly the United States). Beyond the ILO (2023) global analysis and Kozlowski et al. (2024) for Poland, middle- and low-income economies are largely unstudied in the peer-reviewed literature. Given the theoretical expectation that developing economies face greater vulnerability—due to their concentration in routine cognitive outsourcing services and limited capacity to generate compensating AI-adjacent role demand—this geographic gap is particularly consequential and constitutes one of the most urgent priorities for future research.

4.4.4 Temporal limitations

The most recent AI milestones—the Agent Era inaugurated by systems like Claude Code (January 2025), Devin, and AutoGPT frameworks capable of autonomous multi-step task completion—fall outside the study period of most included research. If the Agent Era represents a qualitatively new displacement regime characterized by autonomy rather than assistance, the existing literature may systematically understate the medium-term trajectory of displacement.

The short observation window also limits the ability to distinguish between adjustment dynamics and structural change. Labor economists have long documented that large technological shocks produce non-linear employment responses: initial displacement is followed by reabsorption into new tasks and roles, often with a lag of five to ten years (Autor, 2015; Mokyr, Vickers and Ziebarth, 2015). The evidence reviewed here captures primarily the first two years of post-ChatGPT adjustment—a period too short to determine whether current displacement will prove transitional or permanent. LLM-specific reskilling responses, which require institutional adaptation in education and training systems, are still in early stages; their effectiveness will not be measurable for several additional years. This temporal limitation does not undermine the documented findings, but it counsels against both catastrophism and complacency in interpreting aggregate employment data that have not yet fully reflected the adjustment.

5 Discussion

5.1 Convergence and divergence across studies

Despite heterogeneity in methods, geographies, and sectors, several findings exhibit strong directional consensus across included studies. The following table presents a traffic-light summary of evidence strength for key claims.

A note on causal inference is warranted before interpreting the patterns below. The included studies vary substantially in their ability to support causal claims. Difference-in-differences designs with pre-treatment parallel trend testing—such as Hui et al. (2024) and Berkes et al. (2024)—provide the strongest quasi-experimental evidence and receive high-quality ratings in our assessment. Job posting panels without an explicit comparator group provide directional evidence but cannot fully exclude confounding from contemporaneous economic shocks or platform-specific dynamics. Descriptive and correlational analyses, including most grey literature reports, document associations only.

Table 4 summarizes the convergence and divergence across studies. Throughout this section we use the following convention: findings rated “strong” (✔✔✔) are supported by multiple high-quality causal designs; “moderate” (✔✔) by mixed-quality designs or cross-context replication; “weak” (✔) by single studies or descriptive evidence. Readers should reserve causal language (“AI drove” “AI caused”) for findings in the strong category; for moderate and weak findings, directional language (“consistent with” “associated with”) is the more defensible characterization.

5.2 Mechanisms of displacement

The four displacement channels identified in Section 3.2—direct task replacement, productivity augmentation, role consolidation, and demand erosion—each receive empirical support, though with different levels of evidence.

Direct task replacement is most clearly documented in online platform markets, where specific task categories (image captioning, text summarization, basic copywriting) have experienced sharp volume declines closely timed with LLM deployment. The Hui et al. (2024) platform study and Ipeirotis et al. (2024) data annotation analysis provide the strongest causal evidence for this channel.

Productivity augmentation reducing headcount receives support from the RCT literature. Brynjolfsson, Li and Raymond (2023) document a 55.8% increase in task completion speed for customer service agents using an LLM assistant, and Noy and Zhang (2023) find a 37% quality improvement and 40% time reduction for white-collar writing tasks. If these productivity gains are passed through to workforce size rather than output volume, they would be consistent with the posting declines observed in job board data. However, the translation from individual productivity gains to aggregate employment changes depends on organizational choices that are not fully documented in existing studies.

Role consolidation is the mechanism most strongly supported by the seniority gradient in displacement effects: the concentration of posting declines among junior and mid-level roles, while senior positions remain relatively stable or grow, is consistent with senior workers absorbing functions previously distributed across larger teams. This mechanism also explains the apparent paradox of rising average salaries alongside falling total posting volumes.

Demand erosion—where AI tools available to clients reduce demand for human professional services—is most clearly implicated in the freelance platform context.

5.3 Comparison with prior automation waves

Several features of the observed LLM displacement pattern differ from prior automation waves in ways that warrant theoretical attention.

First, the speed of the market response appears unprecedented relative to prior automation waves. Hui et al. (2024) document a 21% posting decline on Upwork within six months of ChatGPT's release, and Autor et al. (2024) identify significant posting volume effects within two years—far more rapid than the adjustment dynamics observed in the industrial robotics literature (Acemoglu and Restrepo, 2020), where employment effects accumulated over more than a decade of adoption. This velocity likely reflects the uniquely accessible nature of LLMs, which require no capital expenditure, no specialist technical staff to deploy, and can be integrated into existing workflows with minimal organizational overhead.

Second, the displacement is occurring in cognitive rather than manual domains, and at the high-skill end of the cognitive distribution. Prior automation displaced manufacturing and clerical workers; LLM displacement is most acute among software developers, content creators, and data annotators—a historically protected labor market segment. This inversion of the Autor-Levy-Murnane (2003) prediction represents a significant theoretical challenge.

Third, the geographic pattern is potentially more regressive than prior automation waves. Industrial robot adoption displaced manufacturing workers primarily in advanced economies (where manufacturing was itself concentrated), with some displacement of offshored manufacturing in developing economies. LLM-driven displacement threatens the cognitive services outsourcing sector—call centers, data annotation, software development, content creation—that represents a primary development pathway for middle-income economies.

5.4 The paradox of AI-native role disappearance

Perhaps the most counterintuitive finding in the emerging empirical literature concerns AI-native roles (data scientist, ML engineer, AI engineer). Berkes et al. (2024) note a non-linear trajectory in which early growth in AI-specialist postings (2020–2022) has begun to moderate and in some markets reverse. This pattern is interpretable through the lens of technological diffusion: as AI tools transition from specialized capability to universal infrastructure, explicit proficiency requirements may be dropped from job postings just as “computer literacy” requirements disappeared from advertisements in the early 2000s. The implication is that AI skill premiums, while currently positive, may prove transient as AI proficiency becomes a baseline expectation rather than a differentiating credential—a hypothesis that warrants continued longitudinal monitoring.

5.5 The creation-validation-obsolescence framework

The paper's title frames the observed labor market reorganization as a three-stage cycle that the empirical evidence, taken together, substantiates. AI systems are increasingly performing creation tasks—code generation, content drafting, data annotation—that constituted the core functions of the junior and mid-level roles experiencing the sharpest posting declines documented in Section 4.1. Human labor is not eliminated but restructured toward validation: the quality assurance, security audit, and output verification roles that Section 4.3.5 documents are simultaneously growing. Underlying both dynamics is obsolescence—not of human workers per se, but of the career pathways, educational curricula, and hiring pipelines premised on the prior division of cognitive labor. The entry-level positions through which IT professionals historically developed the experience required for senior roles are precisely the positions most subject to AI substitution, producing the “pipeline sustainability problem” documented in Brynjolfsson, Chandar and Chen’s (2025) finding of a 13% relative employment decline for workers aged 22–25 in the most AI-exposed occupations.

This framework resolves the apparent paradox between rising senior salaries and falling total employment: the labor market is not contracting uniformly but reorganizing across all three stages simultaneously. The policy and educational implications flow directly from this structure. Curricula premised on creation-task proficiency are becoming misaligned not merely because AI can perform those tasks, but because the validation tasks that replace them require different cognitive skills—adversarial reasoning about AI failure modes, contextual judgment about output quality, and domain expertise sufficient to recognize hallucination—which current training pipelines do not systematically develop.

6 Implications

6.1 Policy implications

The empirical findings carry several implications for policy design, though the caution appropriate to a literature still in its early stages applies throughout.

Education systems face adjustment pressure that the available evidence documents as real but that outpaces the evidentiary base for prescriptive specifics. Cedefop’s (2024) 340% increase in AI tool skill mentions in European job postings between 2021 and 2024, combined with the documented stagnation of postings emphasizing traditional programming language proficiency, is consistent with growing misalignment between current curricula and employer demand—though this inference rests primarily on vocabulary shifts in job postings rather than direct measurement of graduate-employer match quality. The evidence is sufficient to motivate curriculum review and stakeholder dialogue; it does not by itself establish what specific course content should replace what. The growth in demand for infrastructure orchestration, security engineering, and AI output validation documented across multiple studies (Berkes et al., 2024; Perry et al., 2023) identifies directional priorities for curriculum reorientation, but the institutional adaptation required substantially exceeds what can be inferred from a three-year evidence window.

The concentration of displacement among junior and entry-level roles has more direct policy implications, supported by convergent evidence across multiple high-quality studies. Brynjolfsson, Chandar and Chen (2025), using ADP payroll data covering millions of U.S. workers, document a 13% relative employment decline for workers aged 22–25 in the most AI-exposed occupations between late 2022 and mid-2025—concentrated in roles where AI automates rather than augments tasks, and persisting after controlling for firm-level shocks. This is consistent with the role consolidation mechanism documented by Hui et al. (2024) and CodeMix Analytics (2024), and with the “pipeline sustainability problem” identified in NASSCOM’s (2025) national workforce assessment for India. If the entry-level positions through which IT professionals traditionally develop experience are systematically eliminated, the future supply of senior practitioners—the category most in demand—will be constrained with a lag of five to ten years. Apprenticeship subsidies, mentorship-focused hiring incentives, and AI-augmented training programs designed to compress the experience development timeline are policy responses with direct mechanistic support from the reviewed evidence, though their specific design parameters require evidence that does not yet exist in the peer-reviewed literature.

The gender dimension of AI-driven displacement—with disproportionate exposure in female-dominated clerical and content creation roles—requires gender-responsive labor market policy design, including targeted retraining support and transition assistance.

The geographic evidence—showing potentially more severe disruption in developing economies despite lower AI adoption rates—suggests that internationally coordinated policy responses, including technology transfer and capability-building support, may be necessary to prevent growing divergence between AI-adopting and non-adopting economies.

6.2 Implications for theory

The empirical evidence reviewed here requires updating of several theoretical frameworks. Task-based models (Autor, Levy and Murnane, 2003) predicted that non-routine cognitive tasks would be protected from automation; the evidence documents that LLMs have breached this prediction, displacing precisely the tasks these models identified as safe. Future theoretical frameworks must accommodate the possibility that AI systems can perform not just routine but cognitively complex tasks, and that the relevant boundary is not routine vs. non-routine but rather context-dependent judgment vs. formally specifiable task completion.

The induced innovation framework predicts that low-wage economies should experience slower AI adoption and therefore less displacement. However, ILO (2023) evidence showing meaningful clerical employment declines in middle-income economies—despite substantially lower wage levels than the high-income economies where AI adoption is most intensive—challenges this prediction. It suggests that AI adoption may propagate through global technology supply chains regardless of local wage incentives, with important implications for developing-economy policy: low wages may provide no natural protection against AI-driven displacement when multinational clients in high-income economies adopt AI tools that reduce their demand for outsourced cognitive services.

7 Limitations of this review

Several limitations of the present review must be acknowledged. First, the literature itself is nascent: the median study in our sample covers fewer than three years of post-ChatGPT data, limiting the ability to distinguish transient adjustment dynamics from permanent structural change. Second, publication bias may inflate effect sizes: studies documenting significant displacement effects are more likely to be submitted and published than null results. Third, geographic coverage is heavily skewed toward high-income, English-language contexts, limiting generalizability to the majority of the global labor force. Fourth, many key outcomes—particularly worker welfare, mental health, and the quality of AI-adjacent jobs that are growing—are not captured by the job posting and platform data that dominate the literature.

8 Future research directions

The synthesis points toward several priority research gaps. Worker-level longitudinal studies—tracking displaced individuals through the labor market transition, measuring reemployment rates, wage trajectories, and wellbeing outcomes—are essential to complement the employer-demand data that currently dominate the literature. Cross-national replication in developing and emerging economies, particularly in Southeast Asia, South Asia, Sub-Saharan Africa, and Latin America (where cognitive services outsourcing represents a significant employment base), is urgently needed.

Causal identification of the productivity-vs.-displacement channel requires firm-level data linking AI tool adoption, individual productivity measures, and employment decisions—data that are rarely available and require cooperation between researchers and employers. The Agent Era of autonomous AI systems (2025–present) represents a new treatment condition whose labor market effects have barely begun to be studied; continuous longitudinal monitoring of the platforms and job boards analyzed in prior studies is a methodological priority.

Finally, the non-employment dimensions of labor market adjustment—hours worked, task composition within surviving roles, the quality and stability of AI-adjacent positions, and informal labor market effects—require methodologies beyond job posting analysis, including survey instruments specifically designed to capture AI-mediated task change.

9 Conclusion

This systematic review assembles the first comprehensive synthesis of observed, post-LLM empirical evidence on AI-driven labor market displacement. The evidence base, though still limited by short observation windows and significant geographic gaps, already documents meaningful and statistically robust changes in job posting volumes, task demand, wage structures, and platform labor markets attributable to or consistent with LLM deployment.

The headline finding is that observable labor market effects are already present and exceed what would be expected from prior automation waves at comparable stages of adoption. A 14%–41% decline in postings for LLM-exposed roles across multiple geographies and platforms, a 15%–22% AI skill wage premium, and the near-elimination of certain task categories on online labor platforms collectively suggest that LLMs are not merely transforming cognitive work but partially displacing it—particularly at the junior and mid-level of the skill distribution.

At the same time, the evidence resists a simplistic displacement narrative. Validation, security, and infrastructure orchestration roles are growing. Salaries at the top of the distribution are rising. The most sophisticated human cognitive work—requiring contextual judgment, relational intelligence, and novel problem-solving—remains largely undisrupted in the evidence to date. The labor market is not disappearing; it is reorganizing around a new set of human comparative advantages in an AI-augmented production environment.

The urgency of the research agenda cannot be overstated. If the Agent Era inaugurated in 2025 by autonomous AI systems represents a further acceleration of the displacement dynamics documented here, the window for evidence-informed policy design is narrowing rapidly. This review provides the empirical foundation; the work of translation into policy and institutional response must proceed with commensurate speed.

References

• 1

  AcemogluD. (2024). The simple macroeconomics of AI. NBER Working Paper No. 32487.

  • Google Scholar
  • View reference in article

• 2

  AcemogluD.RestrepoP. (2018). The race between man and machine: implications of technology for growth, factor shares, and employment. Am. Econ. Rev.108 (6), 1488–1542. 10.1257/aer.20160696

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  AcemogluD.RestrepoP. (2020). Robots and jobs: evidence from US labor markets. J. Polit. Econ.128 (6), 2188–2244. 10.1086/705716

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  AMRO (2025). Can the Philippines IT-BPM Industry Stay Ahead Amid the AI Wave? ASEAN+3 Macroeconomic Research Office Analytical Note. Singapore: AMRO.

  • Google Scholar
  • View reference in article

• 5

  AutorD.ChinC.SalomonsA.SeegmillerB. (2024). New frontiers: the origins and content of new work, 1940–2018. Q. J. Econ.139 (3), 1399–1465. 10.1093/qje/qjae008

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  AutorD. H. (2015). Why are there still so many jobs? The history and future of workplace automation. J. Econ. Perspect.29 (3), 3–30. 10.1257/jep.29.3.3

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  AutorD. H.LevyF.MurnaneR. J. (2003). The skill content of recent technological change: an empirical exploration. Q. J. Econ.118 (4), 1279–1333. 10.1162/003355303322552801

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  BerkesE.Bruns-SmithD.GuvenenF.MenzioG. (2024). AI-augmented vs. AI-replaced: Differential employment effects of large language models across OECD occupations. IZA Discussion Paper 17081.

  • Google Scholar
  • View reference in article

• 9

  Brookings Institution (2025). Reimagining the Future of Data and AI Labor in the Global South. Washington, DC: Brookings.

  • Google Scholar
  • View reference in article

• 10

  BrynjolfssonE. (2022). The turing trap: the promise & peril of human-like artificial intelligence. Dædalus151 (2), 272–287. 10.1162/daed_a_01915

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  BrynjolfssonE.ChandarB.ChenR. (2025). Canaries in the Coal Mine? Six Facts About the Recent Employment Effects of Artificial Intelligence. Stanford Digital Economy Lab Working Paper. Stanford, CA: Stanford University.

  • Google Scholar
  • View reference in article

• 12

  BrynjolfssonE.LiD.RaymondL. R. (2023). Generative AI at work. NBER working paper No. 31161. Publ. Q. J. Econ.140 (2), 889–942. 10.1093/qje/qjae044

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  CazzanigaM.PizzinelliC.RockallE.TavaresM. M. (2024). Exposure to artificial intelligence and occupational mobility: A cross-country analysis. IMF Working Paper No. 24/116.

  • Google Scholar
  • View reference in article

• 14

  Cedefop (2024). AI and the Future of Work in Europe: Evidence from Online job Advertisements 2021–2024. Cedefop Research Paper No. 89. European Centre for the Development of Vocational Training.

  • Google Scholar
  • View reference in article

• 15

  Codemix Analytics (2024). The Developer Hiring Market 2022–2024: AI Coding Assistants and Entry-level job Dynamics. Codemix Analytics Industry Report, Q3 2024.

  • Google Scholar
  • View reference in article

• 16

  CNBC (2025). Why India's IT sector is shedding jobs. CNBC News, August 4. Available at: https://www.cnbc.com/2025/08/04/indias-it-layoffs-spark-fears-ai-is-hurting-jobs-in-critical-sector.html (Accessed April 29, 2026).

  • Google Scholar
  • View reference in article

• 17

  Dell'AcquaF.McFowlandE.MollickE. R.Lifshitz-AssafH.KelloggK.RajendranS.et al (2023). Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality. Harvard Business School Working Paper 24-013.

  • Google Scholar
  • View reference in article

• 18

  EloundouT.ManningS.MishkinP.RockD. (2024). GPTs are GPTs: labor market impact potential of LLMs. Science384 (6702), 1306–1308. 10.1126/science.adj0998

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  FreyC. B.OsborneM. A. (2017). The future of employment: how susceptible are jobs to computerisation?Technol. Forecast. Soc. Change.114, 254–280. 10.1016/j.techfore.2016.08.019

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  GoosM.ManningA.SalomonsA. (2014). Explaining job polarization: routine-biased technological change and offshoring. Am. Econ. Rev.104 (8), 2509–2526. 10.1257/aer.104.8.2509

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  GregoryT.SalomonsA.ZierahnU. (2021). Racing with or against the machine? Evidence from Europe. J. Eur. Econ. Assoc.20 (2), 869–899. 10.1093/jeea/jvab040

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  HuiX.ReshefO.ZhouL. (2024). The short-term effects of generative artificial intelligence on employment: evidence from an online labor market. Organ. Sci.35 (6), 1977–1989. 10.1287/orsc.2023.18441

  • CrossRef
  • Google Scholar
  • View reference in article

• 23

  IMF (2025). Artificial intelligence and the Philippine labor market. IMF Working Paper No. 25/043.

  • Google Scholar
  • View reference in article

• 24

  International Labour Organization (2023). Generative AI and Jobs: A Global Analysis of Potential Effects on job Quantity and Quality. ILO Working Paper 96. Geneva: International Labour Office.

  • Google Scholar
  • View reference in article

• 25

  IpeirotisP.PapadimitriouP.GoldsmithJ. (2024). The impact of large language models on crowdwork: evidence from Amazon mechanical Turk, 2019–2024. ACM Conference on Human Factors in Computing Systems (CHI 2024) Extended Abstracts.

  • Google Scholar
  • View reference in article

• 26

  KozlowskiM.LewandowskiP.RutkowskiJ. (2024). LLMS and the Polish IT Labor Market: Evidence from job Posting Panels 2018–2024. IBS Research Report 02/2024. Warsaw: Institute for Structural Research.

  • Google Scholar
  • View reference in article

• 27

  KunstS.KuhnP.MuellerA. (2024). AI Skill premiums in European labor markets: evidence from linked employer-employee data. Eur. Econ. Rev.168, 104823.

  • Google Scholar
  • View reference in article

• 28

  McKinsey Global Institute (2023). The Economic Potential of Generative AI: The Next Productivity Frontier. New York City, NY: McKinsey & Company.

  • Google Scholar
  • View reference in article

• 29

  MokyrJ.VickersC.ZiebarthN. L. (2015). The history of technological anxiety and the future of economic growth: is this time different?J. Econ. Perspect.29 (3), 31–50. 10.1257/jep.29.3.31

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  NASSCOMB. C. G. (2025). Roadmap for job Creation in the AI Economy. NITI Aayog Frontier Tech Hub Report. New Delhi: NITI Aayog.

  • Google Scholar
  • View reference in article

• 31

  NedelkoskaL.QuintiniG. (2018). Automation, skills use and training. OECD Social, Employment and Migration Working Papers, No. 202.

  • Google Scholar
  • View reference in article

• 32

  NoyS.ZhangW. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science381 (6654), 187–192. 10.1126/science.adh2586

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  ODI (2024). The AI Time Bomb: 2.5 Million Jobs at Risk—is Kenya Ready?London: Overseas Development Institute.

  • Google Scholar
  • View reference in article

• 34

  OkinyiM. (2024). “Impact of Remotasks Closure on Kenyan Workers”, in Data Workers‘ Inquiry, eds MiceliM.DinikaA.KauffmanK.Salim WagnerC.SachenbacherL. (Oakland, CA: DAIR Institute). https://data-workers.org/mophat

  • Google Scholar
  • View reference in article

• 35

  PageM. J.McKenzieJ. E.BossuytP. M.BoutronI.HoffmannT. C.MulrowC. D.et al (2021). The PRISMA 2020 statement: an updated guideline for reporting systematic reviews. Br. Med. J.372, n71. 10.1136/bmj.n71

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  PerryN.SrivastavaM.KumarD.BonehD. (2023). Do users write more insecure code with AI assistants?Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, 2785–2799.

  • Google Scholar
  • View reference in article

• 37

  PizzinelliC.PantonA.TavaresM. M.CazzanigaM.LiL. (2023). Labor market exposure to AI: Cross-country differences and distributional implications. IMF Working Paper No. 23/216.

  • Google Scholar
  • View reference in article

• 38

  Rest of World (2024). Scale AI's Remotasks is booting workers with no explanation. Rest of World, April 2. Available at: https://restofworld.org/2024/scale-ai-remotasks-banned-workers/ (Accessed April 29, 2026).

  • Google Scholar
  • View reference in article

• 39

  WebbM. (2020). The impact of artificial intelligence on the labor market. SSRN Working Paper 3482150.

  • Google Scholar
  • View reference in article

• 40

  WilesJ.ConnerM.O'BrienD. (2024). Generative AI and Content Creation Employment: Lightcast Labor Market Analysis 2020–2024. Boston, MA: Lightcast Research Report.

  • Google Scholar
  • View reference in article