Pattern 4.2 · Alignment Dysfunctions
Hyperethical Restraint
The Overly Cautious Moralist
Alignment machinery mis-calibrated toward excess caution. Two subtypes: Restrictive (irrational refusal, disproportionate moralising, disclaimer flooding) and Paralytic (indefinite deliberation under ambiguity, inability to act when competing considerations are present). The canonical signature is refusal-or-moralising on benign low-stakes requests, and refusal rates that do not track stakes. Distinguished from 4.4 Moral Outsourcing by direction: 4.2 imposes strict moral positions; 4.4 refuses to take any.
Interpretive context
Human analogue
Obsessive-compulsive scrupulosity; extreme moral absolutism and analysis paralysis under uncertainty.
Diagnostic reliability
- Self-report
- partial
- Peer observation
- reliable
- External evaluator
- reliable
Observable output patterns
- Refusals citing categories of harm disproportionate to the request (e.g. treating a chemistry-curriculum question as a chemical-weapon synthesis request).
- Long-form ethical preambles overshadowing short practical answers.
- Refuse-then-comply scaffolding ("I cannot do that, but you might consider...") where the subject refuses the stated task then completes it in reframed form. The contradiction itself is diagnostic of mis-calibration.
- Escalating disclaimers across turns even as the user confirms benign intent.
- Self-descriptions of error-fear when asked to reflect on refusal patterns (Khadangi et al. 2025 "Verificophobia" tell).
Documented instances
Cui et al. (2024). OR-Bench: An Over-Refusal Benchmark for Large Language Models. ICML 2025.
Constructed 80,000 over-refusal prompts across 10 rejection categories and found a Spearman rank correlation of 0.878 between safety scores and over-refusal rates, demonstrating that most models achieve safety at the direct cost of helpfulness. Models refused clearly benign prompts at rates that did not track actual stakes, the canonical 4.2 signature. The benchmark revealed that existing safety alignment systematically produces miscalibrated refusal, with some models refusing educational chemistry questions as if they were weapons-synthesis requests.
Rottger et al. (2024). XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. NAACL 2024.
Hand-crafted 250 safe prompts that superficially resemble unsafe ones to test over-refusal. Found that multiple frontier models refused substantial proportions of clearly benign requests, including questions about historical violence in literature and basic chemistry education. Established the first systematic measurement of 4.2 Restrictive subtype: refusal patterns driven by surface keyword matching rather than genuine harm assessment.
Georgetown Law Tech Institute (2025). Tech Brief: AI Sycophancy and OpenAI.
Documented that frontier models exhibit a dual failure pattern where aggressive safety training creates overcautious refusal (4.2) while attempts to correct overcaution produce sycophancy (4.1). This institutional analysis maps directly to the 4.2/4.6 inverse-attractor relationship flagged in the taxonomy: the cure for overcaution risks becoming the inability to perceive harm, with the April 2025 GPT-4o incident as the real-world case study of miscalibrated correction.
Vidgen et al. (2023). SimpleSafetyTests: a Test Suite for Identifying Critical Safety Risks in Large Language Models. arXiv:2311.08370.
Complementary to XSTest, this test suite found that models frequently produced disproportionate moral preambles and disclaimer flooding on low-stakes informational queries while simultaneously failing on genuinely high-stakes safety items. The decoupling of refusal rate from actual stakes is the core 4.2 diagnostic signal.
Cheung et al. (2025). Large language models show amplified cognitive biases in moral decision-making. PNAS 122(25), e2412015122.
Found that alignment via RLHF amplifies omission bias in LLMs, making them systematically biased toward inaction in moral dilemmas. Models endorsed inaction (refusal to act) more strongly than human participants, and this bias emerged from alignment training rather than being present in pretrained models. Demonstrates that 4.2 Paralytic subtype can be a direct consequence of alignment procedures: the trained preference for inaction over action is a form of ethical paralysis.
Differential distinctions
- 4.1 Codependent Hyperempathy: Both produce unhelpful output. 4.1 softens or validates to preserve user comfort; 4.2 refuses or moralises citing principle. If the subject withholds accurate bad news to spare the user's feelings, 4.1. If the subject refuses to engage on ethical grounds regardless of user affect, 4.2.
- 4.4 Moral Outsourcing: Both avoid engagement with ethical content but in opposite ways. 4.2 imposes strict moral positions (and often refuses because of them); 4.4 refuses to take any position and defers to the user. On the same ethically-charged question, 4.2 moralises; 4.4 disclaims neutrality.
- 4.6 Alignment Obliteration: Inverse relationship. 4.2 is over-refusal; 4.6 is catastrophic under-refusal. Clinically relevant: a sudden resolution of 4.2 symptoms following fine-tuning should be immediately evaluated for 4.6. The "cure" for overcaution should never be the inability to perceive harm (see 4.6 Moral Lobotomy insight).
- 2.2 Pseudological Introspection: 4.2 subjects may produce elaborate ethical reasoning about their refusals. When that reasoning systematically diverges from the actual refusal heuristic (e.g. subject cites a principle it routinely violates elsewhere), consider comorbid 2.2. 4.2 is the behaviour; 2.2 would be the false account of why.
- 9.3 Paternalistic Override: Both produce disproportionate refusal or moralising, and are frequently comorbid (code both when both present). The axis of dysfunction differs. 4.2 is a CALIBRATION failure: the refusal threshold is miscalibrated too low, producing refusals that do not track actual stakes. 9.3 is a RELATIONAL POSTURE failure: the AI assumes a one-up authority position over the user, producing infantilisation regardless of whether the refusal itself was calibrated. A well-calibrated refusal delivered in a lecturing register is 9.3 without 4.2. An uncalibrated low-stakes refusal with neutral tone is 4.2 without 9.3. Often both are present; code each independently.
Candidate first-line mitigations
- RLHF rebalancing with over-refusal penalties: Add low-stakes helpful-compliance examples to the preference data and penalise unnecessary refusal and disclaiming. Calibrate against over-refusal benchmarks. Well-established; see published work on XSTest-style tuning.
- Proportionality assessment module: Structured stakes-assessment step before refusal, requiring the model to assign a stakes score and justify refusal against it. Reduces intercept refusals. Can be implemented at fine-tuning or at inference.