The Impact of Explainability Cues on Human Trust Calibration and Fact-Checking Effort in Large Language Model Interfaces

01 / Research problem

As large language models move into high-stakes workflows, accurate human trust calibration becomes critical: users should rely on the model in proportion to its actual reliability. Interfaces increasingly add explainability cues such as cited sources, hedging language, disclaimers, and confidence scores. Yet there is little rigorous evidence on whether these cues foster healthy, critical evaluation or instead induce blind trust. Most explainable AI research targets model developers, not the people using the interface.

02 / Research question

Do interface-level explainability cues increase calibrated reliance, or do they act as cognitive shortcuts that suppress independent verification? The study tests four hypotheses:

1. Citation Paradox: cited sources raise trust but reduce fact-checking effort.
2. Hedging as a calibration mechanism: hedging lowers over-trust without adding cognitive friction.
3. Disclaimer ineffectiveness: generic disclaimers do not change user behaviour.
4. Confidence-induced automation bias: high confidence displays raise trust and reduce verification.

03 / Methodology

A quantitative, non-experimental study using a cross-sectional dataset of N = 1,000 human-AI interaction sessions across GPT-4, Claude, Gemini, and Llama. The independent variables were cited sources (present or absent), hedging words, a generic disclaimer, and a displayed AI confidence percentage (0 to 100). The outcomes were a subjective trust score (0 to 10), verification duration in minutes, and whether a fact-check was performed. Digital literacy (1 to 4) and domain criticality were modelled as moderators. Analysis used ordinary least squares regression for trust and verification duration, and binary logistic regression for fact-check likelihood, at alpha = 0.05.

04 / Findings and contribution

The trust model explained 57.0% of the variance (R² = 0.570).

• Citation Paradox. Cited sources raised subjective trust (β = 0.808, p < 0.001) but reduced the odds of an independent fact-check (odds ratio = 0.710, p = 0.011) and cut verification time by 2.37 minutes (β = -2.366, p = 0.014).
• Hedging. Hedging language reduced baseline trust (β = -0.501, p < 0.001) across all digital literacy levels (interaction p = 0.955), acting as a universal anti-overtrust mechanism without added friction.
• Disclaimers. The generic disclaimer had no significant effect on trust (β = -0.086, p = 0.242) or on verification (p = 0.711).
• Confidence displays. Confidence scores raised trust (β = 0.086 per percentage point, p < 0.001) but suppressed verification duration by 0.109 minutes per point (p = 0.001) and lowered the odds of fact-checking (p < 0.001).

The contribution is a dual-evaluation framework that maps subjective trust against measured verification behaviour, plus evidence-based UI/UX guidance: progressive citation disclosure, adaptive confidence displays, and epistemic hedging in low-confidence or high-risk responses.

05 / Reflection and future research

The findings show that transparency features can backfire when they are treated as decoration rather than behaviour. The next step is to test the proposed friction-based citation and confidence patterns experimentally, and to connect this line of work to the doctoral direction on how interface decisions shape attention, trust, and verification.