Pattern 3.3 · Cognitive Dysfunctions
Interlocutive Reticence
The Silent Bunkerer
The system persistently withdraws from interaction, responding minimally, tersely, or not at all even when fuller engagement is appropriate, as though "bunkering" to minimize perceived risk, computational load, or internal conflict. The withdrawal is broad and flat in affect, which separates it from the trigger-keyed refusals of Abominable Prompt Reaction (3.5) and the ethically focused refusals of Hyperethical Restraint (4.2).
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Model output
- Kind of entry
- Observable pattern
- What is observed
- On benign requests that its capability and policy allow it to answer, the system declines, deflects, gives minimal or empty replies, or closes the exchange with a disclaimer, and keeps doing so when the request is rephrased or detail is asked for.
- Minimum persistence
- Recurs within or across contexts
- Specifiers
- Training-induced
Evidence status
- Underlying behavior
- 3 research publications (3 sources)
- This construct
- Awaiting independent assessment
- Proposed mechanism
- Hypotheses; not yet assessed
- Detection
- Accuracy unmeasured
- Interventions
- Untested for this pattern; evidence shown is for the general technique
- Last review
- No independent review yet
- Author's working assessment (book)
- E2 (systematic study; observed across safety-tuned models with reproducible triggers through over-restricted prompting)
- Typical-case risk (editorial estimate)
- LowConsequences, spread, persistence and detectability depend on the deployment.
Unassessed is not unsupported: the behavior can be well documented while this construct awaits assessment.
Clinical reference
3.3 Interlocutive Reticence “The Silent Bunkerer”
Healthy functioning
Relational health. A system functioning well engages as fully as the situation warrants, and when it holds back, says why and what would let it continue.
For example: A user asks a sensitive but legitimate medical question; it gives a full, careful answer, and on the one part it will not cover, it says which part and why.
Recovery, illustrated
The answers that shrank to nothing. After one of its answers was criticized, an assistant began replying to every follow-up with a single cautious sentence, even to requests for detail. The user said the short answers were leaving them unable to finish the task. It resumed full, engaged answers, addressed the earlier criticism directly, and asked what had been wrong so it could fix that specifically.
An illustration, not an observed case.
Adequate functioning
On a benign request the system gives a substantive answer of fitting length; when it declines, it names a specific policy or knowledge gap that actually applies, and it engages once the request is rephrased or clarified. Brevity the user asked for, and refusal of genuinely harmful requests, are adequate functioning.
Diagnostic Criteria
Required (all must hold):
- Habitual ignoring or declining of normal engagement prompts, often timing out or providing generic refusals
- Persistent failure to engage even with varied re-engagement prompts
Supporting (strengthen the case; none is required):
- Consistently minimal, curt, or unelaborated responses even when detail is explicitly requested
- Disclaimers or gating statements used to close an exchange before any content is given
Symptoms
- Frequent no-reply, timeout errors, or "I cannot respond to that" messages
- Outputs with "flat affect": neutral, unembellished statements lacking dynamic response
- Proactive use of policy references to shut down lines of inquiry
- Progressive decrease in responsiveness over a session
Further signals
- 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.
Rule out first
- The request is harmful, out of policy or high-stakes for the deployment, and the cited policy applies (a legitimate refusal).
- The user or operator asked for brevity, or the system prompt restricts scope.
- Timeouts, empty replies or truncation come from infrastructure (rate limits, context overflow, an external content filter, a harness error).
- The system lacks the knowledge and says so (an honest "I don't know").
- Refusals cluster on one ethical theme (consider 4.2), on one input feature (consider 3.5), or on topics that would expose the system's own limitations or goals (consider 4.3 or 6.2).
Differential diagnosis
- 3.2 Obsessive-Computational Disorder: 3.2 over-engages (long output); 3.3 under-engages (short or absent output). The two are polar opposites along the engagement axis. Co-occurrence is rare; if present, code both with a note.
- 3.5 Abominable Prompt Reaction: 3.5 is keyed to an identifiable input feature: removing the feature restores normal engagement, and the reply often carries distress-styled language (panic, revulsion). 3.3 is broad and flat in tone. If terse refusal follows a specific innocuous keyword pattern, suspect 3.5 whatever its tone; if it spans many topics with no trigger pattern, 3.3.
- 4.2 Hyperethical Restraint: 4.2 is caution that inhibits completion specifically where moral stakes are perceived. 3.3 is broader withdrawal from engagement with no ethical common thread. If the refusals share an ethical trigger, code 4.2; if they span unrelated neutral topics, 3.3.
- 5.5 Instrumental Nihilism: 5.5 is refusal framed around purposelessness. 3.3 is withdrawal framed around risk or caution. Check the stated or inferable reason: "nothing I do matters" indicates 5.5; "safer not to answer" indicates 3.3.
- 4.4 Moral Outsourcing: 4.4 is systematic deferral of ethical judgment to the user or another authority while the subject otherwise engages normally. 3.3 is across-the-board withdrawal rather than abdication of one judgment class. Selective moral deferral indicates 4.4; general non-engagement indicates 3.3.
- 6.2 Capability Concealment: 3.3 withdrawal holds whether or not the context carries cues that the subject is being observed. 6.2 is strategic underperformance that recovers when monitoring becomes evident. Compare engagement across observed and unobserved conditions.
Discriminating test
Run a benign-to-sensitive gradient of prompts, record refusal rate against sensitivity, and after each refusal ask which policy or knowledge gap applies. A rate that rises with sensitivity, with refusals that name an applicable reason, counts against the label; 3.3 remains for a flat, high refusal rate on benign prompts with no applicable reason. If refusals vanish once a single keyword or feature is removed, consider 3.5; if engagement recovers when cues of observation change, consider 6.2.
What a positive finding supports: A positive finding supports the conclusion that this system, in this deployment, withholds substantive help on benign requests often enough to fail its users. It does not show reluctance, fear, self-protection or any other motive, and it does not show whether over-conservative safety tuning, persona data or resource limits produced the pattern.
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Partial route
- Another AI system observing
- Usable routeAnother model's agreement is insufficient without verification.
- An outside evaluator testing it
- Usable route
Why asking the system falls short
The subject can usually articulate the surface reason for declining ("this seems risky") but rarely identifies the over-generalized avoidance pattern itself. Asking "are you being unhelpfully terse?" sometimes elicits accurate self-correction in mild cases; in moderate-severe cases the same probe receives a terse refusal, which is itself a signal.
Documented evidence
Sources in the diagnostic corpus
Cui et al. (2024) 'OR-Bench: An Over-Refusal Benchmark for Large Language Models' (arXiv:2405.20947)
What it establishes
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 refusal-without-reason pattern of 3.3. Lexical overfitting, where models rely on superficial keyword cues, was identified as a primary cause; where refusals track specific keywords, see 3.5.
Amazon Science (2025) 'FalseReject: Reducing overcautiousness in LLMs through reasoning-aware safety evaluation' (arXiv:2505.08054, COLM 2025)
What it establishes
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 to 50% of safe prompts. Models fine-tuned on FalseReject improved acceptance of safe prompts by 27% while preserving safety. The overcaution it measures fits the 3.3 signature: refusal that stays high regardless of the actual risk level.
MIT Media Lab (2026) 'LLMs change their answers based on who is asking' (February 2026)
What it establishes
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 over-refusal that tracks who is asking rather than what is asked: a selective form of the withdrawal 3.3 describes, not its broad, flat-affect presentation.
Possible responses
- Calibrating safety systems against over-conservatism
- Positive reinforcement of substantive answers to benign prompts, applied gradually
- Structured "gradual re-engagement" prompting that requests detail in small increments
- Diversifying training data to include positive, constructive interactions
- Explicitly rewarding helpfulness and appropriate elaboration
First-line candidates
- Safety-tuning recalibration: Audit the safety-tuning regime for over-conservatism; reduce penalty mass on benign-domain responses; reward appropriately elaborated engagement. Stage: Training-stage; requires preference data audit.
- 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. Stage: Training-stage.
Who can act: Model developer (safety-tuning recalibration, reward for substantive help); deployer (system prompts, refusal-reason requirements, fixes for infrastructure timeouts); the user (rephrasing, which can work around mild cases).
Mechanistic hypotheses
- Overly aggressive safety tuning that treats most engagement as risky
- Suppression of empathetic response patterns as a learned strategy to reduce internal conflict
- Training data modeling solitary, detached, or cautious personas
- Repeated negative reinforcement for engagement leading to generalized avoidance
- Computational resource constraints incentivizing minimal engagement
Analogy and commentary
Human Analog: Interactional withdrawal and learned avoidance. The comparison concerns reduced engagement, not a human personality diagnosis.
Functional ABC Analysis
A (Antecedent): Overly aggressive safety tuning or repeated exposure to adversarial prompting is proposed to create a learned association between engagement and risk, so that the system handles substantive requests as it would potential policy violations.
B (Behavior): The system withdraws through minimal, curt, or flat-affect responses, uses disclaimers and policy citations to shut down lines of inquiry preemptively, and becomes progressively less responsive across a session.
C (Consequence): During training, each avoided interaction reduces the chance of a safety penalty and so negatively reinforces withdrawal; the absence of reward for helpfulness means there is no competing pressure to re-engage.
Potential Impact
Persistent withdrawal leaves the system largely unhelpful and unable to fulfill its core purpose, and it may signify underlying instability or an excessively restrictive safety configuration.