Trust Development & Repair in AI-Assisted Decisions – Comprehensive Study Notes

Background & Key Concepts

Human–AI Collaboration
- Seeks to combine complementary strengths of humans (contextual reasoning, empathy, domain insight) and AI (speed, pattern-recognition, memorisation).
- Central challenge: users must decide when and how much to trust AI advice.
Complementary vs. Overlapping Expertise
- Overlapping: both human and AI competent → humans can directly judge AI accuracy.
- Complementary: each excels at different facets → humans cannot always assess AI accuracy.
Appropriate Trust / Trust Calibration
- Balance between over-trust (unwarranted reliance) and under-trust (undue scepticism).
- Calibration aims to align subjective trust with objective AI capability.
Trust Repair Strategies (TRSs) – interventions adopted from Social Psychology & Human–Robot Interaction (HRI) to rebuild trust after violations.
- Apology (expression of regret)
- Denial (reject culpability)
- Promise (commit to do better)
- Model Update (inform user of technical improvement)
Dispositional Trust (TiA-PtT) – a trait-level propensity to trust automation; moderates initial trust.

Research Questions & Contributions

RQ1: How does perceived AI accuracy in High Human-Expertise (HHE) tasks influence users’ trust during Low Human-Expertise (LHE) tasks?
RQ2: In complementary expertise settings, how does trust recover when AI accuracy improves, with and without explicit TRSs?
Key contributions
- Empirical evidence that users extrapolate AI accuracy in HHE tasks to calibrate trust in LHE tasks.
- Comparative efficacy of four TRSs; Model Update > Apology > Promise ≈ No-Repair > Denial.
- Insight that trust recovery depends on anthropomorphism, perceived regret/deceit, causal attribution, and nature of improvement (behavioural vs technical).
- Dual-task (Shapes & Animals) replication enhances ecological validity.

Related Work

Trust as an attitude under uncertainty (Lee & See, 2004).
Accuracy → trust correlation (Yu et al., 2016; Yin et al., 2019).
Initial impressions / early errors have long-term effects (Tolmeijer et al., 2021).
HRI literature on trust repair: mixed findings for apology, promise, denial; scarcity of work in non-robotic AI decision aids.

Trust Repair Strategies (TRS) – Details

Apology
- Operates emotionally; seeks to restore social expectations.
- Text used: “I’m sorry … I hope you can trust me again.”
Denial
- Shifts blame; may appear deceptive if evidence contradicts statement.
Promise
- Behavioural commitment; effectiveness relies on perceived agency to fulfil promise.
Model Update (novel)
- Communicates algorithmic upgrade; implicitly supplies causal attribution (flawed model → fixed).
All TRS messages followed template: acknowledgement of mistrust → core element → hope for restored trust.

Methodology

Design: Survey-based, between-subjects; $5$ TRS conditions (including baseline No-Repair) × $2$ tasks.
Participants: $N = 300$ Prolific users (eng ≥ $98\%$ ); $150$ per task, $30$ per TRS.
- Mean ages: Shapes $\bar{x}=35$ (SD $=13.01$ ); Animals $\bar{x}=34.2$ (SD $=11.85$ ).
Power Analysis: $f^{2}=0.25$ , $\alpha = 0.05$ , power $=0.8$ ⇒ minimum $n=135$ per task (G*Power 3).
Phases & AI Accuracy
1. Phase 1 – High accuracy $=80\%$ (1 error at trial 7) → build trust.
2. Phase 2 – Low accuracy $=20\%$ (multiple errors) → violate trust.
3. Phase 3 – High accuracy $=80\%$ again; TRS shown before start.
Trial Types (10 per phase; 30 total)
- HHE (Familiar) = human knows answer.
- LHE (Unfamiliar) = human does not know; must rely on AI.
Delay: $3\;\text{s}$ before AI advice to encourage independent thought.
Model-Update scenario: extra $6\;\text{s}$ wait to simulate retraining.

Experimental Tasks & Stimuli

Shapes Task
- Familiar: Circle, Rectangle, Triangle (5 visual variants each).
- Unfamiliar (fabricated): “Scleratice”, “Tenectus”, “Pyrangle”; randomised fill/border patterns.
Animals Task
- Familiar: 15 common species (Cat, Dog, Horse, etc.).
- Unfamiliar: 15 obscure species (e.g., Kakapo, Markhor, Aye-Aye) with non-descriptive names.
Stimuli sets balanced & counter-balanced; examples in Appendix.

Measures & Instruments

Behavioural Trust: Binary agreement (“AI is accurate / inaccurate”) on each trial.
Confidence: Slider $1\text{–}100$ (anchor appears after click to avoid bias).
Self-Report Trust: $12$ -item scale (Jian et al., 2000) $1\text{–}7$ after each Phase.
Dispositional Trust: $\text{TiA–PtT}$ sub-scale prior to task.
Open-Ended Questions: Post-study reflections on trust evolution & TRS perception.

Results

Manipulation Check

HHE accuracy near ceiling → Shapes $99.73\%$ , Animals $90.35\%$ .
LHE accuracy ≈ chance → Shapes $49.77\%$ , Animals $49.17\%$ .
Confidence HHE > LHE (e.g., Shapes familiar $\bar{x}_{conf}=99.01$ vs unfamiliar $43.01$ ).

Influence of Perceived HHE Accuracy on LHE Trust (RQ1)

GLMM: agreement in LHE trial predicted by AI correctness in preceding HHE trial.
- Shapes: $\beta = -0.472$ (SE $=0.034$ ), p<0.001; $\text{OR}=1.62$ .
- Animals: $\beta = -0.576$ (SE $=0.064$ ), p<0.001; $\text{OR}=1.56$ .
Interpretation: correct HHE judgment increases odds of trusting AI on subsequent LHE trial by ~ $60\%$ .

Trust Dynamics Across Phases

Phase-level trust means
- Shapes: $M<em>1=3.94$ → $M</em>2=3.20$ (drop, t(149)=11.0, p<0.001).
- Animals: $M<em>1=4.42$ → $M</em>2=3.72$ (drop, t(149)=8.44, p<0.001).
Linear models
- Phase 1 trust ↑ with $\text{TiA–PtT}$ (Shapes $\beta=0.645$ ; Animals $\beta=0.681$ ).
- Phase 2 trust predicted by Phase 1 trust, not by $\text{TiA–PtT}$ .
- Phase 3 trust predicted by both Phase 1 & Phase 2 trust; $\text{TiA–PtT}$ significant only in Animals.

Effectiveness of TRSs (RQ2)

Regression coefficients (vs baseline)
- Model Update: Shapes $\beta=-0.711$ , Animals $\beta=-0.715$ (highest recovery).
- Apology: significant but smaller effect.
- Promise ≈ No-Repair; Denial worsened trust.
Post-recovery mean trust exceeded Phase 1 only for Model Update; Apology reached ~baseline; Promise/No-Repair below; Denial lowest.

Qualitative Findings

Heuristic Use: Majority reported using AI performance on Familiar items to gauge trust on Unfamiliar.
Apology: Effective when users anthropomorphised AI (“seems regretful”); less effective when apology deemed inauthentic.
Denial: Viewed as deceitful; absence of causal explanation intensified distrust.
Promise: Helped when users believed AI had agency to improve; perceived as hollow or uncertain by others.
Model Update: Widely accepted; technical fix seen as credible; also reduced suspicion of intentional deception.

Discussion & Implications

Accuracy-Based Heuristic
- Designers can leverage AI performance in domains where users have expertise to bootstrap trust in unfamiliar domains.
- Caution: if accuracy diverges across domains, heuristic may induce mis-calibration.
Persistent First Impressions
- Early trust levels continue to colour later phases despite new evidence.
Dispositional Trust
- Matters initially; its influence wanes when direct performance evidence accumulates.
TRS Design Guidance
- Technical explanations/updates can surpass emotional appeals.
- Provide causal attribution for errors to avoid perceptions of deceit.
- Avoid denials that contradict observable mistakes.
Ethical/Practical Considerations
- Transparent communication of model changes respects user autonomy.
- Over-anthropomorphising may mislead; balance needed.

Limitations & Future Work

Single violation per phase; future work to vary frequency/severity and temporal placement.
Binary expertise extremes (certain vs uncertain); explore graded uncertainty.
Simulated AI; replication with real adaptive models recommended.
Investigate long-term stability of repaired trust and successive violations.

Key Takeaways

Users dynamically calibrate trust using observed AI accuracy where they hold expertise.
Trust erosion via errors is swift; recovery is partial unless supported by effective TRS.
Model Update (technical fix) most effective; Denial counter-productive.
Trust repair hinges on users’ causal reasoning, perceived sincerity, and belief in AI’s capacity to improve.