The DOI matches its Crossref record; 24 of the 24 references it cites were found in a registry or an index; 2 of the 16 sources its claims cite were checked for support.
Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth
Open accessAbstract
The Illusion of Explanatory Depth (IOED), i.e., the tendency to overestimate the coherence and depth of one’s understanding, tends to increase when people search for explanations online. However, it is less clear whether this applies to information obtained from AI chatbots. We tested whether receiving information from a chatbot magnifies IOED and how it affects the quality of participants’ explanations. University students (N = 102) were presented with four questions and predicted how well they could explain them. However, the GPT group preliminarily asked for explanations a custom version of ChatGPT and received pre-controlled answers; the no-GPT group received the same texts directly; the control group received no materials. Afterward, they wrote their explanations and rated them. Results showed the biggest difference between the initial prediction and subsequent self-evaluation in the GPT group. Moreover, coder evaluations and text analysis measures, including length, lexical diversity, and semantic similarity, revealed that GPT group produced less accurate explanations than no-GPT group.
Overview
CoChat Summary
TL;DR
Receiving explanations from a custom ChatGPT increased the Illusion of Explanatory Depth compared with receiving identical explanations as plain text. GPT-assisted participants also produced shorter, less lexically diverse, and less accurate explanations than the no-GPT group.
Key findings
- The pre-to-post self-rating drop was smallest in the GPT group: relative to GPT, the reduction was larger in no-GPT (b = 0.54, SE = 0.25, t = 2.20, p = .028) and control (b = 0.89, SE = 0.25, t = 3.60, p < .001).
- Coder-rated explanation quality was lower in GPT than no-GPT: means were 2.47 (SD = 0.91) for GPT, 2.75 (SD = 0.89) for no-GPT, and 1.8 (SD = 0.71) for control; no-GPT ratings were significantly higher than GPT (b = 0.83, SE = 0.41, z = 2.01, p = .044), and control ratings were significantly lower (b = -1.86, SE = 0.43, z = -4.31, p < .001).
- Text measures favored no-GPT over GPT: GPT responses averaged 43.82 words, no-GPT +15.82 words (~59.64), control -10.27 words (~33.55, p = .081); CTTR was 2.71 in GPT, +0.29 (~3.0) in no-GPT, -0.28 (~2.43) in control; semantic similarity was 0.22 in GPT, +0.03 (~0.25) in no-GPT, -0.06 (~0.16) in control.
- Across all text measures, the no-GPT condition consistently showed higher values than GPT, indicating longer responses, richer vocabulary, and higher semantic fidelity to the prompts.
Methods
- Online experiment with N = 102 university students (M age = 21.3, women = 36), randomly assigned to GPT, no-GPT, or control groups; all answered four general-knowledge questions from Eliseev & Marsh (2023): leap years, glass making, zippers, and moon phases.
- Participants first predicted how well they could explain each question, then GPT and no-GPT completed an induction phase: GPT used a custom ChatGPT returning identical pre-generated answers from ChatGPT-4.5, no-GPT received the same texts directly, and control received no materials; afterward all wrote explanations and rated them on a 7-point scale.
- Analyses used linear mixed-effects models in lme4 with condition, time, and their interaction; coder ratings used a cumulative link mixed model on a 4-point scale with two coders (final squared Cohen’s κ = .82); text analysis measured word count, Corrected Type-Token Ratio, and TF-IDF cosine semantic similarity.
Limitations
- The paper states that simulation-based sample-size estimation was not entirely reliable because comparable IOED literature was lacking, so it relied on Körner et al. (2024) and increased the target to 34 participants per group.
- The paper states that further research is needed to examine the mechanism more deeply and to test individual differences such as metacognitive sensitivity and cognitive reflection that may moderate or mitigate the effect.
- Evident from the text: the sample was university students, the study was online, only four questions were used, and the GPT condition involved a custom ChatGPT with precontrolled answers, so generalizability to other populations, domains, and real-world chatbot interactions is not established.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
2 Valid · 1 Not valid · 10 Not checked · 27 Own results and interpretations
The cited sources support only part of this claim — read each source’s verdict below.
In the paper
“we relied on the calculation by Körner et al. (2024), who, after reviewing a set of previously published IOED experiments, concluded that for studying betweensubjects effects in the IOED paradigm, a total of 62 participants, or about 31 per group (for two groups), is sufficient for power 0.8.”
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Sample size rationale
What the sources say
- Partly supportedcites Körner et al. (2024) sourceExperimental — accuracy not yet measured
Study 1, Method reports a required sample of 62 for a between-subjects ANOVA at 80% power, but that calculation assumes an effect size of f = .36. It does not establish that 62 is sufficient for between-subjects effects in the IOED paradigm generally: the authors targeted at least 120 to detect smaller effects and allow for exclusions. (LLM, conf=0.88)
plainn = 62- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
- Partly supported
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“and even more so in the control condition, b = 0.89, SE = 0.25, t = 3.60, p < .001.”
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Results
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“whereas ratings in the control condition were significantly lower, b = -1.86,SE = 0.43, z = -4.31,p < .001.”
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Results
plainstates an association- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Across conditions, post-test ratings were overall lower than pre-test ratings (β = -1.32,SE = 0.20, t = -6.56,p < .001).”
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Results
plain- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: unclear — It is unclear whether the paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
A result the paper reports from its own data — it cites no other source for it, so there is none to check. The quote shows it is in the paper.
In the paper
“Crucially, the interaction between condition and time indicated that the reduction from preto post-test in the GPT condition was significantly smaller in the no-GPT condition, b = 0.54, SE = 0.25, t = 2.20, p = .028”
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Results
hedged- Asserted: passes — The authors put it forward as true.
- Checkable: passes — It can be checked against evidence.
- Load-bearing: passes — The paper’s conclusions depend on it.
- Based: passes — The paper gives a reason why it holds.
15 more claims
Create a free account to see all 40 claims and every source verdict.
- Probably not a substantial claim. 17
- It may state two separate claims as one. 7
- The quote only partly states this claim. 1
Statistics recomputed
Every reported test result, recomputed from its test statistic and degrees of freedom. A result shows as not adding up only once an independent check has confirmed it.
- Consistentz = 2.01reported p = .044 · recomputed p = .044
“z = 2.01, p = .044”
- Consistentz = -4.31reported p < .001 · recomputed p < .001
“z = -4.31,p < .001”
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Rated minor: it would not touch a result — a value no result rests on, a p-value that stays on the same side of the significance level, or a conflict that is clerical, resolved elsewhere in the paper, or outside its own findings.
In the paper
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Results
“3. Mixed effect regression analysis outcomes for all three dependent variables; number of words (Word count), lexical diversity (CTTR) and semantic similarity (Cosine distance) between response and prompts.”
§ Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth > Data analysis
“Semantic similarity was quantified with cosine similarity (Turney & Pantel, 2010), ranging from 0 (no overlap) to 1 (identical).”
Citations
Cites
The publisher hasn’t deposited this paper’s reference list.
Research works printed but not linked: 1
- Salton, G., & McGill, M. J. (1983). Introduction to modern information retrieval. McGraw-Hill.
Metadata
- Authors
- Aslanov, I., Felmer, P., & Guerra, E.
- Affiliations
- University of Chile (CL)
- Publisher
- Center for Open Science
- Published
- 2025-10-04
- OpenAlex
- W4414819064
- Cited by
- 0
- References
- 0
- Type
- posted content
- Language
- en
- Open access
- green
- doi.org
- Locations
- PsyArXiv (OSF Preprints) (submittedVersion)
- Funders
- China Institute of Atomic Energy
- Agencia Nacional de Investigación y Desarrollo (ID24I10105, AFB240004, FB0003, 2024 AFB240004, FB210005)
- Fondo de Fomento al Desarrollo Científico y Tecnológico
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Explainable Artificial Intelligence (XAI)
- Social Sciences › Decision Sciences › General Decision Sciences › Decision-Making and Behavioral Economics
- Social Sciences › Psychology › Developmental and Educational Psychology › Educational Strategies and Epistemologies
- Keywords
- illusion of explanatory depth, overconfidence, AI chatbots, GPT, AI-generated explanations, self-assessment, explanation quality, online information seeking
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.
The claim verdicts this report counts are experimental: their accuracy is not yet measured.
A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).
A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).
Cited sources are checked with a language model’s help (gpt-6-sol).
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
How to cite
Cite CoChat for the checks on this page, and the paper for its own findings.
- This page
- CoChat. (2026). Citation check: Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth. https://app.cochat.ai/papers/10.31234/osf.io/8psgf_v1
- The paper
- Aslanov, I., Felmer, P., & Guerra, E. (2025). Overconfidence without Understanding: AI Explanations Increase the Illusion of Explanatory Depth. https://doi.org/10.31234/osf.io/8psgf_v1
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.