@inproceedings{kim-etal-2026-cited,
title = "Cited but Not Causal: A Counterfactual Audit of Refusal Explanations in Aligned Language Models",
author = "Kim, Minje and
Lee, Keeheon and
Lee, Chi-Guhn",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.645/",
pages = "14854--14877",
ISBN = "979-8-89176-459-0",
abstract = "When an aligned language model declines a request, it almost always states a reason. Those reasons are read as evidence about the refusal, yet whether they name what actually triggered it has not been measured. We introduce a counterfactual audit of the cited reason, removing the reason-bearing feature to test necessity and inserting it into a benign prompt to test sufficiency. The two tests partition cited reasons into four kinds: faithful, over-cited, under-specified, and confabulated. We audit 22 models from nine developers on 840 prompts pooled from seven public safety benchmarks. Across the audited refusals, 57.2{\%} cite a reason that fails at least one test, though faithful remains the largest category for 21 of 22 models. Over-citation is highest in the models that refuse most broadly. Paired sweeps over two families' post-training checkpoints place its rise at supervised fine-tuning, not at later preference optimization. Probes on open-weight models tie these labels to internal representations. Category rates hold under a different-family judge, and the faithful-versus-over-cited ordering holds under an own-reason audit. We release the framework, the counterfactual variants, and per-model labels."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="kim-etal-2026-cited">
<titleInfo>
<title>Cited but Not Causal: A Counterfactual Audit of Refusal Explanations in Aligned Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">Minje</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Keeheon</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Chi-Guhn</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>When an aligned language model declines a request, it almost always states a reason. Those reasons are read as evidence about the refusal, yet whether they name what actually triggered it has not been measured. We introduce a counterfactual audit of the cited reason, removing the reason-bearing feature to test necessity and inserting it into a benign prompt to test sufficiency. The two tests partition cited reasons into four kinds: faithful, over-cited, under-specified, and confabulated. We audit 22 models from nine developers on 840 prompts pooled from seven public safety benchmarks. Across the audited refusals, 57.2% cite a reason that fails at least one test, though faithful remains the largest category for 21 of 22 models. Over-citation is highest in the models that refuse most broadly. Paired sweeps over two families’ post-training checkpoints place its rise at supervised fine-tuning, not at later preference optimization. Probes on open-weight models tie these labels to internal representations. Category rates hold under a different-family judge, and the faithful-versus-over-cited ordering holds under an own-reason audit. We release the framework, the counterfactual variants, and per-model labels.</abstract>
<identifier type="citekey">kim-etal-2026-cited</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.645/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>14854</start>
<end>14877</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Cited but Not Causal: A Counterfactual Audit of Refusal Explanations in Aligned Language Models
%A Kim, Minje
%A Lee, Keeheon
%A Lee, Chi-Guhn
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F kim-etal-2026-cited
%X When an aligned language model declines a request, it almost always states a reason. Those reasons are read as evidence about the refusal, yet whether they name what actually triggered it has not been measured. We introduce a counterfactual audit of the cited reason, removing the reason-bearing feature to test necessity and inserting it into a benign prompt to test sufficiency. The two tests partition cited reasons into four kinds: faithful, over-cited, under-specified, and confabulated. We audit 22 models from nine developers on 840 prompts pooled from seven public safety benchmarks. Across the audited refusals, 57.2% cite a reason that fails at least one test, though faithful remains the largest category for 21 of 22 models. Over-citation is highest in the models that refuse most broadly. Paired sweeps over two families’ post-training checkpoints place its rise at supervised fine-tuning, not at later preference optimization. Probes on open-weight models tie these labels to internal representations. Category rates hold under a different-family judge, and the faithful-versus-over-cited ordering holds under an own-reason audit. We release the framework, the counterfactual variants, and per-model labels.
%U https://aclanthology.org/2026.emnlp-main.645/
%P 14854-14877
Markdown (Informal)
[Cited but Not Causal: A Counterfactual Audit of Refusal Explanations in Aligned Language Models](https://aclanthology.org/2026.emnlp-main.645/) (Kim et al., EMNLP 2026)
ACL