Technical Summary: Generative AI Impacts on Intra-urban Inequality and Skill Premium in Beijing

Problem Statement

Generative Artificial Intelligence (GenAI) represents the first automation wave capable of performing high-cognitive tasks at scale, yet its specific effects on intra-urban inequality and the skill premium remain largely unknown. While previous technological waves (e.g., computerization, industrial robots) reinforced spatial polarization by clustering high-value jobs in city cores, it is unclear whether GenAI will act as a "great leveler" dispersing opportunity to urban peripheries or deepen core–periphery divides. Existing literature suffers from three critical gaps: (1) a lack of fine-grained, intra-urban spatial analysis, as most studies are national-scale and cross-sectional; (2) a scarcity of dynamic evidence linking exposure to synchronous changes in skill supply and wages, particularly using causal identification strategies like natural experiments; and (3) reliance on single large language models (LLMs) or expert elicitation to measure exposure, which introduces model-specific bias. This study addresses these gaps by investigating the spatial, structural, and causal mechanisms of the GenAI shock in Beijing, a global tech hub with a pronounced core–periphery structure.

Methodology

Data and Study Area

The study utilizes a unique dataset of 4,995,615 online job postings from Beijing spanning 2018 to 2024, sourced from the RESSET Enterprise Big Data Platform. These postings are matched to Beijing's neighborhood-level administrative units (regulatory planning units). The dataset is enriched with multi-source geospatial data (nightlight intensity, NDVI, POI density, land-use ratio) and demographic controls.

Measurement of GenAI Exposure

To overcome the limitations of single-model assessments, the authors developed a Retrieval-Augmented Generation (RAG) pipeline:

1. Occupational Mapping: Job titles and descriptions were mapped to China's 2022 Occupational Classification Dictionary (1,605 standard occupations) using semantic vector retrieval (Qwen3-Embedding-4B), achieving >93% accuracy.
2. Ensemble Exposure Scoring: Five state-of-the-art LLMs (ChatGPT-4o, Gemini-2.5-pro, Claude-3.5-Sonnet, GLM-4, Deepseek-R1) independently assessed each occupation's exposure level (ranging from No Exposure to Multimodal Capability Exposure).
3. Aggregation: The final GenAI Exposure Index for each neighborhood was constructed by aggregating the mean assessments from the five models, reducing individual model bias.

Causal Identification Strategy

To isolate the causal effect of GenAI from concurrent shocks (e.g., tech regulation, pandemic recovery, real estate downturn), the authors employ a Pre-determined Difference-in-Differences (DID) design:

• Treatment Variable: GenAI exposure measured in 2018 (pre-determined, four years before ChatGPT's release in late 2022). This rules out reverse causality and anticipatory responses.
• Post-Treatment Period: 2023–2024.
• Controls: The model interacts the 2018 exposure with a post-2022 dummy, while controlling for concurrent shocks (also interacted with the post-period) and neighborhood/time fixed effects.
• Robustness Checks:
  • Event Study: Validates the parallel trends assumption using pre-treatment coefficients (2020–2021).
  • Randomization Inference: Permutation tests (500 iterations) to confirm the wage penalty is not a statistical artifact.
  • Instrumental Variable (IV): A Bartik shift-share IV constructed using national AI job growth weighted by 2018 local industrial composition.
  • Mechanism Analysis: Triple-DID specifications and interaction terms to test "de-skilling" (education moderation) and "crowding" (job-market heat) channels.

Key Results

1. Spatial Lock-in and the "Golden Triangle"

GenAI exposure is not evenly distributed but is spatially locked into Beijing's core districts. High-exposure areas form a "golden triangle" comprising Zhongguancun Science City, Financial Street, and the Guomao CBD, where exposure scores consistently exceed 0.3. This concentration intensifies the intra-urban AI divide. While a selective spillover occurred in municipal sub-centers post-2023, outer suburban counties remain low-exposure, creating a "spatial poverty trap" in low-exposure zones and a "high-skill trap" in the core.

2. The "High-Skill Trap" and Wage Stagnation

Contrary to the Skill-Biased Technological Change (SBTC) hypothesis, high-exposure neighborhoods exhibit a decoupling of skill supply and wages:

• Talent Inflow: These areas continue to attract highly educated workers (average schooling >12 years).
• Wage Penalty: Despite the influx of talent, wages in high-exposure neighborhoods stagnated and declined after 2022. Post-ChatGPT, high-exposure wages fell to ~13,673 CNY/month, narrowing the gap with lower-exposure areas.
• Causal Evidence: The DID analysis reveals that a one-standard-deviation increase in pre-determined GenAI exposure is associated with a ~13.1% to 17.5% decline in wages post-shock. This effect remains robust after controlling for concurrent economic shocks.

3. Mechanisms: De-skilling and Crowding

The wage penalty is driven by two reinforcing channels:

• Task De-skilling: GenAI lowers the cognitive threshold for tasks previously reserved for high-skilled workers (e.g., coding, copywriting, data analysis). Regression results show a significant negative interaction between GenAI exposure and education levels; for the most highly educated workers, GenAI acts as a substitute rather than a complement, eroding their bargaining power.
• Labor-Market Crowding: The steady inflow of educated talent into AI-exposed fields, combined with compressed task spaces, creates an oversupply of workers relative to high-value tasks. High "job-market heat" significantly suppresses the GenAI skill premium, confirming a "crowding-driven wage penalty."

Significance and Contributions

The paper claims four primary contributions to the literature and policy discourse:

1. Measurement Innovation: It provides the first fine-grained, neighborhood-level map of GenAI exposure, utilizing a multi-model ensemble approach to mitigate bias and offering an intra-urban perspective previously missing from national-scale studies.
2. Causal Identification: By leveraging the quasi-exogenous shock of ChatGPT's release and pre-determined 2018 exposure, the study constructs a rigorous DID design supported by randomization inference and IV strategies, moving beyond correlation to causal interpretation.
3. Theoretical Challenge: The findings challenge the prevailing SBTC theory, documenting a "high-skill trap" where technological exposure suppresses wage growth for high-skilled workers through de-skilling and crowding, rather than uniformly increasing the skill premium.
4. Policy Implications: The study offers an empirical basis for inclusive AI governance in global tech hubs. It suggests that policymakers should move beyond generic infrastructure or skills training to address spatial monopolies (e.g., establishing AI zones in sub-centers), pivot higher education toward non-codifiable skills (collaboration, critical reasoning), and implement targeted labor protections for highly exposed occupations to prevent talent outflows.

The authors conclude that while GenAI may not eliminate jobs wholesale, it fundamentally restructures the logic of returns to education and agglomeration, creating a new form of intra-urban inequality that requires specific governance interventions.