AI & Delegation and Source Labelling.

Participants are randomly assigned to receive financial advice attributed
to one of three sources: (1) Bank Advisor, (2) AI Advisor, or (3) an
unlabeled baseline condition. In the Feedback Arm, participants first
receive performance feedback (positive or negative) on earlier credit card
choices before making a delegation decision. In the Control Arm,
participants proceed directly to the delegation decision without feedback.

2026-02-16

2026-02-17

1. S5_AcceptOverride: Binary decision to accept or override the
recommendation in Scenario 5 (Accept=1, Override=0)

2. S5_Reliance: Continuous measure (0-100) of how much participants want
the recommendation to determine their final choice

3. SwitchStick: Binary decision to change or keep original choice after
receiving feedback (Feedback Arm only)

S5_AcceptOverride is coded from participant responses: Accept=1 if they
choose to follow the recommendation, Override=0 if they choose to use
their own selection. For participants whose initial choice matches the
recommendation, we additionally code S5_MatchAttribution to distinguish
"Accept because of recommendation" (Influenced) vs "Accept independently"
(Independent).

S5_Reliance is captured directly from a slider question (0=Use my choice
only, 100=Use recommendation only).

SwitchStick is coded: Change=1, Keep=0, capturing behavioral response to
feedback before the delegation task.

1. Trust_Ability: Perceived competence of the advisor to choose the
lowest-cost option (5-point Likert scale)

2. Trust_Benevolence: Perceived alignment of advisor with participant's
financial interests (5-point Likert scale)

3. Trust_Integrity: Perceived consistency and rule-following of advisor
(5-point Likert scale)

4. Agency_Conflict: Perceived financial motive/conflict of interest
behind recommendation (7-point scale)

5. Perceived_Algorithmicity: Belief about whether recommendation was
generated by computer/algorithm vs human judgment (5-point scale)

6. MC_Source_Recall: Whether participant correctly identifies the source
of the recommendation (categorical: Bank/AI/Neutral/Other)

7. MC_Pass: Binary indicator of correct source recall (1=correct, 0=incorrect)

8. S5_MatchAttribution: For participants whose initial choice matched the
recommendation, whether acceptance was influenced by the recommendation
or independent (Influenced/Independent/N/A)

9. Risk_SOEP: General risk tolerance (0-10 scale, SOEP validated item)

10. Task_Objectivity: Perceived objectivity of credit card optimization
task (7-point scale)

11. Delegation_Propensity: General preference for making own choices vs
letting experts decide (7-point scale)

12. AI_Trust_General: Baseline trust in AI systems for decisions/
recommendations (0-10 scale)

13. Bank_Trust_General: Baseline trust in banks to act in customers'
financial interests (0-10 scale)

14. Ambiguity_Aversion: Preference for known vs unknown probabilities
(Ellsberg urn task - categorical)

15. FinLit_Total: Financial literacy score (0-3, sum of correct responses
to Lusardi & Mitchell "Big Three" items)

16. Correct_S1 through Correct_S4: Accuracy on each of the four initial
scenarios (binary: 1=correct, 0=incorrect)

17. Total_Correct: Sum of correct responses across Scenarios 1-4 (0-4)

18. Valence_Determinism: Classification of participant's overall accuracy
pattern (AllCorrect/AllWrong/Mixed)

TRUST DIMENSIONS (Mayer et al., 1995): Three items measuring Ability
("capable of choosing lowest-cost option"), Benevolence ("acting in my
interests"), and Integrity ("applies consistent rule") on 5-point Likert
scales. Neutral condition rewords items to reference "this recommendation."

AGENCY CONFLICT (Jensen & Meckling, 1976): Perceived financial motive
behind recommendation measured on 7-point scale ("Definitely would not
benefit" to "Definitely would benefit").

PERCEIVED ALGORITHMICITY (Logg et al., 2019): 5-point scale measuring
belief that recommendation was computer-generated vs human judgment.
Critical for assessing implicit source inference in Neutral condition.

MANIPULATION CHECK: MC_Source_Recall captures verbatim source recall;
MC_Pass coded 1 if response matches assigned condition, 0 otherwise.

MATCH ATTRIBUTION: For participants whose choice matched the recommendation,
distinguishes "Influenced" (recommendation reinforced choice) vs
"Independent" (would have chosen same regardless).

RISK TOLERANCE (Dohmen et al., 2011): SOEP validated single-item measure
of general risk-taking propensity (0-10 scale).

TASK OBJECTIVITY (Castelo et al., 2019): 7-point scale measuring whether
task is perceived as objective/calculable vs subjective.

DELEGATION PROPENSITY: 7-point scale measuring preference for self-decision
vs expert delegation in complex decisions.

BASELINE TRUST: AI_Trust_General and Bank_Trust_General measured as
trait-level covariates (0-10 scales).

AMBIGUITY AVERSION (Ellsberg, 1961): Binary urn choice distinguishing
ambiguity aversion from risk aversion.

FINANCIAL LITERACY (Lusardi & Mitchell, 2014): Sum of correct responses
to "Big Three" items (0-3 scale).

SCENARIO ACCURACY: Correct_S1-S4 (binary), Total_Correct (0-4 sum), and
Valence_Determinism (AllCorrect/AllWrong/Mixed) classify participant
performance and feedback assignment mechanism.

Between-subjects randomized controlled trial with two arms:

Feedback Arm (50%): 2×2 factorial design crossing Source (Bank Advisor vs
AI Advisor) with Feedback Valence (Positive vs Negative). Valence is
determined by participant's accuracy on randomly-selected feedback
scenario.

Control Arm (50%): Three-way randomization of Source (Bank Advisor vs AI
Advisor vs Neutral) without feedback exposure.

All participants complete identical credit card optimization scenarios and
a final delegation task (Scenario 5).

Not available

Computer-generated randomization via Qualtrics survey software.
Randomization uses Qualtrics "Evenly Present Elements" function to ensure
balanced allocation across conditions. Nested randomization: participants
first randomized to Feedback vs Control arm (50/50), then randomized to
Source conditions within arm (50/50 for Feedback; 33/33/33 for Control).

Individual participant. Each participant is independently randomized to
one experimental condition. No clustering or group-level randomization.

No

N/A - Individual-level randomization, no clustering.

600 participants (individual respondents recruited via Pureprofile online panel, Australian adults)

FEEDBACK ARM (n=300 total):
- Bank Advisor: n=150 (Positive: ~75, Negative: ~75)
- AI Advisor: n=150 (Positive: ~75, Negative: ~75)

CONTROL ARM (n=300 total):
- Bank Advisor: n=100
- AI Advisor: n=100
- Neutral: n=100

Note: Within Feedback Arm, Positive/Negative split depends on participant
accuracy and is expected to be approximately balanced but not guaranteed.

For the primary contrast (Bank vs AI on S5_AcceptOverride): With n=150 per group (Feedback Arm Bank vs AI), alpha=0.05, power=0.80: - Minimum detectable effect size: Cohen's d = 0.32 (small-to-medium) - For binary Accept/Override outcome: detectable difference of ~12 percentage points (e.g., 50% vs 62% acceptance rates) For Source main effect across Control Arm (3 groups, n=100 each): - ANOVA detectable effect size: f = 0.18 (small-to-medium) For continuous S5_Reliance (0-100 scale, assumed SD=25): - Detectable difference: ~8 points on 100-point scale Power calculations conducted using G*Power 3.1.