CoChat

The Illusion of Explanatory Depth in Generative AI: Miscalibration, User Profiles, and the Persistence of Retrieval-Based Mental Models

Open access
Alon, Lilach · 2026 · SSRN Electronic Journal

The DOI matches its Crossref record; 44 of the 44 references it cites were found in a registry or an index.

Abstract

Mental models of Generative AI (GenAI) shape how users interpret outputs and decide when to verify information, yet users often infer how these systems work from surface behavior rather than mechanistic understanding. Drawing on the illusion of explanatory depth (IoED), this mixed-methods study examined whether perceived understanding changes after an explanation attempt and whether self-evaluations align with demonstrated explanatory understanding. A large adult sample (N = 2,926) rated their understanding of how GenAI produces answers, wrote an open-ended explanation, and then rated their understanding again. Explanations were rubric-scored for explanatory quality and coded for recurring themes. Results showed substantial miscalibration. Self evaluations changed minimally after the explanation attempt (pre M = 3.90; post M = 3.95). In contrast, demonstrated understanding was low (M = 2.20), revealing a large gap between perceived and explanation-based understanding. Explanations were dominated by retrieval-based misconceptions: 54.03% described GenAI as a web search engine and 9.60% as an internal database, whereas mechanistic generative explanations (e.g., token prediction) were rare (2.22%). A two-step cluster analysis identified three user profiles: over-estimators (51.9%) reported high confidence but provided weak, retrieval-framed explanations; calibrated explainers (13.5%) paired high confidence with strong mechanistic accounts; and underconfident low-knowledge participants (34.6%) reported low confidence and also provided weak explanations. These findings suggest that confidence-explanation gaps can persist even after an explanation attempt and that fluent interaction may reinforce familiar search metaphors. The study underscores the need for AI literacy that explicitly contrasts retrieval and generation to support calibrated reliance in human-AI interaction.

Overview

Integrity signals

Registry

Crossref · 10.2139/ssrn.6172315

Cites

44 of 44 references verified0 citation errors · 0 not checkable

Statistics

Recomputing p-values needs the paper’s full text.

Verification

Claims the paper makes

We couldn’t retrieve this paper’s full text, so its claims and self-check weren’t run. The registry and citation checks still apply.

Citations

Cites

Show all 44 references

Research works printed but not linked: 3

  1. Y Elsayed, 2024, ChatGPT and the illusion of explanatory depth, Proceedings of the Annual Meeting of the Cognitive Science Society
  2. N Vaupoti?, 2022, Gaining insight through explaining? How generating explanations affects individuals' perceptions of their own and of experts' knowledge, International Journal of Science Education, 12, 42
  3. J Wilson, 2023, Perceptions of explanation completeness help decrease knowledge overestimation, Proceedings of the Annual Meeting of the Cognitive Science Society

Metadata

Authors
Alon, L.
Affiliations
Tel Hai Academic College (IL)
Corresponding author
Alon, Lilach
Venue
SSRN Electronic Journal
Publisher
Elsevier BV
Published
2026
ISSN
1556-5068 (ISSN-L)
OpenAlex
W7127417011
Cited by
0
References
30
Type
posted content
Language
en
Open access
Locations
SSRN Electronic Journal (acceptedVersion)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
  • Social Sciences › Psychology › Developmental and Educational Psychology › Educational Strategies and Epistemologies
  • Physical Sciences › Computer Science › Information Systems › Information Retrieval and Search Behavior
Keywords
illusion of explanatory depth, generative AI, mental models, calibration error, mechanistic explanation, AI literacy, user profiles, human-AI interaction, explanation quality

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.