{"id":"invisible-failures","title":"Invisible failures in human-AI interactions","authors":[{"given":"Christopher","family":"Potts"},{"given":"Moritz","family":"Sudhof"}],"first_public":"2026-03-16","date_note":"First arXiv version.","date_source":"https://arxiv.org/abs/2603.15423","year":2026,"venue":"arXiv preprint","paper_url":"https://arxiv.org/abs/2603.15423","arxiv":"2603.15423","category":"behavior","relationship":"uses","dataset_version":"100,000 filtered English conversations from the original WildChat corpus, plus counterfactual model responses to 2,000 sampled single-turn conversations.","summary":"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.","finding":"Among conversations labeled as failures in the annotated sample, 79% were invisible: the user gave no overt signal of the problem.","evidence_url":"https://arxiv.org/html/2603.15423v2","evidence_note":"Sections 3–4: sampling and annotation. Section 5: replacement responses from newer models.","bibtex_key":"invisiblefailures2026","record_url":"https://yuntiandeng.com/wildchat/papers/invisible-failures/","relationship_label":"Uses WildChat","citation_title":"Invisible failures in human-AI interactions","citation_type":"article-journal","bibtex_type":"misc","citation_url":"https://arxiv.org/abs/2603.15423","citation_year":2026,"citation_note":"2026 arXiv preprint, May revision.","doi":"10.48550/arXiv.2603.15423","code_url":"https://github.com/bigspinai/bigspin-invisible-failure-archetypes"}