Objective: This study investigates whether access to Generative Artificial Intelligence (GenAI) – specifically large language models – improves the labor productivity of knowledge workers, and examines how task complexity, worker skill level, and automation bias moderate or mediate this effect.

Methods: We employed a mixed-methods design including (1) a pre-registered randomized controlled trial (RCT) with 320 knowledge workers from three Chinese technology companies over four weeks, comparing a treatment group with GPT-4 access to a control group without AI; (2) a two-wave longitudinal survey of 1,103 knowledge workers who used GenAI weekly; and (3) qualitative follow-up interviews. Objective productivity was measured via task-tracking systems. Hypotheses were tested using ANCOVA, mixed-effects models with quadratic terms, and structural equation modeling (SEM) with bootstrapped confidence intervals.

Results: GenAI access increased average productivity by 23.7% (p<0.001, η²=0.37). However, task complexity exhibited an inverted-U moderation: medium-complexity tasks gained 34.7% (p<0.001), while high-complexity tasks showed a 14.2% productivity loss (p=0.01). Worker skill heterogeneity was pronounced: mid-performers gained 41.2% (p<0.001), low-performers gained 13.8% (p=0.010), and top-performers showed no gain (-2.5%, p=0.680). Automation bias mediated the relationship, offsetting 18.9% of potential productivity gains, with 31% of users reporting uncritical acceptance of AI outputs.

Conclusions: Generative AI improves knowledge worker productivity on average, but this effect is highly conditional. Universal AI mandates are counterproductive; organizations should target medium-complexity tasks, mid-level performers, and implement verification protocols to counter automation bias. The findings support a contingent task-skill-technology fit framework.