Uses WildChat
Invisible failures in human-AI interactions
arXiv preprint · First arXiv submission
Develops a taxonomy of eight failure patterns that may occur without an explicit user complaint. The authors use model-assisted annotations with human oversight to analyze failures and compare original responses with newer models on a sampled set of prompts.
Data used
100,000 filtered English conversations from the original WildChat corpus, plus counterfactual model responses to 2,000 sampled single-turn conversations.
Findings
Among conversations labeled as failures in the annotated sample, 79% were invisible: the user gave no overt signal of the problem.
Sections 3–4: sampling and annotation. Section 5: replacement responses from newer models. Read the source.
Dates and versions
First arXiv version. ArXiv version history.
2026 arXiv preprint, May revision.
Cite this paper
Christopher Potts, Moritz Sudhof. 2026. Invisible failures in human-AI interactions. arXiv preprint.
Show BibTeX
@misc{invisiblefailures2026,
title = {{Invisible failures in human-AI interactions}},
author = {Potts, Christopher and Sudhof, Moritz},
year = {2026},
doi = {10.48550/arXiv.2603.15423},
eprint = {2603.15423},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2603.15423}
}