GenAI could squeeze high-skill workers in city centers

Generative AI has started transforming urban labor markets in ways that challenge long-held assumptions about technology, education and wages. Rather than simply rewarding skilled workers, the latest evidence suggests that GenAI may compress the wage premium in high-exposure occupations, especially in dense urban cores where highly educated workers and AI-vulnerable cognitive tasks are concentrated.

A new study, titled "Generative AI impacts on intra-urban inequality and skill premium in Beijing" and posted on arXiv, uses Beijing as a detailed case study of a broader urban labor-market shift. Researchers analyzed nearly 5 million job postings from 2018 to 2024, built a neighborhood-level GenAI Exposure Index using five large language models, and tested how exposure affected wages, skill demand and spatial inequality after ChatGPT's release.

GenAI exposure is becoming an urban inequality problem

The findings point to a wider challenge for global technology hubs: GenAI exposure is unlikely to spread evenly across cities. It tends to cluster in the same places that already concentrate knowledge work, advanced services, finance, software, research and business management. These are the districts where cognitive, creative and communicative tasks are most common, and where workers are most likely to use or compete with AI systems. This pattern complicates the idea that digital technology automatically weakens geography.

Although GenAI tools can be accessed from almost anywhere, the work most exposed to them often remains tied to urban cores. Dense business districts still offer talent networks, firm clusters, knowledge spillovers, venture capital, research institutions and face-to-face interaction. As a result, AI may reinforce the economic power of city centers instead of dispersing opportunity to peripheral areas.

Beijing illustrates this pattern clearly. The study finds that GenAI exposure is concentrated in the city's high-value core, especially around Zhongguancun Science City, Financial Street and the Guomao central business district. These areas form a high-exposure zone built around innovation, capital management and business services. By contrast, lower-exposure neighborhoods are more common in outer suburban, ecological-conservation and legacy industrial zones, where traditional manufacturing and lower-skill services are more common.

The same pattern could matter in other AI hubs, including San Francisco, London, Singapore, Seoul, Tokyo, Bengaluru and Shenzhen. Cities with strong technology cores may see GenAI intensify existing spatial divides. Core districts may absorb both the benefits and disruptions of AI adoption, while peripheral neighborhoods may be partially shielded from immediate labor displacement but also excluded from productivity gains.

This creates a new kind of urban AI divide, which is not only about who has access to tools, but also about which neighborhoods have the firms, workers, infrastructure and institutional networks needed to convert those tools into economic opportunity. Areas without those assets may remain low-exposure, low-opportunity zones, even as the urban core becomes more AI-intensive.

The Beijing evidence also shows that AI diffusion can be selective. Some spillover appeared in Beijing's municipal sub-center after 2023, suggesting that emerging districts can capture AI-related activity when they have enough infrastructure and absorptive capacity. But distant peripheral areas remained weakly exposed. That suggests that AI-driven growth may not automatically spread outward without targeted policy support.

The high-skill trap challenges assumptions about education and pay

The key finding is not simply that AI exposure is concentrated in city centers. It is that high-exposure areas can attract skilled workers while wages stagnate or decline. This pattern challenges the traditional view that new technologies mainly reward highly educated labor.

For decades, many economists described technological change as skill-biased. Under that model, advanced technologies raise demand for educated workers, increase productivity and widen the wage premium for high-skill jobs. GenAI may operate differently because it reaches tasks that were previously considered hard to automate: coding, writing, translation, research, data analysis, documentation, design support and professional communication.

The study calls this emerging pattern a high-skill trap. It occurs when skilled workers continue to move into AI-exposed occupations and locations, but the wage returns to those skills weaken because GenAI lowers the scarcity value of the tasks they perform. In simple terms, the labor market still pulls educated workers into exposed sectors, but those workers no longer receive the same wage gains because AI can assist, replicate or standardize parts of their work.

In Beijing, high-exposure neighborhoods continued to attract highly educated workers, and exposure rose with average years of schooling. However, wages in the most exposed areas weakened after peaking in 2021. After ChatGPT's release, the high-exposure group fell further, reaching 13,673 yuan per month in 2024. The wage gap between high-, medium- and low-exposure areas narrowed mainly because wages in exposed areas declined.

This pattern has broader relevance. In many cities, workers in AI-exposed jobs are not necessarily low-skill workers. They include programmers, analysts, designers, copywriters, researchers, consultants, finance professionals and junior white-collar employees. These groups may face pressure not because their jobs disappear immediately, but because parts of their work become easier for others to perform with AI assistance.

The study identifies two mechanisms behind the wage pressure.

• De-skilling: GenAI can reduce the difficulty of tasks that once required advanced training. If AI tools allow less-experienced workers to perform basic coding, drafting, translation, analysis or content production, the market value of those tasks may fall. Highly educated workers remain valuable, but the tasks that once justified higher pay may become less scarce.
• Crowding: As AI-exposed fields attract more educated workers, competition increases. If more people can produce similar AI-assisted outputs, employers may have less reason to pay a premium. Workers may need to do more, learn more tools and compete harder simply to maintain their position. In crowded fields, productivity gains can therefore translate into wage pressure rather than higher pay.

Together, these mechanisms suggest that GenAI may compress wages at the top of some urban labor markets. It may help less-educated or less-experienced workers cross skill barriers, while reducing the advantage held by workers whose tasks overlap strongly with AI capabilities.

Cities need AI policies that protect workers and widen opportunity

The Beijing study strengthens the causal case by treating ChatGPT's release as a major GenAI shock. The researchers used pre-2020 exposure measures and a difference-in-differences design to compare wage changes across neighborhoods with different levels of earlier GenAI exposure. They also controlled for other major shocks, including technology-sector regulation, pandemic recovery and real-estate adjustment.

The key result is a post-shock wage penalty in more exposed neighborhoods. In the baseline model, a one-standard-deviation increase in 2018 GenAI exposure was associated with a wage decline of about 13.1% after the shock. With controls for concurrent disruptions, the estimated decline rose to about 15.1%. Pre-treatment wage trends were statistically similar, while the negative break appeared after the GenAI shock, particularly in 2023.

AI governance cannot focus only on national productivity, firm innovation or computing infrastructure. GenAI is also an urban policy issue. Its effects vary across neighborhoods, occupations and skill groups, and the early evidence suggests that highly exposed urban cores may face a difficult mix of talent concentration, task substitution, wage compression and labor-market crowding.

City government's first priority should be to prevent AI opportunity from becoming locked inside already dominant districts. AI demonstration zones, digital infrastructure subsidies, startup support and applied AI training should not be limited to established business cores. Secondary centers, suburban innovation districts and peripheral employment zones need better access to AI tools, cloud resources, technical training and industry partnerships.

The second priority is education reform. Generic reskilling programs may not be enough if they train workers in codifiable skills that GenAI can quickly assist or replicate. Education and workforce programs need to focus more strongly on complex problem-solving, domain expertise, critical reasoning, team coordination, human judgment, ethics, negotiation and cross-cultural communication. These skills are harder to automate and more likely to remain valuable as AI tools spread.

The third priority is labor protection for exposed occupations. Highly educated workers in AI-exposed fields may need transition support just as much as lower-skilled workers affected by earlier automation waves. Cities and national governments may need wage insurance, re-employment funds, skill-transformation subsidies and stronger protections for entry-level white-collar workers whose career ladders could narrow as AI handles routine professional tasks.

The fourth priority is better labor-market monitoring. The Beijing study shows the value of fine-grained job-posting data, neighborhood-level analysis and task-level AI exposure measures. Other cities will need similar systems to track which occupations are being reshaped, where wage pressure is emerging, and whether AI adoption is increasing or reducing inequality.

The findings do not mean that GenAI will permanently reduce wages for all skilled workers. New AI-complementary jobs may emerge, firms may reorganize work, and workers may shift toward tasks that require deeper expertise. However, the early evidence suggests that the transition will not be frictionless. The wage premium attached to education and cognitive skill can weaken when AI reduces the scarcity of those skills.