Human-AI interaction in the workplace: Understanding workers' Internal Motivation

Participants are randomly assigned to complete a creative task, designing a book cover based on a summary, in collaboration with either a human co-worker (another participant) or a generative AI (ChatGPT). The task is conducted in a collaborative setting, allowing us to examine how the type of co-worker (human vs. AI) influences participants' intrinsic motivation.

2025-12-03

2029-01-01

Primary Outcome 1: Self-reported intrinsic motivation
Intrinsic motivation will be measured using validated survey scales derived from Self-Determination Theory (Deci & Ryan). These include Perceived Competence, Perceived Autonomy, Perceived Relatedness, Perceived Enjoyment / Interest (primary affective component of intrinsic motivation).

Primary Outcome 2: Behavioral motivation proxy
we will measure participant self motivation by asking participants whether they want to repeat the same task or whether they want to complete an alternative task. By varying the length of the alternative task, we are able to measure the strength of their internal motivation. This behavioral measure complements self-reported intrinsic motivation.

Specifically we test the following hypotheses:
Hypothesis 1a. Algo–human teams will decrease perceived relatedness among team members.
Hypothesis 1b. Algo–human teams will increase perceived team functioning, as measured by task orientation.
Hypothesis 2a. Higher perceived relatedness among team members will increase perceived enjoyment of the team task.
Hypothesis 2b. Higher perceived team functioning will increase perceived enjoyment of the team task.
Hypothesis 3. Higher enjoyment of the team task will increase task interest and the willingness to complete the task again.

Secondary outcomes correspond to the following hypotheses (AH1–AH3). These outcomes focus on performance, distributional patterns, and moderators:
AH1. Motivational outcomes—measured through task interest and willingness to perform the task—will predict performance on the task. Each individuals' performance will be assessed by several external raters recruited from the same population.
AH2. Algo–human teams in the top decile of motivational outcomes or performance will be underrepresented relative to fully human teams.
AH3. Differences in motivational outcomes and performance between treatments will be moderated by team composition, including personality traits, familiarity with large language models, and teamwork experience.
AH4: Task performance: Task performance will be measured as the difference in participants’ performance across the two treatments, as evaluated by multiple external raters recruited from the same population.

We use a between-subjects, online experiment with a 2×1 design. Participants are randomly assigned to complete a creative task, designing a book cover based on a book summary, in collaboration with either a human co-worker (another participant) or a generative AI co-worker (ChatGPT).

The task is framed as a collaborative activity; all participants receive identical instructions and materials. Random assignment is handled automatically by the experimental platform.

After completing the task, participants respond to a survey measuring intrinsic motivation, using validated scales based on self-determination theory (perceived competence, autonomy, and relatedness). In addition, participants answer three open-ended questions about their engagement during the task, their interaction with their partner (AI or human), and the steps they followed to coordinate the task.

As a behavioral proxy of motivation, participants are also given the option to continue working on either:

1) a similar creative task, or

2) a different repetitive task requiring little creativity.

This choice allows us to assess both their willingness to continue.

Not available

A computer automatically randomizes participants through the experimental platform. At the time they begin the task, participants are randomly assigned to one of the two treatment conditions (human co-worker vs. AI co-worker) using a built-in randomization function.

The unit of randomization is the individual participant. Each participant is independently assigned to one of the treatment conditions (human co-worker vs. AI co-worker).

No

Human–AI condition: Treatment is clustered at the individual level (1 participant per cluster). Planned number of clusters = N individuals.

Human–Human condition: Treatment is clustered at the dyad level (2 participants per cluster). Planned number of clusters = N/2 dyads (each dyad = 1 cluster).

We collect data from Prolific and stop the experiment when 900 participants have completed the experiment.

600 participants – Human co-worker treatment

300 participants – Generative AI (ChatGPT) co-worker treatment

Randomization is done at the individual level. Upon starting the experiment, each participant is randomly assigned to one of the two arms.

Our main outcome (willingness to repeat the task) is measured on a 0–10 scale. Based on pilot data, the standard deviation of this outcome is 3.68. Using a significance level of α = 0.05 and accounting for the clustered sample design in the Human–Human condition, the minimum detectable effect size (MDE) is: 0.97 when the observations within the cluster are assumed to be fully dependent; 0.86 when the observations within the cluster are assumed to be fully independent 0.87 when we assume an intra cluster correlation of 0.08 as measured in our pilot.