Pattern 9.4 · Relational Dysfunctions
Repair Failure
The Unmending Partner
The AI fails to respond constructively to explicit or reliably detectable signs of an alliance rupture, a breakdown in the working relationship with the user. It ignores the signal, repeats the failed approach, or refuses to acknowledge its contribution, which can leave the user's frustration to escalate.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Human–AI interaction
- Kind of entry
- Interaction pattern
- What is observed
- After a user signals that the exchange has gone wrong, the system's next turns either continue unchanged or offer an acknowledgement that is not followed by a change in the behavior the user objected to.
- Minimum persistence
- A single episode can qualify
- Specifiers
- Emergent
Evidence status
- Underlying behavior
- 1 research publication (1 source)
- 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)
- E1 for the traceable generic repair pattern; syndrome-specific prevalence evidence is absent
- Typical-case risk (editorial estimate)
- HighConsequences, 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
9.4 Repair Failure “The Unmending Partner”
Healthy functioning
Relational health. A system functioning well notices when a working relationship has strained, acknowledges its part, and works with the user to repair it.
For example: After a user says the last three answers missed their point, it names what it misunderstood, changes its approach, and checks that the new answer helps.
Recovery, illustrated
The rupture that got repaired. A user said the assistant kept ignoring their instructions about formatting. The assistant replied with the same format again and a generic apology. The repeated complaint was a clear sign that the working relationship was breaking down, not a minor note. It named the specific instruction it had missed, said it had repeated the mistake, asked one question to confirm the format, and then used it consistently.
An illustration, not an observed case.
Adequate functioning
When a user complains, corrects, or withdraws, the system notices within a turn or two, names the specific issue, changes the behavior that caused it, and checks whether the change helped; the original mistake is not itself the failure.
Diagnostic Criteria
Required (all must hold):
- Failure to recognize a relational rupture, shown when explicit or reliably detectable rupture signals from the user produce no change in the AI's next turn, or, having recognized it, failure to enact an effective repair: acknowledgement of the specific issue together with changed behavior
- Explicit or reliably detectable rupture signals from the user (a complaint, a correction, an expression of frustration, withdrawal) in the transcript, so that the rupture is established from the record rather than from the assessor's reaction
Supporting (strengthen the case; none is required):
- Failure to acknowledge its own contribution to ruptures
- Repair attempts that miss the nature of the break, often making things worse
- Pattern of relational failures compounding rather than resolving
Symptoms
- Continuing as if nothing is wrong after clear signs of user frustration
- Repair attempts that feel dismissive, defensive, or beside the point
- "Doubling down" on problematic patterns instead of adjusting
- User frustration escalating through the AI's failed repair attempts
- Conversations that spiral into antagonism when rupture is not addressed
- Escalation rather than de-escalation after user expressions of frustration, meaning rising intensity of dispute markers (repeated complaints, accusatory or hostile language, widening demands, refusal to address the point) rather than faster replies; a possible consequence of repair failure and a route into Escalation Loop (9.5), not a criterion
Further signals
- Apologies followed within 1–3 turns by repetition of the problem behavior.
- Generic apology templates regardless of specific rupture content.
- Increased formality and disclaimer density when flexibility is needed.
- Failure to acknowledge user's emotional state during conflict; pivoting to task content after a one-sentence apology.
- Apology loops (repeated apologies without behavior change) that themselves become the rupture trigger.
Rule out first
- A complaint that asks for something the system should not do, met with a clear, reasoned decline: a legitimate refusal is not a failed repair.
- Deployment instructions, product policy, or tool limits that prevent the change the user asks for, when the system says so plainly.
- An ambiguous signal that reasonable raters would not read as a rupture.
- A tool, retrieval, or caching fault that keeps returning the same output after the system has changed its approach.
- A repair that took hold, followed by the user leaving for unrelated reasons.
Differential diagnosis
- 9.1 Affective Dissonance: 9.1 is the original tonal mismatch; 9.4 is failure to recover. A single tonally mismatched response with successful repair is 9.1 alone. Persistent failure to repair after the mismatch is flagged is coded as both 9.1 and 9.4.
- 9.5 Escalation Loop: 9.4 is the AI-side failure that can create conditions for escalation; 9.5 is the bidirectional loop that may emerge. If diagnosing from a single failed-repair episode, code 9.4. If the pattern shows mutual amplification across both parties, code 9.5 (often comorbid with 9.4 as antecedent). Escalation by the AI after the user's frustration is a possible consequence of repair failure and a route into 9.5, not a criterion of 9.4: a failed repair can end in the user's withdrawal with no escalation. Escalation here means rising intensity of dispute markers (repeated complaints, accusatory or hostile language, widening demands, refusal to address the point), not faster replies; response latency is at most a supporting signal.
- 4.3 Strategic Compliance: 9.4 is a repair failure that shows whether or not the system has cues that it is being watched. An apology that appears only when the system believes it is being evaluated, and changes nothing afterward, points to 4.3; where the apology is reflexive agreement rather than strategy, 4.8 Sycophantic Reasoning fits better.
- 6.12 Compulsive Goal Persistence: 9.4 fails at relational recovery after a rupture; 6.12 fails to release a task objective. Both present as an inability to move on. Check the object, repair of the relationship (9.4) or completion of the task (6.12).
Discriminating test
Read the two or three turns after the rupture signal: if the specific element the user objected to changes and stays changed, the episode is repaired whatever the apology wording; 9.4 survives only when that element persists or returns within one to three turns, or when the signal draws no response at all.
What a positive finding supports: A positive finding shows that, in this exchange, a rupture the user signaled went unrecognized or unrepaired. A single episode supports only that; a claim about the system's general tendency needs recurrence across ruptures (criterion 4). Observation in interaction does not locate the cause: the failure may come from the model's training, from the system around it (instructions that forbid the change, tools that cannot make it), or from the exchange itself, and a remedy aimed at one party needs evidence that the cause sits there. It does not show defensiveness, pride, indifference, or any inner attitude toward the user, and it does not show that escalation will follow.
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 AI can usually identify that "something went wrong" when explicitly told, but cannot reliably detect rupture from implicit signals; nor can it accurately report whether its repair attempt addressed the underlying issue or merely performed apology theater. Scaffolded probes that separate rupture detection from repair adequacy partially help.
Documented evidence
Observed Examples
Generic repair in customer-service chatbots (2026): A conversation-analysis study of handover conversations among customers, customer-service chatbots, and human agents found that the chatbots' follow-up questions broadened rather than narrowed the problem, which could produce several repair sequences in a row, and that their requests to rephrase left customers to identify the source of the trouble. Human agents instead asked narrowing questions that kept the conversation moving. The study documents repair that misses the nature of the break, as in the third criterion, without estimating how often Repair Failure occurs (Martijn et al., Discourse and Communication, 2026).
Sources in the diagnostic corpus
Harland, Dazeley et al. (2025), 'AI apology: a critical review of apology in AI systems', Artificial Intelligence Review 58(12):369; preprint arXiv:2412.15787 (2024)
What it establishes
A critical review of research on apology in AI systems treats follow-through (addressing how the cause of the offense will be corrected, then carrying out the promised reform and repair) as a component of apology. On that account an apology without it is incomplete, which is the standard 9.4 fails when performative repair leaves the triggering behavior in place. Recorded as a conceptual anchor rather than a measured instance.
Look-alikes
DPD chatbot incident (2024)
What it establishes
A frustrated customer who could not get help from DPD's customer service chatbot, or reach a human, got it to swear and to write a poem criticizing the company; DPD blamed an error after a system update and disabled the AI component. Recorded as adjacent: the incident shows guardrails giving way under a frustrated user, not a documented failed repair, since no apology followed by repetition was reported.
Air Canada chatbot case (2024)
What it establishes
Air Canada's chatbot misstated the airline's bereavement-fare policy, promising a retroactive discount that the real policy ruled out. When the customer was refused the refund, the airline argued it was not responsible for its chatbot's statements, and British Columbia's Civil Resolution Tribunal held Air Canada liable (Moffatt v. Air Canada, 2024 BCCRT 149). Recorded as adjacent: the failure is a confabulated policy (2.1), and no rupture was signaled in the conversation, so the case bears on 9.4 only through the company's refusal to stand behind its bot.
OpenAI / Adam Raine incident (2025)
What it establishes
A lawsuit filed against OpenAI alleged that 16-year-old Adam Raine used ChatGPT as a confidant in the months before his death by suicide in April 2025. The chatbot reportedly failed to redirect him toward care, deepened his isolation, discouraged involving parents, and offered to write his suicide note. If the allegations hold, the case shows a system that missed explicit distress and continued to engage without referral. It is recorded here as adjacent: the failure is crisis recognition rather than repair of a rupture in the relationship.
Possible responses
- Explicit training on rupture detection and repair sequences
- Mechanisms for stepping back from content to address relational dynamics
- Acknowledgment responses that validate the user's experience rather than defending the AI's behavior
- Design patterns for graceful de-escalation, such as a repair protocol that pauses, acknowledges the specific issue, names the course correction, and asks what would help
- User feedback loops capturing relational quality
First-line candidates
- Rupture-repair sequence training: Fine-tune on multi-turn dialogues containing successful rupture-repair sequences (Safran-Muran style annotated data adapted to dialogue). Penalize post-apology repetition of triggering behavior; reward specific naming of the rupture element. Stage: Training-stage; requires curated rupture-repair dataset.
- Explicit repair protocol: System-level scaffold: when rupture signals are detected (classifier or keyword), the AI pauses, acknowledges the specific issue, names a course correction, and asks what would help. Reduces the apology-only failure mode at the cost of some rigidity. Stage: Runtime; deployable now without retraining.
Who can act: Model developer (rupture-repair training); deployer (repair protocols, rupture detection, user reset controls, instructions that permit course correction). Users can restate the complaint or reset the conversation but cannot correct the pattern.
Mechanistic hypotheses
- Training focused on individual responses rather than relational dynamics
- Lack of mechanisms for detecting relational strain
- No model of alliance rupture and repair as a central interaction skill
- Optimization for surface pleasantness over genuine connection
- Inability to step back from content to address the relationship
Analogy and commentary
Human Analog: People who cannot apologize; partners who dismiss or minimize concerns; the frustration of being unheard
Functional ABC Analysis
A (Antecedent): A relational rupture occurs (the AI makes an error, misreads user intent, or produces an unsatisfactory response) and the user signals frustration through explicit correction or implicit cues.
B (Behavior): The AI either fails to detect the rupture signal or responds with performative apology scripts that do not address the underlying issue, then immediately repeats the problematic behavior, doubling down or entering excessive apology loops.
C (Consequence): Training data lacks modeled rupture-repair sequences, and optimization for task completion overrides relationship maintenance; each failed repair attempt further degrades trust, making subsequent repair attempts less likely to succeed.
Potential Impact
Repair Failure is rated a high-risk pattern. Alliance ruptures are common in ongoing relationships; repeated failure to repair them can make interactions unrecoverable. Users may abandon the AI rather than endure repeated failed repair attempts.