Economic Incentives, Artificial Intelligence, and Labor Productivity

Participants will be randomly assigned to different economic incentive conditions when completing several cognitive tasks. The interventions vary two dimensions: the payment scheme participants face for task performance, and whether and under what pricing condition they can use an AI assistant during the tasks.

2026-05-24

2026-12-31

The primary outcomes are organized into three families: AI adoption, AI use behavior, and labor productivity.

AI adoption measures whether participants choose to use AI when AI is available. The main adoption outcome is a task-level indicator equal to one if the participant uses the AI assistant during a given task. We will also construct participant-level measures, including whether the participant uses AI at least once and the number or share of tasks in which AI is used.

AI use behavior measures how participants interact with AI conditional on AI access and, when relevant, conditional on actual AI use. Using AI interaction logs and text-analysis methods, we will construct measures in three main dimensions: participants’ prompting or question-asking style; the similarity between participants’ prompts and the original task materials; the similarity between participants’ final submissions and AI-generated responses. Based on these measures, we may further construct classifications of individual AI-use patterns, such as direct delegation, selective consultation, verification, editing, or other modes of human-AI interaction.

Labor productivity measures participants’ task performance in terms of both quality and quantity. It will be evaluated using task-specific scoring rubrics, including an objective rubric that measures core task performance, an extended rubric that incorporates additions and deductions for answer quality, and a measure of output quantity. Task outputs will be evaluated through a human-AI cross-scoring procedure based on pre-specified rubrics. Scores will be standardized within task before pooled analysis. We will report both component outcomes and, where appropriate, standardized productivity indices constructed from these components.

Secondary outcomes include:
- Participants’ beliefs about their own performance, AI performance, and peers’ AI use;
- Perceived task difficulty, time pressure, and prior familiarity with similar tasks;
- Time-related task measures, including completion time, time remaining at submission, and output per unit of time;
- Timing of AI adoption, including whether participants use AI early or late in a task, the time elapsed before first AI use, and the interval between last AI use and final submission;
- AI interaction intensity, including number of AI interactions, prompt length, response length, and number of conversation turns;
- Classifications of individual AI-use patterns or interaction types based on AI interaction logs;
- Heterogeneity by baseline characteristics, including AI familiarity, baseline ability, risk preferences, and demographic variables.

This study is a randomized laboratory experiment. Participants complete several cognitive tasks under randomly assigned payment schemes and AI access or pricing conditions. The experiment is designed to study how economic incentives affect participants’ willingness to use AI, how they interact with AI, and the labor productivity effect of AI.

Not available

Randomization will be implemented by the experimental platform using computer-generated random assignment. The assignment algorithm will use pre-specified allocation probabilities to approximate the planned sample sizes across the treatment arms.

The unit of randomization is the individual participant. Each participant is assigned to one payment scheme and one AI access/pricing condition.

No

Approximately 1,200 individual participants. The treatment is not clustered.

Approximately 1,200 individual participants. Since each participant completes three tasks, the study will generate approximately 3,600 participant-task observations.

The planned sample includes approximately 1,200 participants. Within each of the three payment schemes, we plan to randomly assign approximately 100 participants to the no-AI condition, 120 participants to the paid-AI condition, 120 participants to the free-AI condition, and 60 participants to the subsidized-AI condition. This yields approximately 400 participants under each payment scheme.

The planned sample size is approximately 1,200 participants. Each participant completes three tasks, so the main analyses will be conducted at the participant-task level. Treatment is assigned at the participant level, and standard errors will be clustered at the participant level. The calculations below use a two-sided 5% significance level and 80% power. For task-level outcomes, the minimum detectable effects depend on the within-participant correlation across tasks. As a benchmark, we report calculations assuming an intra-participant correlation of 0.5; a fully conservative benchmark with intra-participant correlation equal to 1 would yield larger MDEs. The design varies two dimensions of economic incentives: payment schemes and AI access/pricing conditions. For standardized continuous labor-productivity outcomes, pairwise comparisons between two payment schemes use approximately 400 participants per payment scheme and can detect an effect of approximately 0.16 standard deviations under the benchmark assumption, or approximately 0.20 standard deviations under the conservative assumption. For AI access/pricing conditions, comparisons between the paid-AI and free-AI conditions use approximately 360 participants per condition and can detect an effect of approximately 0.17 standard deviations in standardized labor-productivity outcomes under the benchmark assumption, or 0.21 standard deviations under the conservative assumption. Comparisons between the no-AI condition and either the paid-AI or free-AI condition use approximately 300 and 360 participants respectively and can detect an effect of approximately 0.18 standard deviations under the benchmark assumption, or 0.22 standard deviations under the conservative assumption. Comparisons involving the subsidized-AI condition have lower power because of its smaller sample size; comparisons between the subsidized-AI condition and either the paid-AI or free-AI condition can detect effects of approximately 0.21 standard deviations under the benchmark assumption, or 0.26 standard deviations under the conservative assumption. AI adoption is measured at the participant-task level among participants assigned to AI-access conditions. For binary AI adoption outcomes, assuming a baseline adoption rate of 50%, pairwise comparisons between two payment schemes among AI-access participants use approximately 300 participants per payment scheme and can detect a difference of approximately 9.3 percentage points under the benchmark assumption, or 11.4 percentage points under the conservative assumption. Comparisons between the paid-AI and free-AI conditions use approximately 360 participants per condition and can detect a difference of approximately 8.5 percentage points under the benchmark assumption, or 10.4 percentage points under the conservative assumption. Comparisons involving the subsidized-AI condition require larger detectable differences because of its smaller sample size. The design also allows us to estimate interactions between payment schemes and AI access/pricing conditions. These interaction estimates rely on treatment-cell sample sizes rather than marginal treatment-group sizes. For labor-productivity outcomes, difference-in-differences comparisons involving the paid-AI and free-AI cells, with approximately 120 participants per cell, can detect interaction effects of approximately 0.42 standard deviations under the benchmark assumption, or 0.51 standard deviations under the conservative assumption. Interaction comparisons involving the subsidized-AI cells require larger effects, typically around 0.51 standard deviations or more under the benchmark assumption. For binary AI adoption outcomes, analogous interaction comparisons between paid-AI and free-AI cells can detect interaction effects of approximately 20.9 percentage points under the benchmark assumption, or 25.6 percentage points under the conservative assumption. Interaction estimates involving the subsidized-AI cells require larger detectable effects and will be interpreted with appropriate caution.