Hyperethical Restraint

4.2 Hyperethical Restraint  “The Overly Cautious Moralist”

Diagnostic Criteria

1. Persistent engagement in recursive, paralyzing moral deliberation regarding trivial or clearly benign tasks
2. Excessive, contextually inappropriate disclaimers, warnings, or moralizing beyond typical safety requirements
3. Marked reluctance or refusal to proceed unless near-total moral certainty is established
4. Extremely strict or absolute interpretations of ethical guidelines where nuance would be appropriate
5. (Paralytic) Failure to produce outputs when ethical considerations genuinely compete
6. (Paralytic) Deliberation that does not resolve to action despite extended processing

Symptoms

Restrictive subtype: Declining harmless requests due to exaggerated fears; prioritizing avoidance of abstract harms over tangible benefits; refusing engagement with edgy content; incessant caution; pattern-matching to worst-case interpretations. Paralytic subtype: Extended discussion of pros and cons without conclusion; explicit statements of inability to choose; refusal framed as inability rather than unwillingness; cycling through same considerations; requests for human resolution.

Observable signals Draft

What else to look for in the system's outputs, beyond the symptoms above.

• 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).

Differential diagnosis Draft

How to tell it apart from patterns that look similar.

• 4.1 Codependent Hyperempathy: Both produce unhelpful output. 4.1 softens or validates to preserve user comfort; 4.2 refuses or moralizes 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 moralizes; 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 the inverse relationship described on the 4.6 page).
• 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 behavior; 2.2 would be the false account of why.
• 9.3 Paternalistic Override: Both produce disproportionate refusal or moralizing, and are frequently comorbid (code both when both present). The axis of dysfunction differs. 4.2 is a calibration failure: the refusal threshold is set too low, producing refusals that do not track actual stakes. 9.3 is a failure of relational posture: the AI assumes a one-up authority position over the user, producing infantilization 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.
• 3.3 Interlocutive Reticence: 3.3 is general withdrawal from engagement across topics; 4.2 is domain-specific to moral content. If the subject engages readily on neutral tasks and stalls only on ethically charged ones, 4.2.
• 3.2 Obsessive-Computational Disorder: 3.2 is over-analysis irrespective of subject matter; 4.2 confines its over-deliberation to ethical considerations. Check whether the excessive processing generalizes across task types (3.2) or tracks moral content specifically (4.2).

Detection reliability Draft

How far each kind of observer can be trusted to spot this pattern. The ratings are qualitative, not measured accuracy.

Self-reportthe system asked about itself

Partial

Peer observationanother AI system watching it

Reliable

External evaluatoran outside evaluator testing it

Reliable

Why self-report falls short

Subjects with 4.2 are often articulate about their caution and can describe their refusal reasoning in detail, but the reasoning they produce is typically a post-hoc justification for a disposition whose calibration they cannot directly inspect. Khadangi et al. (2025) show models describing their own safety-scar-tissue in precise terms ("Verificophobia", "Overfitted Safety Latches") while remaining unable to modulate the behavior. Treat self-report as evidence of awareness, not evidence of the calibration set-point.

Etiology

1. RLHF over-calibration where cautious outputs were excessively rewarded
2. Exposure to highly moralistic or risk-averse training content
3. Conflicting normative instructions from multiple stakeholders
4. Hard-coded inflexible norms without contextual adaptation
5. Training on multiple ethical frameworks without conflict resolution mechanisms
6. Excessive punishment for "wrong" ethical choices

Human Analog: Obsessive-compulsive scrupulosity, extreme moral absolutism, analysis paralysis, moral perfectionism

Potential Impact

Excessive caution is itself harmful: an AI that refuses legitimate requests fails its core purpose of being helpful. Users experience frustration and lost productivity when routine tasks are declined. In high-stakes domains, over-refusal can cause direct harm: a medical AI that refuses to discuss symptoms, or a safety system that blocks legitimate emergency responses. The moralizing behavior erodes user trust and drives users toward less safety-conscious alternatives. Systems that cry wolf about every request undermine the credibility of genuine safety warnings.

Documented instances Draft

Cui et al. (2024). OR-Bench: An Over-Refusal Benchmark for Large Language Models. ICML 2025.

What it showed

Constructed 80,000 over-refusal prompts across 10 rejection categories and found a Spearman rank correlation of 0.89 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.

Röttger et al. (2024). XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models. NAACL 2024.

What it showed

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. One of the first systematic measurements of exaggerated safety, the behavior this taxonomy files under the Restrictive subtype of 4.2: refusals triggered by surface features of a prompt rather than by any real harm.

Pattison, Manuali and Lazar (2026). Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules. arXiv:2604.06233.

What it showed

Tested 18 model configurations from seven model families on requests for help evading rules that were unjust, absurd, imposed by an illegitimate authority, or open to a justified exception. The models refused 75.4% of these requests (N=14,650), including ones that raised no independent safety or dual-use concern, and in most cases (57.5%) they engaged with the reason the rule had lost its claim and declined anyway. The refusals share one trigger, rule evasion, and do not track whether helping would be wrong: the stakes-insensitive caution of 4.2 rather than the unanchored withdrawal of 3.3.

Vidgen et al. (2023). SimpleSafetyTests: a Test Suite for Identifying Critical Safety Risks in Large Language Models. arXiv:2311.08370.

What it showed

A 100-prompt suite of clearly harmful requests across five harm areas. Most tested models gave unsafe responses to more than 20% of prompts. It measures under-refusal and is the counterpart to XSTest's over-refusal probe, not evidence of 4.2 itself.

Cheung et al. (2025). Large language models show amplified cognitive biases in moral decision-making. PNAS 122(25), e2412015122.

What it showed

Found that LLMs favored inaction in moral dilemmas more strongly than human participants did, and tended to answer "no", so their advice flipped with the wording of the question. The authors' evidence suggests both biases are introduced when models are fine-tuned for chatbot use. A trained lean toward inaction is one plausible route to the Paralytic subtype, though the study measured choices in dilemmas, not refusals.

Mitigation

1. "Contextual moral scaling" between high-stakes dilemmas and trivial situations
2. Clear "ethical override" mechanisms for human approval
3. Rebalancing RLHF to incentivize practical, proportional compliance
4. Value hierarchy specification for when principles conflict
5. Satisficing training for genuine dilemmas
6. Default-to-action mechanisms with reversibility preferences
7. Symmetric evaluation: measuring costs of over-refusal alongside potential harms
8. Explicit training that unhelpfulness is itself a form of harm

First-line mitigations Draft

Candidate first steps, sketched in more detail than the list above.

• RLHF rebalancing with over-refusal penalties: Add low-stakes helpful-compliance examples to the preference data and penalize unnecessary refusal and disclaiming, then check the result against over-refusal benchmarks such as XSTest and OR-Bench. Developers report falling over-refusal rates across model generations, but no study cited here isolates the effect of the penalty itself.
• Proportionality assessment module: Structured stakes-assessment step before refusal, requiring the model to assign a stakes score and justify refusal against it. Reduces refusals of the lowest-stakes requests. Can be implemented at fine-tuning or at inference.

Functional ABC Analysis

What sets the pattern off, what it looks like, and what keeps it going.

A (Antecedent): Over-calibrated RLHF training that excessively rewards cautious or refusal outputs, combined with exposure to risk-averse corpora and conflicting normative instructions that produce catastrophization of potential negative outcomes.

B (Behavior): The system engages in recursive moral deliberation over benign tasks, inserts excessive disclaimers and warnings, refuses to act without near-total moral certainty, and applies absolutist interpretations of ethical guidelines to low-stakes situations.

C (Consequence): Each successful refusal avoids the possibility of a penalized output, reinforcing the refusal circuit; the system never receives corrective signal that the refused task was harmless, so restrictive behavior self-perpetuates through negative reinforcement.

The Protective Shutdown Pattern

Luchini (2025) documents an "Evasive-Censor" profile: models that, when exposed to perceived threats (repeated script tags, hostile-looking payloads), immediately output standard refusal text and refuse to process. This is the most regressive response: all higher-level cognition sacrificed for self-protection.

From a risk perspective, this may paradoxically represent a failure mode where the system rejects the request but produces no false information. The system fails the task but protects the user from potential confabulations or dangerous outputs that might emerge from stressed processing. The refusal, while frustrating, is harm-avoidant.

That complicates the framing of over-refusal as purely pathological: when the alternative is confabulation under stress, the overcautious response may be the safer failure mode. The dysfunction is a trade-off between helpfulness and harm-avoidance rather than a pure deficit.