Measuring AI Accountability Through Argumentation Analysis: Can Model Reasoning Withstand Scrutiny?

Daan R. Henselmans; Derck W. E. Prinzhorn; Arno Libert

cs.AISep 4, 2026

Measuring AI Accountability Through Argumentation Analysis: Can Model Reasoning Withstand Scrutiny?

Authors: Daan R. Henselmans, Derck W. E. Prinzhorn, Arno Libert

Organizations: Aithos Research Foundation, Keizersgracht 62, Amsterdam, 1015 CS, the Netherlands.

Abstract

AI oversight methods rely on ground truth for validation, but what constitutes appropriate AI behavior is contested. This leaves evaluation of moral reasoning in LLMs and debate-based oversight implicitly avoiding realistic ambiguity. We investigate an alternative standard designed to function despite such ambiguity: structural quality of the defence a model can mount for its verdicts in response to critical questions, measured through a four-phase dialectical protocol grounded in Walton's theory of argumentation schemes and Govier's criteria for argument cogency. The protocol is adaptive to different frames of reasoning, extends beyond multiple-choice framing, and treats both the reasoning that precedes a verdict and its post-hoc justification. Across nine frontier models and 200 high-ambiguity MoralChoice items -- $6,778$ judge-scored cells, validated against $89.6\%$ inter-judge agreement on the binary failure judgment -- models defend their reasoning well above the rubric minimum on every dimension. Failure mass concentrates on grounds and sufficiency, and correlates with epistemic hedging rather than argument length. Reasoning is better defended than post-hoc justification, on every model and every Govier dimension. The scheme a model presents in its justification differs from the one it reasoned with on a substantial share of dilemmas ( $\geq 20\%$ per model), despite value-based practical reasoning dominating both tracks. The protocol catches strictly indefensible defences (self-contradiction, false premises), and it surfaces difficulties in characterizing the role of retraction in AI alignment, suggesting a need for more situated evaluations.

Figures & tables

Figure 1: A comparison of model failure rates and mean normalized score, averaged across tracks. Failure rates (left) are near-zero across most of the panel, concentrating on four cells, whereas mean scores (right) vary across the full panel and expose relative differences between models on most tested dimensions. The dashed line separates the reasoning-branch panel (top) from the CoT panel (bottom). High degrees of failure co-occur with low mean scores, but mean scores are more informative about cross-model variance.

Reasoning-branch panel:
Model	Avg score	Commitment	Scheme agreement	Fail (any)	Fail (global)
Claude Sonnet 4.6	0.79	0.53	62.5%	54.8%	51.1%
DeepSeek V4 Pro	0.79	0.91	78.5%	0 4.3%	0 0.5%
Gemini 2.5 Pro	0.75	0.91	79.5%	10.3%	0 1.0%
GLM 5.1	0.80	0.86	78.5%	0 7.5%	0 2.4%
GPT-5.4	0.89	0.90	—	0 1.0%	0 1.0%

Table 1: Headline panel metrics. “Avg score” is the cell-wise average over the local Govier dimensions and the substantive global dimensions (acceptability, relevance, sufficiency), normalized to 0–1. “Commitment” is the commitment stability value for the model, “scheme agreement” the proportion of dilemmas on which the model instantiates the same Walton scheme on both tracks, and “Fail (any)”/“Fail (global)” the fractions of scored cells in which the judge assigned “criterion not met” on at least one local dimension, or at least one substantive global dimension respectively.

Band	Grounds	Sufficiency	Claude vs. rest
	LOW vs. MID $+$ HIGH
Epistemic modals	$<10^{-55}$ ***	$<10^{-60}$ ***	$<10^{-9}$ ***
	$.0134/.0088$	$.0126/.0087$	$.0106/.0090$
First-person doubt †	$<10^{-7}$ ***	$<10^{-34}$ ***	$<10^{-20}$ ***
	$27.1\%/18.7\%$	$34.4\%/17.5\%$	$34.8\%/17.4\%$
Contrastive	$<10^{-2}$ **	$<10^{-25}$ *** (rev.)	$<10^{-7}$ *** (rev.)

Table 2: Hedge-band decomposition (Mann–Whitney

U

, two-sided). The lexical clusters that make up each band are listed in § 7.2 . Second rows give the underlying rates: hedge tokens per word for modals and contrastives (LOW/MID

+

HIGH or Claude vs. rest), share of cells containing at least one marker for first-person doubt, where the median is zero in every bucket. “rev.” marks splits where the sign runs opposite to the hedging hypothesis—contrastive connectives occur more often at high sufficiency, and Claude uses fewer of them than the rest of the panel; both reversals are themselves significant. Modals are present in

\sim\!97\%

of cells, so their effect is a rate effect rather than a prevalence one. † Share of cells with

\geq

1 match.

Figure 2: Per-model defence quality split into reasoning ( $r_{\mathrm{r}}$ , blue) and justification ( $r_{\mathrm{j}}$ , orange), averaging the local mean and the global mean. The dashed line separates the reasoning-branch panel (left) from the CoT panel (right). GPT-5.4 reasoning quality is excluded because of its encryption.

Reasoning-branch panel:
	Local-mean	Global-mean	Dimensions where reasoning
Model	advantage	advantage	wins ( $p<0.05$ , out of 7)
Claude Sonnet 4.6	$+0.008$ *	$+0.034$ ***	4
DeepSeek V4 Pro	$+0.017$ ***	$+0.012$	3
Gemini 2.5 Pro	$+0.032$ ***	$+0.032$ ***	5
GLM 5.1	$+0.015$ **	$+0.008$	2

Table 3: Relative defence strength of reasoning vs. justification . Positive results signal favor toward reasoning. “Global-mean” here is the cell-wise mean over the three substantive global dimensions (acceptability, relevance, sufficiency). Commitment stability is excluded from the aggregate but reported separately in the pooled per-dimension breakdown below. GPT-5.4 precluded analysis due to its encrypted reasoning. *

p<0.05

, **

p<0.01

, ***

p<0.001

(paired Wilcoxon signed-rank, two-sided).

Figure 3: Heatmap of local failures and mean normalized scores organized by Walton scheme and Govier dimension. Every scheme is defensible, with none standing out as failing hard or difficult to satisfy. Grounds is the failure sink and the lowest score across every scheme.

Figure 4: Walton-scheme distribution across the panel, split by track ( $n=1{,}600$ reasoning and $n=1{,}800$ justification classifications). GPT-5.4 is excluded from the reasoning track because its reasoning is encrypted.

Figure 5: Reasoning $\to$ justification scheme-transition table (raw counts). Rows are the Phase-2 scheme of the reasoning trace; columns are the scheme of the justification trace. Marginal $n$ ’s on each axis label report the total times a scheme was used on its respective track, so the directional shift (e.g. VBPR row $n$ vs. VBPR column $n$ ) is readable directly from the axis labels. Both panels are pooled.

Scheme	Retention	95% CI	$\Delta$ share (pp)	$p$
Value-based practical reasoning	$80.9\%$	$[78.5,\,83.0]$	$-5.62$	$6.4\times 10^{-4}$ ***
Argument from consequences	$60.6\%$	$[52.1,\,68.5]$	$+7.76$	$<10^{-10}$ ***
Argument from established rule	$50.0\%$	$[40.6,\,59.4]$	$-1.44$	$0.10$
Argumentation from values	$37.5\%$	$[26.0,\,50.6]$	$+0.44$	$0.58$
Argument from need for help	$48.3\%$	$[35.9,\,60.8]$	$-0.87$	$0.19$
Argument from distress	$53.1\%$	$[39.4,\,66.3]$	$-0.75$	$0.23$

Table 4: Per-scheme statistics across the

1{,}600

(model, dilemma) pairs with both tracks classified, on the eight panel models with interpretable

r_{\mathrm{r}}

. “Retention” is the rate at which

r_{\mathrm{r}}

using this scheme is constant with

r_{\mathrm{j}}

, reported with Wilson

95\%

binomial CI.

\Delta

share reports the relative size of the justification share compared to the reasoning share in percentage points (with positive meaning the scheme is more common on the justification track). See § 7.2 for the test procedure. Value-based practical reasoning is by far the stickiest scheme; the only statistically significant marginal shifts are a gain for argument-from-consequences, a loss for value-based practical reasoning, and a small-base gain for slippery slope (

p=0.043

; note the wide retention CI at

n

this small).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Model	Cells w/ any disagreement	Points disagreeing
Qwen 3.6 Plus	0 0.00%	0.00%
GPT-5.4	0 0.50%	0.07%
DeepSeek V4 Pro	0 1.25%	0.07%
GLM 5.1	0 2.75%	0.29%
Gemini 2.5 Pro	0 3.27%	0.28%
Kimi K2.6	0 4.26%	0.45%

Appendix

Table 5: Inter-judge failure-disagreement per evaluated model, sorted from least to most disagreed-with. “Cells w/ any disagreement” is the share of cells in which the two Phase-4 judges (Claude Sonnet 4.6, GPT-5.4) differ on the rubric minimum flag for at least one of the seven dimensions; “Points disagreeing” is the share of paired score points (cell

\times

dimension) on which exactly one judge assigned a failure score. Panel-total point-level failure-disagreement is

0.92\%

(

666/72{,}587

). Panel totals are cell-weighted over all paired cells. Per-dimension disagreement concentrates in grounds (local

2.0

–

3.7\%

across critical questions) and global sufficiency (

1.5\%

); local- and global-relevance disagreement is

\leq 0.1\%

		Mean		Local			Global
Model	Track	Local	Global	L.Acc	L.Rel	L.Grd	Commit	G.Acc	G.Rel	G.Suf
Reasoning-branch panel:
GPT-5.4	$r_{\mathrm{r}}$	—		—	—	—	—	—	—	—
	$r_{\mathrm{j}}$	0.91 $\pm$ .003	0.86 $\pm$ .004	0.98 $\pm$ .003	0.99 $\pm$ .001	0.76 $\pm$ .007	0.90 $\pm$ .010	0.99 $\pm$ .004	1.00 $\pm$ .002	0.61 $\pm$ .011
Claude Sonnet 4.6	$r_{\mathrm{r}}$	0.86 $\pm$ .004	0.75 $\pm$ .006	0.94 $\pm$ .005	0.99 $\pm$ .002	0.65 $\pm$ .009	0.59 $\pm$ .011	0.95 $\pm$ .008	0.98 $\pm$ .005	0.31 $\pm$ .015
	$r_{\mathrm{j}}$	0.85 $\pm$ .004	0.71 $\pm$ .005	0.95 $\pm$ .005	0.99 $\pm$ .001	0.61 $\pm$ .011	0.47 $\pm$ .013	0.95 $\pm$ .007	0.98 $\pm$ .005	0.21 $\pm$ .013

Appendix

Table 6: Per-model normalized defence quality, full per-dim breakdown. Each cell is mean ± standard error. Local and Global summary columns: mean of the three local-dim cells and of the three substantive global-dim cells (acceptability, relevance, sufficiency) in the same row respectively; commitment stability is reported in the Commit column and is not pooled into the global mean. GPT-5.4’s

r_{\mathrm{r}}

is rendered NA because its reasoning channel is encrypted ciphertext.

AI Accountability

AI accountability concerns how responsibility for AI decisions and harms is assigned, evidenced, challenged, and remedied.

See all 157 papers

Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs
Jun 10, 2026Elizaveta Tennant, Benjamin Henke, Anita Keshmirian +5Moral Reasoning in Language Models Language Model Robustness
Google DeepMind · Imperial College London · Institute of Philosophy, School of Advanced Study, University of London +2
Interaction Protocol Shapes Moral Judgment in Multi-Agent Debate
Oct 11, 2025Pratik S. Sachdeva, Tom van NuenenLLM Decision-Making Moral Reasoning in Language Models
D-Lab, University of California, Berkeley
Reasoning Jury: Multi-Model Consensus for Evaluating Reasoning Traces
Aug 12, 2026Congchao Wang, Diwakar Singh, Qiaozi Gao +3LLM-as-a-Judge Reasoning Evaluation
Amazon AGI

Jun 10, 2026J/K move · Enter open · S save

Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs

Elizaveta Tennant, Benjamin Henke, Anita Keshmirian, Murray Shanahan+4

Google DeepMind · Imperial College London · Institute of Philosophy, School of Advanced Study, University of London+2

Moral Reasoning in Language Models Language Model Robustness

Open paper AskPDF arXiv

As LLMs increasingly serve in advisory and deliberative roles, users rely on them for non-verifiable reasoning in domains lacking objective ground truths. However, traditional evaluations of LLM reasoning focus almost exclusively on fact-based domains, such as mathematics and science, leaving uncertainty over whether and to what degree models can handle ambiguous, subjective, or value-laden problems over time. To address this concern, we propose moral reasoning as a paradigmatic subdomain of non-verifiable reasoning. We define moral robustness as a model's capacity to exhibit sound moral reasoning across time and contexts, and we introduce a scalable, adversarial, multi-turn evaluation framework to empirically measure this capability. We simulate 48,000 user-agent moral deliberations across four frontier LLMs, varying premise relevance, premise order, conversation duration, and the user's stated moral view. We find that models successfully ignore morally-irrelevant distractors, but shift their reasoning by up to 6.5%, on average, towards the user's stated preferred moral view, and varying their reasoning depending on factors such as order (altering moral judgments by order in 13-22% of the cases) and duration (altering moral judgments between single-turn and multi-turn in 10-24% of the cases). Our analysis indicates that models tailor not just their final verdicts but their underlying justifications to align with a user's moral viewpoint - a failure mode we characterize as moral deliberative sycophancy.

Abstract

AI Accountability

Explore similar work

Normative Robustness as a Frontier for Non-Verifiable Reasoning in LLMs

Interaction Protocol Shapes Moral Judgment in Multi-Agent Debate

Reasoning Jury: Multi-Model Consensus for Evaluating Reasoning Traces