AI-Powered Promises: The Influence of ChatGPT on Trust and Trustworthiness

2023-06-14

2023-10-14

- "In", and "Roll" rates in each treatment arm (Communication, and GPT).
- The frequency that messages contain promises (this will be rated independently by external coders).

- Joint outcome ("IN", "ROLL")
- first-order beliefs about trust (fraction of subjects playing “IN”), trustworthiness (fraction of subjects playing “ROLL”), and whether the message was generated by GPT.
- Whether participants choose to rewrite the message generated by GPT
- participants’ social value orientation
- participants’ attitudes toward artificial intelligence
- correlations between behaviors across the two periods
- risk aversion
- demographic characteristics

We will run a trust game with hidden action.

Two players will play as Trustee or Trustor.

The treatments differ in whether an artificial-intelligence-based language model can be used in order to communicate.

We carry out a hidden action trust game over two treatments. Participants will be randomly matched in pairs of two (player A and player B) at the start of the session. Each player will be told the experiment consists of a comprehension quiz, two decision tasks, a survey, and then finally payment feedback. They will be told that one of the two decision tasks will be randomly chosen as a payment.

After instructions and the quiz, the first decision task starts with a phase where player B can send a written message to player A. In the “GPT” treatment player B has access to GPT-4 in generating this message, which both players are aware of. We preload the instructions for the game into the application and tell the application that it will be asked for advice by Participant B. We also instruct participants on some 'GPT basics', to account for limited experience with this novel technology.

Afterward, simultaneously:

- Each player A will see the message player B wrote and will indicate whether they wish to choose “In” or “Out”. If A chooses “Out”, the game ends. If A chooses "In" player B's choice will decide the final outcome.
- Afterwards, each player A will make an incentivized guess as to the probability that player B chooses “Roll”. [if he is in the GPT treatment, he will also be asked what his confidence is that the message was written using GPT, this is not incentivized]

- Each player B will indicate whether they wish to choose “Roll” or “Don’t Roll” (a computerized 6-sided die). Note that at this point, B will not know whether A has chosen “In” or “Out”; However since B’s decision will only make a difference when A has chosen “In”, we ask B’s (for the purpose of making this decision) to suppose that A has chosen “In”.
- Afterwards, each player B will make an incentivized guess as to the probability that player A has chosen “In”.

Decision Task 2 will be a repeat of Decision Task One where roles are switched. Players will not know this in advance, although participants know that a second decision task was coming. Each player will be grouped with another player (and all players will switch roles). This switching will be kept within clusters of four, or six at maximum if the participant number in a session is not divisible by four.

- Within each session, there is a randomized pairing before decision task 1 occurs. This is done by computer, namely via the method 'group_randomly' in Otree. This shuffles players such that they can end up paired with any random player.

- After decision task one is done, all players will deterministically shift 'roles' in the group before decision task 2, from player A to player B and vice versa. They are then randomly repaired within their cluster (which is decided randomly from the start due to the random allocation into groups, and therefore clusters)

Individual-level randomization into groups of two and clusters of four (or six at maximum).

Yes

For the initial round, 160 participants in groups of two.
For the analysis of the second round, ideally, 70-80 clusters of four (two groups). This depends on how many instances of clusters of 6 we will have.

320 student participants

160 participants in the GPT communication treatment,
160 participants in the Communication treatment