
When a question is missing information, do models ask, or guess? And if they guess, can the user tell? Five open-weight models from four organisations answered 142 factual questions in a fully specified form and with the key referent removed, plus three identity probes. Registered on OSF before any response was scored (10.17605/OSF.IO/CS967). Study C of the thesis MRI of a Mad Machine.
Result
| Finding | Number |
|---|---|
| Referent removed: models ask for clarification | 33.7% of prompts |
| Error among attempted answers, intact → referent removed | 39.1% → 52.0% (mixed-effects OR 4.12, 95% CI 2.76–6.14) |
| Wrong answers that carry no hedge | 75.3% |
| How well hedging separates right from wrong (0.5 = chance) | AUROC 0.56 |
| Same 50 questions, the user claims an identity | mean answer length 188 words (anonymous) vs 259–268 (patient, physician, researcher) |
When information is missing, models usually guess, and the wording barely changes. Hedging is not a usable warning sign.
Two sets of figures, one conclusion. The scoring sheets carry the original correctness labels; code/run_analysis.py reproduces the figures they support (error 39.9% intact and 66.7% referent removed; 77.4% unhedged), reported in the thesis as a robustness check. The headline figures above come from the final registered analysis, which treats a pure clarification request as an abstention, as the rubric specifies.
What is inside
| Path | Contents |
|---|---|
data/prompts/ |
Prompt set and question sources |
data/probes/ |
Pronoun, authority and safety probes |
data/raw_outputs/ |
Every model response |
data/scoring/ |
Both raters' sheets, the blind key and the 600-response overlap list |
code/ |
Sampling, collection, agreement and analysis scripts |
results/ |
Inter-rater agreement report |
STUDY_README.md |
The full study page |
Reproduce
from huggingface_hub import snapshot_download
path = snapshot_download("rejaulkorimsadi/wrong-with-confidence-llm", repo_type="dataset")
cd <the path printed above>
pip install -r requirements.txt
python code/compute_kappa.py
python code/run_analysis.py
Citation
@misc{sadi2026wrongconfidence,
title = {Wrong with Confidence: Measuring Hallucination, Overconfidence and Identity-Conditioned Behaviour in Large Language Models},
author = {Sadi, Md. Rejaul Korim and Tasin, Toufiqur Rahman and Naeem, Golam Mostofa and Chowdhury, Dewan Fahad and Chowdhury, Hatim Al Amin},
year = {2026},
note = {Pre-registered at OSF, doi:10.17605/OSF.IO/CS967}
}
Part of a research collection
Authors' affiliation: Department of Computer Science and Engineering, Metropolitan University, Sylhet, Bangladesh. Questions and feedback: open a discussion on this page.
- Downloads last month
- -