Firms exposed to generative AI recorded modest but significant productivity gains after 2023, amplified in intangible-rich firms, and these gains were accompanied by a drop in labor’s share—suggesting redistributed rents as AI diffuses.

Generative AI Adoption and Firm-Level TFP Growth: Diffusion, Complementarities, and Rent Sharing in US Listed Firms

Alex Wanninger · February 19, 2026 · Journal of Economic Development & Global Markets

openalex quasi_experimental medium evidence 8/10 relevance Summary only summary available; pdf_status=paywall DOI Source PDF

Structured author observations

Linked only from stored provider relations; the raw author line above is never matched by name.

OpenAlex

Latest observation: August 25, 2026

Alex Wanninger provider ID

Semantic Scholar

Latest observation: August 25, 2026

Alex Wanninger provider ID

Firms more exposed to generative AI experienced statistically significant increases in TFP after 2023 (≈0.017) with larger gains for intangible-intensive firms (additional ≈0.011), while exposed firms also saw declines in labor share (≈-0.007).

Citation observations

Cumulative provider counts captured on specific dates; providers are never combined.

0 cumulative citations

OpenAlex · Observed August 25, 2026

View corpus context

0 cumulative citations

Semantic Scholar · Observed August 25, 2026

View corpus context

This paper studies whether the diffusion of generative AI (LLMs) is associated with faster firmlevel productivity growth and how complementary assets shape the gains.Using US listed non-financial firms over 2018-2024 (2,318 firms; 16,226 firm-years), we construct (i) a predetermined industry-level LLM exposure index and (ii) a firm-level adoption proxy from EDGAR 10-K disclosures based on LLM-related keyword intensity.We estimate a differencein-differences specification around the post-2022 generative AI shock.Firms with higher exposure exhibit significantly higher TFP after 2023 ( = 0.017, SE = 0.006).The effect is stronger among intangible-intensive firms (additional = 0.011, SE = 0.004).At the same time, exposed firms display a decline in labor share ( = -0.007,SE = 0.002), consistent with a shift in rent sharing.The results highlight growth gains that are amplified by intangible capital and accompanied by distributional changes.

Summary

Main Finding

Firms more exposed to generative AI (LLMs) experienced faster productivity growth after 2023: a positive effect on firm-level TFP (coefficient = 0.017, SE = 0.006). The productivity gain is larger for firms with high intangible capital (additional effect = 0.011, SE = 0.004). At the same time, exposed firms show a decline in labor share (coefficient = -0.007, SE = 0.002), consistent with a shift in rent sharing away from labor.

(Interpretation: the TFP coefficient of 0.017 is roughly a 1.7% increase in TFP if TFP is measured in logs; the intangible interaction adds ~1.1 percentage points. The labor-share decline of -0.007 is about a 0.7 percentage-point reduction.)

Key Points

Exposure and adoption measures:
- Industry-level LLM exposure: a predetermined index (industry-based) used to capture differential technological susceptibility across industries.
- Firm-level adoption proxy: intensity of LLM-related keywords in EDGAR 10-K disclosures.
Empirical design: difference-in-differences around the post-2022 generative-AI shock (comparison of pre- and post-2023 outcomes by exposure/adoption).
Main estimates:
- TFP increase for higher-exposure firms: 0.017 (SE 0.006), statistically significant.
- Additional TFP gain for intangible-intensive firms: +0.011 (SE 0.004).
- Decline in labor share for exposed firms: -0.007 (SE 0.002).
Interpretation: generative AI adoption/exposure is associated with productivity gains that are amplified by complementary intangible assets, and these gains are accompanied by distributional shifts reducing labor’s share of income.

Data & Methods

Sample: US-listed, non-financial firms, 2018–2024; 2,318 firms, 16,226 firm-year observations.
Measures:
- Predetermined industry-level LLM exposure index (constructed to avoid endogeneity from contemporaneous firm choices).
- Firm-level LLM-adoption proxy based on keyword intensity in 10-K filings (EDGAR).
- Outcome variables include firm TFP and labor share.
Identification:
- Difference-in-differences specification exploiting the post-2022 generative-AI shock timing and heterogeneous cross-industry exposure.
- Interaction with intangible-intensity to test complementarity.
Statistical precision: reported coefficients come with standard errors (TFP: 0.017, SE 0.006; intangible interaction: 0.011, SE 0.004; labor share: -0.007, SE 0.002), all significant at conventional levels.

Implications for AI Economics

Productivity effects:
- Generative AI is associated with measurable firm-level productivity gains soon after diffusion; these gains are concentrated where complementary intangible capital exists.
- Intangible assets (software, R&D, organizational capital) amplify returns to AI, implying complementarities matter for cross-firm heterogeneity in benefits.
Distributional effects:
- The decline in labor share suggests AI-driven productivity gains may accrue disproportionately to capital owners or other non-labor claimants, raising concerns about rising inequality within firms/industries.
Policy and firm strategy:
- Policies should support worker retraining and facilitate labor reallocation to mitigate adverse distributional impacts.
- Encouraging investment in intangible capital (and easing firms’ ability to adopt AI) can magnify aggregate gains, but also risks winner-take-most dynamics.
- Disclosure-based adoption measures matter for monitoring diffusion—regulators and analysts may use filings to track AI uptake.
Research directions:
- Further work should unpack mechanisms (task displacement vs. augmentation), explore heterogeneity by firm size and occupation mix, and examine long-run effects on wages, rents, and market structure.
- Causal identification could be strengthened with additional instruments, firm-level adoption data, or experimental/quasi-experimental variation in AI rollouts.

Caveats: the adoption proxy (10-K keyword intensity) may reflect disclosure behavior as well as actual use; the DiD design mitigates but does not fully eliminate potential confounders.

Assessment

Paper Typequasi_experimental Evidence Strengthmedium — Uses a plausibly exogenous timing shock (post-2022 generative-AI diffusion) and a predetermined industry exposure index to support causal claims, with a large panel of listed firms and DiD estimation; however, firm-level adoption is proxied by 10-K disclosures (which may reflect awareness/hype rather than use), potential violations of parallel trends or omitted time-varying confounders remain possible, and TFP measurement noise can attenuate or bias estimates. Methods Rigormedium — Standard, appropriate empirical strategy (panel DiD, predetermined exposure, heterogeneity analysis by intangible intensity) and a large sample strengthen credibility, but the paper appears vulnerable to measurement error in adoption proxies, possible endogenous disclosure, limited discussion (here) of parallel-trends tests, placebo checks, or alternative identification (e.g., instruments or firm-level adoption timing), which keeps rigor from being rated high. SampleUS-listed non-financial firms, 2018–2024 sample including 2,318 firms and 16,226 firm-year observations; firm-level outcomes include TFP and labor share; key regressors are an industry-level predetermined LLM exposure index and a firm-level 10-K keyword intensity adoption proxy. Themesproductivity labor_markets IdentificationDifference-in-differences comparing firms with higher vs lower predetermined industry-level LLM exposure around the 2023 generative-AI shock, supplemented by a firm-level adoption proxy constructed from EDGAR 10-K LLM-related keyword intensity; includes interactions with firm intangible intensity. GeneralizabilityRestricted to US publicly listed, non-financial firms—may not apply to private, small, or financial firms, Post-treatment window is short (post-2022/2023), limiting inference about long-run effects, Industry-level exposure index may misclassify firm-level usage patterns, 10-K keyword intensity is an imperfect proxy for actual LLM adoption (may capture disclosure behavior or hype), TFP measurement challenges and firm accounting heterogeneity may limit external validity

Claims (5)

Claim	Direction	Outcome	Confidence & Evidence	Details
Firms with higher industry-level LLM exposure exhibit significantly higher total factor productivity (TFP) after 2023. Firm Productivity	positive	Total factor productivity (TFP)	Reading fidelity high Study strength medium	n=2318 0.017 (SE = 0.006) 0.48
The positive TFP effect of LLM exposure is stronger among intangible-intensive firms (an additional effect for intangible intensity). Firm Productivity	positive	Total factor productivity (TFP) — heterogeneous (intangible intensity) effect	Reading fidelity high Study strength medium	n=2318 0.011 (SE = 0.004) 0.48
Firms more exposed to generative AI show a decline in labor share after adoption/exposure. Labor Share	negative	Labor share	Reading fidelity high Study strength medium	n=2318 -0.007 (SE = 0.002) 0.48
The paper constructs a predetermined industry-level LLM exposure index and a firm-level adoption proxy from EDGAR 10-K disclosures based on LLM-related keyword intensity. Adoption Rate	null_result	LLM exposure / firm-level adoption proxy (10-K keyword intensity)	Reading fidelity high Study strength high	n=2318 0.8
Estimates are obtained using a difference-in-differences specification around the post-2022 generative AI shock. Other	null_result	N/A (methodological claim about identification strategy)	Reading fidelity high Study strength high	n=2318 0.8