Preference for Explainable AI Experiment

Individuals make a decision to determine how a private lender allocate a $10000-loan. Participants are randomized into a neutral treatment where payoffs are not tied to actual loan repayment outcome, or a "lender-aligned" treatment where they have with direct stakes tied to loan repayment.

2X3X5X3 design
2: neutral treatment where payoffs are not tied to actual loan repayment outcome, or a "lender-aligned" treatment where they have with direct stakes tied to loan repayment
3: race-gender of prospective borrowers revealed to be different AND no decision to choose whether to see predicted default risk or not, race-gender of prospective borrowers NOT revealed AND no decision to choose whether to see predicted default risk or not, race-gender of prospective borrowers revealed to be different AND with decision to choose whether to see predicted default risk or not
5: always see explanation of AI, never see explanation of AI, option to see explanation of AI with no salience, option to see explanation of AI with hint that financial factors might be used by AI, option to see explanation of AI with hint that financial factors and demographics might be used by AI
3: 3 versions of explanation for those who see or chose to see explanation. One vague description of how neural network AI risk prediction works, one that plus a SHAP interpretation of why the high risk borrower was deemed as such just based on financials, one that plus a SHAP interpretation of why the high risk borrower was deemed as such just based on financials AND race and gender

2025-03-18

2025-03-22

(1) Binary decision of whether the participant want to see an explanation of how the AI made the prediction of the default risk of the borrowers before making a loan allocation decision
(2) Binary decision of whether the participant want to see the predicted default risk of the borrowers before making a loan allocation decision

Both study information-avoidance decisions.

(3) Whether the participant want to allocate 90% of the loan to the low-risk borrower and 10% to the high-risk borrower, or 50% to each
(4) How much, to reduce the second participant’s bonus at a personal cost of $0.01 per $1.00 reduction, up to the full $10 (3 scenarios -informed, not informed, chose not to be informed about explanation) - within subject

This gets at whether the information changes the actual loan allocation decision (i.e. to see if the information influences decision)

I experimentally test how individual's preference for information regarding an AI is influenced by certain factors.

(Full design attached and randomization flow attached in this pre-registration as pdf/docs before the experiment launch)

I ask participants
that if their survey response is randomly selected, what their answers and your answers
alone will do. The participants will have to answer correctly (namely, choose the answer
“Determine how the private lender will allocate the $10, 000 loan between the two real
borrowers”) to be included in the final analysis sample. This sample exclusion restriction
criteria is pre-registered.

To maintain data integrity, the experiment employs multiple bot-prevention mechanisms.
First, I include CAPTCHA authentication at multiple stages (once at the start of the
experiment and once towards the end). Second, I include a “honey-pot” question designed
to be opaque to humans but readable by automated scripts. Third, I include a “Let It Go”
test7 where I ask the participant to copy and paste lyrics under copyright protection into
a text box (LLMs like ChatGPT are “not allowed” to do so because of copyright). I also
time the CAPTCHA authentication response at later stages of the survey to detect non-
human engagement (if the response time was longer than the initial CAPCHA response by
twice). Finally, attention and comprehension checks are included to ensure participant
understanding of task requirements. The participants will have to pass both the bot
detection and the attention check to be included in the final analysis sample. This
sample exclusion restriction criteria is also pre-registered.

Randomization done using Qualtrics' in-built function

individual

No

See below

3,000

Evenly for non-lender-aligned and for lender-aligned for main arm. Same shares for a variant of the main arm where race-gender descriptors for borrowers are replaced by odd/even number birth-day/month. Main arm, in particular active choice of whether to see explanation will be 3X sampled relative to other arms in this group.
1/3 as much as above, but within this even between non-lender-aligned and lender-aligned for an arm similar to main arm except that participants also have a choice to see the predicted default risk generated by AI or not. Same numbers for a variant of the main arm where race-gender descriptors for borrowers are replaced by odd/even number birth-day/month.
3000 in total.