Sunil Venkataram
Back to garden
•
🌰 seedling
AI as Seniority-Biased Technological Change
type/noteaieconomicslabor

AI as Seniority-Biased Technological Change

The paradox

Two bodies of evidence point in opposite directions.

Micro studies: AI equalizes

In controlled experiments, the least experienced workers see the largest gains:

  • Brynjolfsson, Li, Raymond (2025): Less experienced customer-support agents gained 30-35% in productivity; highly skilled agents saw minimal gains or slight quality declines
  • Cruces et al. (2026): AI closed 75% of the education-based productivity gap in a business problem-solving task (1.242 SD gain for lower-education vs 0.834 SD for higher-education)
  • Noy & Zhang (2023): Lower-performing professional writers saw the largest quality and speed improvements
  • Cui et al. (2025): Larger effects for less experienced and more junior developers across three RCTs

The pattern makes intuitive sense. AI substitutes for skills that less-experienced workers lack — pattern recognition, domain knowledge retrieval, structured reasoning. It acts as a “floor raiser” in controlled settings.

Macro data: AI concentrates

In the real world, adoption skews heavily toward the already-advantaged:

  • Anthropic Economic Index: AI usage concentrates among middle-to-upper wage white-collar workers and tasks requiring more education
  • BCG Survey (2025): Managers use AI at nearly twice the rate of frontline workers; only 36% of workers feel properly trained
  • Bick, Blandin, Deming (2024): Among the 40% of Americans who use generative AI, usage correlates with education and income

Why the divergence

The decision to adopt AI involves different factors than the productivity gain conditional on adoption:

  1. Awareness and access. Higher-skilled workers are more likely to know about AI tools, have employer-provided access, and receive training.
  2. Task recognition. Identifying which tasks benefit from AI assistance requires meta-cognitive skill — understanding both the task structure and the AI’s capabilities. Experienced workers are better at this matching problem.
  3. Organizational permission. Senior workers have more autonomy to experiment with new tools. Junior workers face more surveillance, standardized workflows, and institutional friction.
  4. Complementary skills. Effective AI use requires verification, judgment, and prompt engineering — skills that correlate with experience. The METR study (Becker et al. 2025) found AI made experienced open-source developers 19% slower, suggesting even experienced users struggle with integration — but they at least attempt it.

The labor market consequence

The divergence between who benefits and who adopts produces a specific outcome: AI functions as seniority-biased technological change.

For senior workers: Productivity augmentation. More output per hour, expanded scope of individual contribution, justification for higher compensation.

For junior workers: Demand reduction. If a senior worker augmented by AI can do what previously required one senior and two juniors, the juniors do not get hired. Brynjolfsson, Chandar, and Chen (2025) document this directly — early-career workers (ages 22-25) in the most AI-exposed occupations show 15-16% relative employment declines. Hosseini Maasoum and Lichtinger (2025) confirm: junior employment drops sharply in AI-adopting firms, driven by slower hiring rather than separations.

This creates a structural problem for the labor market pipeline. Junior roles serve a dual function: producing output and training the next generation. If AI substitutes for junior output without substituting for junior training, the pipeline of experienced workers thins over time.

Counter-argument

Shen and Tamkin (2025) provide a related but distinct concern: in their study, AI-assisted learners completed tasks faster but scored 17 percentage points lower on subsequent assessments. AI accelerated task completion while undermining the skill acquisition that makes future tasks possible. If this generalizes, AI-augmented senior workers may deplete their own skill development in the long run — though the practical significance of this for already-experienced workers is debatable.

The equalizing effect could still dominate if organizations implement top-down AI training, standardize tool access, and actively push adoption toward less-skilled workers. Several of Alex Imas’s four explanations for the micro-macro disconnect point toward this resolution. But absent deliberate intervention, market forces favor seniority-biased adoption.

Backlinks (4)
Connected Notes