Generative AI substantially narrows education-based productivity gaps: in a randomized trial, GPT-4.1 assistance closes about three-quarters of the performance gap between lower- and higher-education adults on a business problem-solving task, with some gains persisting after AI is removed but a residual human-capital gap remaining.

Main Finding

Generative-AI assistance (GPT-4.1) substantially raises performance on a realistic, incentivized business problem-solving task for adults aged 25–45, with larger gains for lower-education participants. In the absence of AI the high-education group outperforms the low-education group by 0.548 standard deviations (SD); with AI this gap falls to 0.139 SD, i.e., about a 75% closure of the baseline education gap. Follow-up, non-AI testing shows treated participants do not perform worse once AI is removed; lower-education treated participants retain part of their gain (≈0.171 SD), although an education gap re-emerges (≈0.200 SD).

Key Points

• Sample and intervention

  • N = 1,174 adults (Argentina, age 25–45); 520 classified as low-education (high school or <50% postsecondary completion), 654 as high-education (>50% postsecondary completion).
  • Randomized controlled online experiment: access to an embedded GPT-4.1 assistant (treatment) vs. no assistant (control).
  • Task: time-limited, incentivized, workplace-style business problem-solving (reading, data comprehension, reasoning, writing); followed immediately by a non-AI-assisted follow-up module (open-ended root-cause and recall questions).

• Magnitudes (standardized relative to low-education control)

  • AI increases task score by +1.242 SD for low-education individuals and by +0.834 SD for high-education individuals.
  • Baseline (no-AI) education gap = 0.548 SD (high > low); with AI the gap = 0.139 SD (closed ≈75%).
  • Follow-up (no-AI) performance: treated participants are not worse than controls; low-education treated retain about +0.171 SD; follow-up education gap ≈0.200 SD (high > low).

• Mechanisms from chat-log and engagement analysis

  • Lower-education participants obtain substantial assistance from AI (explaining large immediate gains).
  • Higher-education participants use the tool more effectively on several margins (more detailed prompts, structured workflows, better selection/integration of outputs), explaining why the gap does not fully disappear.
  • Intensive AI use predicts strong immediate performance even when engagement (time/effort) is low; however, follow-up carryover is considerably higher only when intensive AI use is combined with sustained task engagement — suggesting gains are not pure delegation.

• Interpretation

  • AI acts partly as a substitute for scarce inputs (writing, structured reasoning, synthesis) — a task-simplification or “expertise-leveling” effect — which compresses effective productivity differences in task execution.
  • Underlying human capital (formal education) still shapes unassisted performance and the effective, sustained use of the tool.

Data & Methods

• Design

  • Preregistered randomized experiment outside firms; treatment = access to an embedded GPT-4.1 assistant while completing the main task; all participants complete an immediate non-AI follow-up.
  • Incentives: up to AR$25,000 (≈US$18) in supermarket gift cards based on relative performance; payment tiers assigned by rank within treatment × education groups.
  • Sample stratified for age and gender balance within education groups; robustness checks use alternative education thresholds and exclusions.

• Outcomes and measurement

  • Main outcome: standardized task score (task graded on content and format; standardization uses low-education control as reference).
  • Follow-up outcome: performance on non-AI-assisted recall and open-ended root-cause questions.
  • Process measures: full chat logs captured; coded measures include prompt detail, number and structure of turns, intensity of AI assistance, and participant engagement (time and effort metrics).

• Analysis

  • Intention-to-treat (random assignment) estimates of treatment effects on task and follow-up performance; subgroup heterogeneity by education.
  • Decomposition of treatment heterogeneity using chat-log features and a four-way classification combining intensity of assistance and engagement.
  • Robustness checks (alternative education definitions, excluding intermediate education cases, distributional checks show treated scores remain dispersed).

• Key validity notes

  • Task is self-contained and designed to be domain-general (reduces role of firm-specific knowledge).
  • Experimental setting trades organizational realism for clean identification of between-education-group effects.
  • Sample is Argentine adults 25–45; results are short-run and task-specific.

Implications for AI Economics

• Redistribution of effective task-execution skills

  • Generative AI can compress between-education productivity gaps in task performance by substituting for certain cognitive and communication inputs, potentially democratizing some task-level capabilities.
  • This suggests AI can reduce observable skill requirements for specific tasks, with implications for who can perform particular jobs and how firms allocate tasks.

• Limits and persistence

  • Human capital remains important: education-based differences persist in unassisted performance and in effective, sustained use of AI. Task-level equalization does not eliminate underlying skill gaps.
  • Short-run gains are evident; long-run labor-market impacts depend on equilibrium adjustments (hiring, wages, task reallocation), adoption patterns, and access to AI.

• Policy and organizational considerations

  • Broad access to high-quality AI tools could reduce performance gaps in task execution; however, equitable access and training in effective AI use are critical (higher-education users extract somewhat more value).
  • Firms and policymakers should focus on complementary investments (training on prompt engineering, workflows, and evaluation of AI outputs) to translate immediate AI-assisted gains into sustained human-capital development.
  • Monitoring is needed for potential adverse dynamics: if tasks are reallocated or wage structures adjust, initial equalizing effects at the task level could have complex distributional consequences.

• Research priorities

  • Study longer-run persistence of gains (skill formation vs. delegation), multiple task types and occupational settings, heterogeneous adoption across firms and regions, and equilibrium labor-market responses (wages, employment, task composition).
  • Investigate how organizational contexts (team workflows, supervision, quality control) interact with AI-induced compression of skill requirements.

Limitations to keep in mind: an outside-firm, short-duration experiment on a single task in Argentina; results may vary across tasks, occupations, AI models, and over longer horizons.