Quantitative effects of explainability modalities on user perceptions

Determine how different forms of explainability in AI systems quantitatively shape user perceptions.

Background

The paper reviews prior Explainable AI (XAI) literature and notes that much of it has been qualitative or theoretical, highlighting a lack of quantitative evidence connecting specific explanation types to measurable user outcomes.

Although the study presents a preliminary quantitative experiment comparing several explanation conditions in a loan approval context, the authors explicitly identify a broader, field-level uncertainty about the quantitative impact of different explainability modalities on user perceptions.

References

Most prior studies are qualitative or theoretical, leaving open questions about how different forms of explainability quantitatively shape user perceptions.

— Preliminary Quantitative Study on Explainability and Trust in AI Systems  (2510.15769 - Sunny, 17 Oct 2025) in Section: Related Work (summary paragraph)

Recognition alone leaves open how users evaluated the explanation.

— How Far Do Capability Cues Travel? Anthropomorphism and Differentiated Trust in a Platform-Embedded AI Assistant  (2609.09713 - Mao et al., 9 Sep 2026) in Section 6.2, “When a rationale serves as evidence of intelligence” (page 16)

Their clinical usefulness has not been assessed with practitioners. Whether the justifications improve a stockperson's decisions or their confidence in the system is an empirical question that requires a user study we have not carried out. We regard the lack of a user study as the most important gap between the present work and a validated claim of trustworthiness.

— When Do Language-Grounded Explanations Help? A Graph-Bottleneck for Farm Monitoring Interpretable Sheep Facial Pain  (2609.20427 - Noor et al., 17 Sep 2026) in Section 7, Limitations

Which the over-trust branch should prefer is a question the present study cannot answer: establishing it requires distinguishing the effect of an explanation from that of a correction in the interaction logs, and that this is the obstacle rather than an insufficient number of events is itself part of what the instantiation shows.

— CRiDiT: Instantiating a run-time testbed for trust calibration in AI-infused systems  (2609.24833 - Ding et al., 21 Sep 2026) in Section 6.1, subsection “Implications for Closed-Loop Calibration Design”

Whether a separate device also changes what people notice is a question for a user study, which we have not run.

— Approve What, Exactly? Effect Binding and Trusted Rendering for Agent Step-Up Authorization  (2610.21063 - Kodathala et al., 15 Oct 2026) in Section 7.3, RQ1(3): the second device adds nothing measurable here