Experimental Design Details
In our experiment, we simulate a scenario where participants assume the role of junior workers in a consulting firm, tasked with providing solutions to address the targeted company's challenges.
In the experiment, following the social class manipulation, each participant is directed to the introduction pages on help-seeking from the supervisor and AI. Then, a total of eight questions is shown sequentially, one from each difficulty group, ensuring that the average difficulty level of each question set is comparable across participants. Each question will be displayed for 30 seconds, during which participants can decide whether to answer independently, seek help from AI, or seek help from the supervisor.
For questions that participants choose to seek help from the supervisor, they need to communicate with the supervisor, who is framed as an expert in the business management consulting field. To do this, participants must go to the help-seeking room/send a text message to explain the question's context to the supervisor to seek help (the participants are required to go to the help-seeking room to talk to the supervisor in the main study while send a text message to the supervisor in the supplemental study), then proceed to the answer submission page. For questions that participants choose to seek help from AI, they must come up with a prompt themselves and enter it on the webpage.
The participants are informed that if the supervisor decides to give advice after communicating with them, the supervisor will give the correct answer to the question with an 80% probability while giving no advice/answer with a 20% probability; if the prompt to AI is effectively enough, AI will also generate the correct answer with an 80% probability while generating no advice/answer with a 20% probability.
In the experiment, we let participants first make help-seeking choices for all 8 questions, then show the response from the supervisor or AI per their choice. This approach prevents biases that could arise from learning and belief updates about the perceived capabilities of the supervisor and AI as participants progress through the question list.
After the help-seeking stage, participants will be directed to the answer submission page, where they need to submit each answer within 15 seconds. Finally, they finish a short follow-up questionnaire to conclude the experiment.