Revisiting AI and inequality

After completing my research paper, “AI and Inequality in the U.S.: Evidence from the American Community Survey Data” in 2025, I have had the opportunity to observe another year of rapid AI development. This provides a useful opportunity to evaluate whether the paper’s assumptions and conclusions remain valid.

Since late 2022, artificial intelligence (AI) has developed rapidly and become increasingly capable of automating both routine and cognitive tasks, such as the release of Sora 2 and GPT image 2. These advances have raised concerns that AI may reshape labour markets by replacing some workers while increasing the productivity and earnings of others, potentially widening existing economic inequality. This scenario mirrors the Industrial Revolution, when steam-powered automation replaced many physical laborers.

AI exposure is a key factor measured in the investigation into the degree of impact of AI on jobs, where my score follows Daron Acemoglu’s paper. Occupations with low AI exposure are less affected by AI, while occupations with high AI exposure can benefit from greater productivity improvements.

The U.S. was chosen as the focus of the research because it exhibits long-standing inequalities in income, race, and gender, making it an appropriate case for examining the distributional effects of AI. The ACS is an ongoing American community survey conducted by the Census Bureau, and it collects data from 3 million households each year. Therefore, studying inequality in the U.S. addresses a classic and persisting issue with support from reliable data.

Through analysis and regression, the variations between genders, race, and nativity are investigated while processing factors such as age. African Americans had a negative coefficient following AI exposure (-0.0286 for p<0.01), which indicates that for each unit increase in AI exposure, Black individuals earn 0.0286 log points less in income growth than non-Blacks. This adds to existing disparities in earnings between African Americans and other races. Latin Americans in the U.S. have close results but slightly narrower gaps. Similar to gender inequality, males had a positive interaction term to AI exposure (0.0124) and baseline advantage as well, indicating an increase in disparities in income for males and non-males through AI and automation in the U.S. On the analysis of nativity, there is a baseline disadvantage for non-natives in occupations, but the existence of AI tools narrowed the income gap between nativity, likely because of the removal of partial language barriers.

A common economic approach to reducing structural inequality is through supply-side policies, such as improving workers' skills and providing training to make them less vulnerable to AI-driven automation. Subsidized training for disadvantaged neighborhoods reduces inequality in the long run, heavily relying on government budget or charity support. This raises a need for short-term fiscal policies, increasing progressive taxation and taxes on the commercial use of AI slows down the takeover of AI in many jobs while supporting the government budget. However, it can be argued that AI-specific taxes are difficult to implement as it faces strong opposition from technology industries and high-income earners.

Reviewing the research, the paper still has limitations from less complete statistical measures, as it only assesses data from 2018 to 2023. When writing the paper, I assumed that occupations with low AI exposure would remain relatively resistant to automation. As mentioned, recent advances in AI capabilities have occurred at a pace that outstrips human skill development. 5 years ago, auto driving was still testing and premature since its earliest attempt in the 1960s. Today, autonomous driving systems have become significantly more advanced, with companies such as Tesla demonstrating increasingly capable self-driving technologies, and recent developments of Robo-run in China replace partial tasks of taxi drivers. Unitree robot is recently developing automated robots to clean houses and rooms that perform tasks of room attendants, which is one of the least exposed occupations listed in the paper (1.52) and once thought unlikely to be replaced.

Looking back, I also underestimated how quickly AI would become capable of performing creative tasks. In 2025, image generation, programming assistants and reasoning models improved far beyond my expectations. Rather than only increasing productivity, AI has begun competing directly with some professional and creative occupations, suggesting that future research should distinguish between complementary and substitutive forms of AI.

Overall, I believe the paper still provides useful evidence that AI can reinforce existing inequalities across race, gender, and income. However, its conclusions are impacted by the context of the data available at the time. AI capabilities are evolving much faster than labour market datasets can be updated; future research should continually reassess which occupations are vulnerable to automation rather than assuming current patterns will remain stable.

The writer is Year A1 of Shenzhen College of International Education (SCIE).