The Privilege of Exposure: Caste and Generative AI in India's Graduate Labour Market

In India, graduates from marginalized castes (SC and ST) are significantly less likely to work in jobs that can be enhanced by AI. Because these AI-ready jobs also command much higher wages, the rise of generative AI is predicted to widen the existing wealth gap between different caste groups.

Most global discussions on artificial intelligence focus on "displacement risk" (the fear that AI will replace humans). This study shifts the focus toward "augmentation potential" (the capacity for AI to enhance human productivity). In many developed economies, the debate centers on whether AI will automate clerical tasks or empower professionals. However, this framework assumes a level playing field regarding who occupies those high-skill roles.

A new study from the Institute for Social and Economic Change explores how social hierarchies intersect with technological shifts. By mapping AI exposure to India’s graduate labor market, the author demonstrates that exposure is a marker of social privilege.

The hidden gradient in AI exposure

Current literature measures the "task content" of an occupation. This refers to the specific duties a worker performs. These are scored against the known capabilities of large language models (LLMs). This approach identifies which jobs are susceptible to being augmented (enhanced by AI) or automated (replaced by AI). While effective at identifying sectoral risks, these models often overlook how social stratification dictates who holds those jobs.

In India, caste remains a primary determinant of occupational attainment. The study finds that the "displacement" narrative misses a crucial reality. For many marginalized graduates, the risk is not being replaced by an algorithm. Instead, they face exclusion from the digital economy entirely. As shown in, the share of Scheduled Caste (SC) and Scheduled Tribe (ST) graduates in the workforce drops steadily as AI exposure increases.

While software developers represent a high-exposure tier, they are dominated by upper-caste graduates. Conversely, construction laborers—who have near-zero AI exposure—contain a much higher proportion of SC/ST workers.

Mapping privilege through occupational sorting

To quantify this gap, the researcher linked international AI exposure indices to India’s local occupational framework. The process followed three main stages:

1. Index Crosswalking: The study utilizes three distinct indices. These include the ability-based AIOE (measuring high-level cognitive tasks) and the task-based GPT exposure (measuring how much an LLM can halve task completion time). These were mapped to India’s National Classification of Occupations (NCO-2015).
2. Data Integration: These scores were merged with the 2025 redesigned Periodic Labour Force Survey (PLFS). This provided a granular view of approximately 83,000 employed graduates.
3. Decomposition Analysis: The study separates the "gap" into two sociological channels: occupational relegation and intra-white-collar sorting.

The first channel, relegation, describes graduates forced into "elementary" occupations. These include farming or manual labor. Such roles require no degree and offer zero AI interaction. The second channel, sorting, describes the distribution within professional sectors. Even when excluding manual labor, the paper finds that upper-caste graduates are disproportionately represented in high-exposure clusters like software, finance, and management.

Evidence of a widening wage gap

The most striking finding is that AI exposure acts as a potent driver of income inequality. The paper reports that exposure to high-skill, augmentable tasks carries a significant wage premium. Specifically, a one standard deviation increase in ability-based exposure (AIOE) is associated with a 20.0 log point increase in monthly earnings. This represents a substantial boost to individual earning power.

The authors illustrate this relationship across the entire income spectrum in .

The "exposure premium" is not uniform. It peaks at the 40th percentile of the wage distribution. This is where junior professional and clerical roles reside. This is the precise segment of the labor market from which SC and ST graduates are systematically absent.

Because the occupations that provide the greatest "AI boost" to wages are mostly held by upper-caste graduates, the technology may compound existing disparities. The authors report that SC graduates are 0.24–0.27 standard deviations less exposed than upper-caste graduates. Meanwhile, ST graduates lag by 0.31–0.37 standard deviations.

Limitations of the counterfactual model

Several caveats limit the scope of these findings. First, the study measures "exposure" as a counterfactual incidence. This describes what would happen if AI were adopted. It does not measure actual, real-world adoption rates. In many parts of India, the infrastructure for widespread generative AI integration is still maturing.

Second, there is a risk of measurement error due to geographical translation. Most exposure indices use US-centric occupational data. The authors acknowledge that a software developer in San Francisco might perform different tasks than one in Bengaluru. However, the high correlation between different indices suggests this effect is minimized. Finally, the use of 3-digit occupational coding averages over many different roles. The authors admit this may slightly dampen the observed gradients.

The verdict: A tool for entrenching hierarchy

Is generative AI a democratizing force or a divider? The answer depends on institutional access. The study suggests that without intervention, AI will function as a "privilege multiplier."

To prevent a widening caste gap, the research implies a need to expand access to high-skill pipelines. This includes moving graduates into the professional sectors where AI's value is realized. Currently, the public sector offers the most parity for disadvantaged groups. However, these roles are largely insulated from the technologies driving modern wage growth. Code for the occupational crosswalk used in this study is reportedly available at github.com/Kpmishra1998/nco2015-ai-exposure.