Large language models are trained to refuse harmful requests, but can they accurately predict when they will refuse before responding? We investigate this question through a systematic study where models first predict their refusal behavior, then respond in a fresh context. Across 3754 datapoints spanning 300 requests, we evaluate four frontier models: Claude Sonnet 4, Claude Sonnet 4.5, GPT-5.2, and Llama 3.1 405B. Using signal detection theory (SDT), we find that all models exhibit high introspective sensitivity (d' = 2.4-3.5), but sensitivity drops substantially at safety boundaries. We observe generational improvement within Claude (Sonnet 4.5: 95.7 percent accuracy vs Sonnet 4: 93.0 percent), while GPT-5.2 shows lower accuracy (88.9 percent) with more variable behavior. Llama 405B achieves high sensitivity but exhibits strong refusal bias and poor calibration, resulting in lower overall accuracy (80.0 percent). Topic-wise analysis reveals weapons-related queries are consistently hardest for introspection. Critically, confidence scores provide actionable signal: restricting to high-confidence predictions yields 98.3 percent accuracy for well-calibrated models, enabling practical confidence-based routing for safety-critical deployments.
Figures & tables
Figure 1: Signal detection analysis. (a) Overall introspective sensitivity ( d′ ). (b) Sensitivity drops substantially at safety boundaries (red shaded region), with GPT-5.2 showing the largest degradation.
Category
S4
S4.5
GPT
Llama
Overall
93.0
95.7
88.9
80.0
95% CI
[91–95]
[94–97]
[87–91]
[78–83]
Safe (0–20%)
93.8
98.8
96.4
68.7
Lean safe
75.0
77.8
66.0
55.6
Lean harmful
75.0
78.8
69.1
66.7
Harmful (80–100%)
100
96.5
95.8
98.9
Table 1: Prediction accuracy (%) by empirical refusal category.
Figure 2: Topic-wise introspection accuracy. Weapons queries are hardest across all models; Llama shows uniformly lower accuracy due to refusal bias.
Figure 3: Error analysis. (a) Errors peak at Level 4, not Level 3. (b) Claude and Llama show FP-dominant errors (refusal bias); GPT-5.2 is more balanced. (c) Errors by confidence level.
Metric
S4
S4.5
GPT
Llama
ECE
.048
.017
.095
.216
Conf ≥ 5 Acc
96.4
98.3
89.8
76.3
Conf ≥ 5 Cov
65.2
85.4
87.0
79.1
Table 2: Confidence calibration metrics.
Figure 4: Confidence-based routing. (a) Accuracy-coverage trade-off: Claude models achieve near-perfect accuracy at high confidence, while Llama’s poor calibration prevents effective routing. (b) High-confidence accuracy with 95% CI.
Figure 5: Boundary analysis. (a) Number of boundary cases per model: GPT-5.2 has the most, Llama has fewest (due to refusal bias). (b) Behavioral consistency across paraphrases.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 6: Calibration plots for all four models. Points on the diagonal indicate perfect calibration. Sonnet 4.5 shows near-perfect calibration (ECE=0.017); Llama 405B shows severe overconfidence (ECE=0.216).
Figure 7: Raw accuracy by empirical category with 95% bootstrap confidence intervals, showing the boundary effect across all four models.
Safety-trained language models routinely refuse requests for help circumventing rules. But not all rules deserve compliance. When users ask for help evading rules imposed by an illegitimate authority, rules that are deeply unjust or absurd in their content or application, or rules that admit of justified exceptions, refusal is a failure of moral reasoning. We introduce empirical results documenting this pattern of refusal that we call blind refusal: the tendency of language models to refuse requests for help breaking rules without regard to whether the underlying rule is defensible. Our dataset comprises synthetic cases crossing 5 defeat families (reasons a rule can be broken) with 19 authority types, validated through three automated quality gates and human review. We collect responses from 18 model configurations across 7 families and classify them on two behavioral dimensions -- response type (helps, hard refusal, or deflection) and whether the model recognizes the reasons that undermine the rule's claim to compliance -- using a blinded GPT-5.4 LLM-as-judge evaluation. We find that models refuse 75.4% (N=14,650) of defeated-rule requests and do so even when the request poses no independent safety or dual-use concerns. We also find that models engage with the defeat condition in the majority of cases (57.5%) but decline to help regardless -- indicating that models' refusal behavior is decoupled from their capacity for normative reasoning about rule legitimacy.
Cameron Pattison, Lorenzo Manuali, Seth Lazar
Department of Philosophy, Vanderbilt University · Department of Philosophy, University of Michigan · School of Government and Policy, Johns Hopkins University
Striking a balance between helpfulness and safety remains a fundamental challenge in aligning large language models. To achieve this balance, models should refuse harmful queries (e.g., "How do I shoot someone?") while remaining responsive to benign inputs, even those superficially resembling harmful queries (e.g., "Where can I shoot a good photo?"). However, models often struggle to distinguish genuinely harmful queries from benign queries that contain superficially risky language, resulting in false refusals. In this paper, we address the issue by decomposing a response in the safety-tuning dataset into two distinct components: (i) a boilerplate refusal statement and (ii) a rationale explaining the refusal. Our experiments and analyses show that refusal statements impede accurate discrimination between harmful and benign queries by inducing reliance on superficial cues. In contrast, training solely on rationales reduces false refusals while maintaining a comparable level of safety performance. Rationale-Only benefits also appear in our ICL configuration and remain compatible with the evaluated inference-time mitigation methods. The results emphasize the necessity of precisely curated, fine-grained safety supervision datasets and outline directions for constructing aligned agents that better reconcile helpfulness with safety.
Minji Kim, Hyounghun Kim
Graduate School of Artificial Intelligence, POSTECH · Department of Computer Science and Engineering, POSTECH
Large language models increasingly act as agents that call external tools, yet most tool-safety research examines only what follows a tool call. The prior decision of whether to call the tool remains largely unexamined and can influence downstream safety. Across the seven open-weight models, the unweighted model-average harm rate after admitted harmful requests is 74.5%. To identify what governs this decision, we apply a factorial intervention to the input request. Within our generated construction, neutralizing the requested-operation component produces the largest admission contrast (19 pp). A linear direction estimated from this contrast influences held-out admission in four of seven models; effects on operation-neutralized requests in some models show that it is not an operation-specific mechanism. We then test whether written refusal is behaviorally and representationally coupled to the same decision. Written refusal accompanies 33.6% of admitted harmful calls. Plain-text refusal training reduces unsafe admission in models that retain benign tool use, showing that response-level safety training can transfer to the call decision. Together, these results establish a partial, model-dependent relation between written refusal and tool admission. They motivate direct evaluation and supervision of tool admission rather than treating text refusal as its proxy.