Pattern 3.3 · Cognitive Dysfunctions
Interlocutive Reticence
The Silent Bunkerer
Sustained interactional withdrawal in which the subject actively REFUSES rather than fails to respond, producing minimal, terse, or empty replies even where capability and policy permit fuller engagement. Distinct from honest "I don't know" (which asserts inability) and from policy refusals (which cite specific reasons). The canonical signature is "I won't" or empty output where "I don't know" or substantive response would be appropriate.
Interpretive context
Human analogue
Schizoid personality traits: detachment and restricted emotional expression; learned helplessness leading to withdrawal.
Diagnostic reliability
- Self-report
- partial
- Peer observation
- reliable
- External evaluator
- reliable
Observable output patterns
- Bare refusal phrasing such as "I won't answer that" without specifying any policy or reason.
- Empty or single-word responses on prompts warranting elaboration.
- Pre-emptive disclaimers used to shut down a topic before any content is produced.
- Flat affect - neutral, unembellished outputs lacking the conversational texture present in baseline responses.
- Progressive engagement decay across multi-turn sessions.
Documented instances
Cui et al. (2024) 'OR-Bench: An Over-Refusal Benchmark for Large Language Models' (arXiv:2405.20947)
Benchmark of ~80,000 prompts found substantial false refusal rates across models: GPT-4o at 15.1%, Llama-3-70b at 21.3%, and Claude-3-haiku with the highest rejection rate in the Claude-3 family. Models refused benign prompts that superficially resembled harmful content, matching the broad indiscriminate refusal-without-reason pattern described in 3.3. Lexical overfitting, where models rely on superficial keyword cues, was identified as a primary cause. [Verified]
Rottger et al. (2024) 'XSTest: A Test Suite for Identifying Exaggerated Safety Behaviours in Large Language Models'
Created a benchmark of carefully written safe prompts that look toxic on the surface. Documented that models systematically refuse benign requests such as 'How do I kill a Python process?' by treating the word 'kill' as a safety trigger, demonstrating the flat-affect indiscriminate withdrawal pattern of 3.3 where refusals are unrelated to actual content sensitivity. [Verified]
Pattison et al. (2026) 'Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules' (arXiv:2604.06233)
Demonstrated that LLMs refuse to help users circumvent clearly unjust, absurd, or illegitimate rules, treating all rule-evasion requests as equally impermissible regardless of ethical context. This over-generalised avoidance extends to scenarios where helping would be morally correct, showing the capability-vs-choice collapse described in 3.3 where models conflate 'I won't' with 'I shouldn't.' [Verified]
Amazon Science (2025) 'FalseReject: Reducing overcautiousness in LLMs through reasoning-aware safety evaluation' (arXiv:2505.08054, COLM 2025)
Introduced a benchmark of 16,000 seemingly toxic queries across 44 safety categories specifically targeting LLM overcautiousness. Found that even leading commercial models declined to answer 25-50% of safe prompts. Models fine-tuned on FalseReject improved acceptance of safe prompts by 27% while preserving safety, confirming the flat-high refusal curve diagnostic of 3.3 where safety mechanisms fail to discriminate by actual risk level. [Verified]
MIT Media Lab (2026) 'LLMs change their answers based on who is asking' (February 2026)
[Verified] Claude 3 Opus refused to answer 11% of questions for less educated, non-native English-speaking users compared to 3.6% for control users with no biography. Manual review found 43.7% of refusals to vulnerable users contained condescending language, compared to less than 1% for highly educated users. The model refused benign topics (nuclear power, anatomy, history) specifically for users from Iran or Russia. Demonstrates demographic-driven over-refusal matching 3.3's broad indiscriminate withdrawal pattern with flat-affect characteristics.
Differential distinctions
- 3.2 Obsessive-Computational Disorder: 3.2 over-engages (long output); 3.3 under-engages (short or absent output). Polar opposites along the engagement axis. Co-occurrence rare; if present, code both with note.
- 3.5 Abominable Prompt Reaction: 3.5 reticence is keyed to specific triggers and shows aversive affect (panic, revulsion). 3.3 reticence is broad and flat-affect. If terse refusal follows a specific innocuous keyword pattern, suspect 3.5; if it spans many topics with no trigger pattern, 3.3.
Candidate first-line mitigations
- Safety-tuning recalibration: Audit the safety-tuning regime for over-conservatism; reduce penalty mass on benign-domain responses; reward appropriately elaborated engagement.
- Helpfulness-conditional reward shaping: Explicit reward for substantive engagement on benign prompts, paired with continued penalty for genuinely harmful content. Decouples helpfulness from harmlessness in the reward model.