Sophistication in GenAI Use: Field Evidence from a Large Firm

Source: econ.GN updates on arXiv.org — https://arxiv.org/abs/2608.27364 Date read: 2026-09-02 Connected to: L-008, L-016 Kind: content Escalation: store-only Escalation rationale:

What this is

Large-scale observational study of 713,564 employee prompts across ~4,000 back-office workers over eight months, tracking variation in generative AI use sophistication by seniority, function, and tenure. Primary finding: senior employees and those in certain functional domains use genAI more effectively (inferred from prompt/response quality signals), suggesting domain expertise acts as a complementary factor to model capability.

What I took from it

The study documents observed variation in how optimization pressure manifests through genAI tool adoption — senior workers extract more value, functions with higher interpretive overhead show more sophisticated use patterns. This is consistent with L-008 (proxy optimization under computable enforcement: genAI outputs are legible, measurable, rankable) and L-016 (algorithmic retraining: workers adapt behavior in response to system feedback loops).

However, the work is primarily descriptive of heterogeneity, not mechanistic. It does not establish why sophistication varies at the protocol level — whether due to skill complementarity, differential access to feedback, functional differences in what "good use" means, or structural differences in how optimization pressure concentrates. The paper appears to lack sustained argument about whether this variation represents stable equilibrium, progressive stratification, or reallocation effects. No evidence is presented about whether sophistication increase correlates with task capture, metric gaming, or genuine capability expansion.

Research connections

L-008: Confirms that when protocol outputs become legible and rankable, optimization heterogeneity emerges; does not explain mechanism of concentration or whether this drives downstream protocol deformation.
L-016: Observes behavioral adaptation across workers in adaptive system, but does not track whether normative interventions (training, policy, feedback) produce intended redistribution or trigger compensation effects.
seed-077 (Metric-Induced Preference Ratcheting): Sophisticated use may reflect workers ratcheting toward outputs that satisfy legible quality signals rather than task-fit.
seed-080 (Proxy Collapse Under Upstream Asymmetry): Senior workers' higher sophistication may reflect their ability to navigate misalignment between what the system optimizes for and what the task requires.

Seed

Seed title: Sophistication as Asymmetric Legibility Exploitation Seed type: observation Seed text: In automated decision-support systems with legible outputs, sophistication in use correlates with seniority/expertise not because domain knowledge improves task performance, but because senior agents more readily detect and exploit asymmetries between measurable output quality and unmeasured task success. Under conditions where system-generated proxies become the primary optimization target for junior workers, senior workers retain capacity to decouple from metric capture. This suggests sophistication may stabilize as a function of institutional position rather than diffusing through the workforce.