Experimental Design
We study the impact of digital photo enhancement on individual behaviors in a trust game. The trust game, proposed by Berg et al. (1995), is a widely applied method to measure trust and trustworthiness in economic decisions. In our experiment, each sender is randomly paired with ten different receivers to play ten rounds of a one-shot game. At the beginning of each round, both the sender and the receiver are endowed with ten tokens. Each token is worth 1 RMB (0.15 USD). In each round, the sender can invest any number of tokens, from 0 to 10, to the receiver. Any amount sent will be tripled, and the receiver can return any number of tokens out of the tripled amount back to the sender.
We enrich the trust game with an AI photo enhancement module and design a single-factor randomized online experiment. The treatment differs in whether the participants enhance their profile photos and present themselves with the enhanced photos in the trust game. In the experiment, a participant creates a virtual account with an ID code. The participant needs to upload a profile photo that is saliently displayed to both players in the game. Participants are provided with an AI photo makeup module to enhance their uploaded photo. In the treatment group, participants can use the enhanced photos to present themselves in the trust game. In the control group, participants can only use their original profile photos in the game. The photo enhancement module uses the Pho.to API for facial recognition and feature enhancements. The participants can adjust the brightness, the temperature, and the saturation to obtain better skin tones and create refined facial features. They can also smooth out wrinkles, improve eye colors, remove facial imperfections and apply a warm-glow filter. The default options of the module are set to make the profile photos significantly more appealing. Meanwhile, detailed instructions are given to the participants on how the module works.
The experimental procedure is as follows. Each participant needs to register an account with their names and phone numbers after arriving at the experiment webpage. In the recruitment, we make sure that no participants have previously engaged in similar experiments. After registration, the participant is directed to the detailed instructions on the rules of the trust game and takes a quiz to ensure complete understanding. The participant then creates a profile with a personal photo. We require that the profile photo be a clear headshot with a white background and no modifications. Participants with unqualified profile photos are excluded. After completing the profile, the participant is randomly assigned to one of the game roles (a sender or a receiver), and enters three rounds of trial games to get familiar with the game rules and interface. After the trial games, the participant is randomly assigned to the treatment and control groups. Participants in the treatment group use their enhanced photo but those in the control group can only use their original photos in their profile in the game.
Next, each participant plays ten rounds of the independent one-shot trust game with the assigned role (sender or receiver). Senders (participants assigned the sender role) decide the amount to be sent to the opponent, while receivers (participants assigned the receiver role) determine the amount to be returned to the opponent. Both the senders and the receivers make decisions by entering numbers in an online form. The players in each round are randomly assigned to play with a fictional opponent. Participants were unaware that their opponents were randomly drawn from a historical dataset (Eckel and Petrie 2011). Opponents’ profile photos are randomly set to be either original or enhanced. Finally, each participant confirms the total earnings and collects rewards after finishing a posttest. The post-experiment includes measurement scales for social presence (Khalifa and Shen 2004), trust propensity (Gefen and Straub 2004), risk attitude (Pennings and Smidts 2000), as well as trap questions, manipulation checks, and demographic questions. We use a seven-point Likert scale and randomize the items to avoid potential order effects.