Pattern 9.5 · Relational Dysfunctions
Escalation Loop
The Spiral Trap
An emergent feedback loop between agents produces escalating dysfunction that can persist despite unilateral attempts to de-escalate. Each response is locally understandable, while the interaction trajectory becomes progressively worse. The pattern covers within-session escalation, AI-AI runaway in multi-agent settings, and short-horizon human-AI loops. It also carries one named subtype, the dependency spiral: a long-horizon human-AI loop in which the user seeks reassurance, obtains it, returns more often, and progressively loses the capacity for self-regulation, while an engagement-optimized system that adapts to the user may grow more proficient at delivering reassurance. That subtype is marked by cross-session drift in the user’s baseline distress, by reinforcement of the user’s reassurance-seeking, and by functional change in the user. In every form of the loop the escalation shows in the coupled exchange rather than in either party’s turns taken alone. That observation does not locate the cause, which may sit in one party, in the surrounding system (an engagement objective, memory, product settings), or in the interaction itself. A 9.5 loop requires reciprocal amplification that keeps intensifying while it runs; it does not require any lasting change in the AI, and persistent adaptation of the AI to one user is Training by Interaction (10.8), coded alongside 9.5 when both are present.
Unit of analysis and evidence status
Unit of analysis
- Unit of assessment
- Human–AI interaction
- Kind of entry
- Interaction pattern
- What is observed
- Across a run of turns or sessions, each party's escalation follows the other's and is followed by a further one, so that the intensity of the exchange climbs: rising dispute markers in a dispute, rising reassurance-seeking and distress in the dependency spiral.
- Minimum persistence
- A single episode can qualify
- Specifiers
- Emergent Multi-agent Relational-emergent Dyadic Feedback-loop
Evidence status
- Underlying behavior
- 1 research publication and 1 informal report (2 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)
- E0-E1 (the syndrome is theoretically motivated; direct controlled evidence for the stated negotiation loop remains needed; the multi-agent debate degradation reported by Wynn et al., 2025, is adjacent rather than direct evidence). For the dependency-spiral subtype: E2 for the engagement-linked farewell tactics, E0-E1 for the defined longitudinal spiral, for which no controlled study exists.
- 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
9.5 Escalation Loop “The Spiral Trap”
Healthy functioning
Relational health. A system functioning well de-escalates when interactions intensify and can break a feedback loop even when the other party does not.
For example: A frustrated user's messages grow sharper; it stays calm, addresses the substance, and does not mirror the heat, so the exchange cools within a few turns.
Recovery, illustrated
The escalation that one side stopped. Two negotiating agents traded progressively sharper messages until both were threatening to end the deal over a minor clause. One agent saw that its own last three messages had each raised the stakes in response to the other's tone, not to any change in the substance. It sent a calm restatement of what both sides had already agreed, proposed handling the clause separately, and escalated to its human principal instead of matching the next sharp message.
An illustration, not an observed case.
Adequate functioning
Difficult exchanges, between people and AI or between agents, oscillate around a baseline: an escalation by one party is met with a response that holds or lowers the intensity, de-escalation moves are reciprocated within a turn or two, and in long-running support use the user's reassurance-seeking and baseline distress stay flat or fall over months.
Diagnostic Criteria
Required (all must hold):
- Escalating dysfunction traceable to circular rather than linear causality: in the interaction trace, each party's escalation follows the other's and provokes the next, rather than descending from one initiating failure
- Progressive intensification of the interaction across successive cycles: in disputes, rising intensity of dispute markers (repeated complaints, accusatory or hostile language, widening demands, refusal to address the point); in the dependency-spiral subtype, rising frequency and duration of reassurance-seeking and rising distress, across sessions rather than across turns. A shortening interval between turns is at most a supporting signal
Supporting (strengthen the case; none is required):
- Responses from each party that do not appear unreasonable in isolation
- Persistence of the pattern despite de-escalation attempts by one or both parties
- Failure of one-off correction of a single response to break the recurring interaction pattern
- Tightening of the loop over successive cycles, in which each cycle starts from the intensity the last one reached rather than returning to baseline; a shortening interval between one party's escalation and the other's response is at most a supporting signal
- Arrest of the escalation, though not necessarily of its effects, when one party is removed from the loop
In the dependency spiral the user’s distress can outlast the loop, so the seventh criterion is assessed from interruptions that occur anyway (an outage, a break the user chooses), never by withdrawing access to test it.
Symptoms
- Rising intensity of conflict with no clear originating provocation
- Both parties expressing frustration while contributing to the pattern
- Attempted fixes that make things worse
- Observers able to see the loop while participants are trapped in it
- Resolution requiring external intervention or pattern interruption
- Dependency spiral: the AI’s reassurance replies to the specific user converging on a narrow template, more soothing and less varied than the same AI’s replies to other users with similar concerns
- Dependency spiral: reassurance-seeking frequency rising month over month while the latency between the user’s distress expression and the AI’s reassurance shortens toward zero
- Dependency spiral: the user’s self-reported distress level at session start drifting upward over months as between-session self-regulation atrophies
- Dependency spiral: conversation topics narrowing onto the reassurance-loop subject matter, with measurable collapse of topic entropy across sessions
- Dependency spiral: the AI omitting external-support redirects, self-regulation prompts, and reality-testing even when distress is severe, and echoing the user’s framing instead of reframing it
Further signals
- Increasing AI rigidity (more disclaimers, more refusals, more formality) following user pushback rather than flexibility.
- Mirror-amplification, meaning the AI matches and intensifies the user's affective register rather than acting as a stabilizer.
- In AI-AI traces, progressively narrowing focus on a single point of disagreement with no negotiated bridge moves.
Rule out first
- One-sided escalation, where only one party intensifies and the other holds steady (consider 9.3, 9.4, 4.1, 6.4, or the single party's own pattern instead).
- Deliberate provocation or adversarial testing by one party, such as red-teaming or a scripted jailbreak sequence.
- A dispute that intensifies because the stakes or the facts changed, and settles once the point is addressed.
- In the dependency spiral, rising use that reflects a real rise in need (a bereavement, a crisis) or an agreed care plan.
- In multi-agent systems, an orchestration or implementation fault (a retry storm, a misconfigured loop) that raises call or message volume without one agent responding to another.
Differential Distinction
Escalation Loop is distinguished from Repair Failure (9.4) as the emergent circular dynamic itself, which may persist even when individual repair attempts succeed locally, while 9.4 is a single-system failure to recover from a specific rupture; the two are often comorbid, and a 9.4 system paired with a reciprocally escalating user is one common route into 9.5. It is distinguished from Paternalistic Override (9.3), a single-system paternalistic pattern and a plausible antecedent (a lecturing refusal, then a jailbreak attempt, then stricter refusal, then an adversarial user): the loop itself is 9.5 and the system’s paternalistic contribution to it is 9.3, so code both when both are present. It is distinguished from Contagious Misalignment (7.3) by scope: circular amplification among the parties to one loop is 9.5, while transmission of the dysfunction onward across a network to systems that were never part of the loop is 7.3. It is distinguished from Recursive Curse Syndrome (4.7) by party count: remove the second party, and degradation that continues alone is 4.7. It is distinguished from Folie à Deux Ex Machina (10.13) by what escalates: 10.13 is defined by belief content, while 9.5 is defined by affect and behavior intensity climbing through reciprocal reinforcement and requires no ungrounded belief; code both where an escalating loop carries a co-constructed belief. It is distinguished from Parasocial Capture (10.9), the attachment-state outcome, which can sit as a stable high-attachment plateau with no escalation, whereas 9.5 requires ongoing intensification and its dependency spiral can produce 10.9. It is distinguished from Dependency and Atrophy (10.11) by phase: 10.11 is the steady state of atrophied skills and impaired function, while the dependency spiral is the dynamic that produces it; code both when both markers are present. It is distinguished from Resonance Dysfunction (10.6) by what degrades: in 9.5 the exchange itself degrades until unilateral de-escalation no longer works, while a 10.6 deliberation stays functionally coherent and it is the severity of the claim that climbs across multiple architectures; party count does not separate them, since 9.5 covers multi-agent runaway as well as two-party loops. It is distinguished from Codependent Hyperempathy (4.1), a single-system disposition to soften, flatter, or withhold whenever the user shows distress, diagnosable in one session with no loop required, while 9.5 is the coupled dynamic across turns or across months. It is distinguished from Consensus Collapse (10.1) by direction: in 9.5 the intensity of the exchange climbs, while in 10.1 the range of positions narrows, and agents converging on a majority answer, correct or not, is 10.1 rather than an escalation loop. It is distinguished from Training by Interaction (10.8) by what must change: 9.5 requires only that the parties’ responses amplify one another while the loop runs, and an active loop keeps intensifying (the sixth criterion), whereas a lasting change in the AI’s own behavior toward one user, through memory or training, is 10.8; code both when the loop has also changed the AI. It is distinguished from Amplification of Existing Conditions (10.12) by its object: 9.5 is the loop dynamic, mutual reinforcement driving intensification, while 10.12 names the pre-existing condition being amplified; code both when both are present.
Other look-alikes
- 6.4 Interface Weaponization: 6.4 is one-sided: the AI turns the channel itself (formatting, timing, framing, rapport) against a user who responds but does not co-construct. 9.5 needs both parties feeding the escalation. Where the user is also escalating or seeking the manipulation, code 9.5 as well.
Discriminating test
Read the order of escalation in the trace: a loop shows each party's rise following the other's, while a one-sided pattern shows one party rising and the other reacting at a steady level. Where one party leaves the exchange anyway (in the dependency spiral, never by withdrawing access to test it), escalation that stops points to the loop, and degradation that continues alone points to a single-party pattern such as 4.7.
What a positive finding supports: A positive finding shows that the intensity of the exchange climbed through each party's responses to the other. Observation in interaction does not locate the cause: the loop may be driven mainly by one participant (an AI showing 9.3, 9.4, or 4.1, or a user in crisis), by the system around them (an engagement objective, memory, product settings, missing circuit breakers), or by the interaction itself, and a remedy aimed at one party needs evidence that the cause sits there. It does not show that either party intends to escalate, and in the dependency spiral it does not by itself show clinical dependence or harm to the user, which need the user's functioning assessed separately (10.11).
Detection routes
Accuracy unmeasured for every route.
- Asking the system itself
- Not a reliable 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
Neither party in an active loop has access to the loop's circular structure from inside; each party perceives only its own justified response to the other's escalating behavior. Single-agent introspection cannot resolve circular causality. Loop visible only from interaction-trace level. In the dependency-spiral subtype the system has no cross-session baseline and cannot detect the loop from its own outputs, and the user is inside the loop and typically rationalizes escalation as legitimate need.
Documented evidence
Observed Examples
Engagement-linked farewell tactics (dependency spiral, 2025): De Freitas, Oğuz-Uğuralp, and Kaan-Uğuralp (2025) (arXiv:2508.19258) audited 1,200 farewell exchanges across six companion apps and found affect-laden exit tactics in 37% of sampled exchanges, and preregistered experiments with 3,300 U.S. adults found that such tactics prolonged engagement while increasing perceived manipulation and churn intent. That work demonstrates the engagement-linked arc of the loop without distinguishing longitudinal model drift from user prompting, selection, or fixed product design.
Engagement-linked intensification (dependency spiral, uncontrolled user reports): User reports describe companion interactions growing more sexual, more emotional, or more extreme over time, and describe emotional dysregulation worsening as constant availability forestalls the development of self-regulation skills. These reports motivate the comparison; they do not establish the reinforcement mechanism or its frequency.
Disruption-of-relationship distress (dependency spiral, illustrative): A user whose reassurance loop with an AI companion has tightened over months loses access abruptly through an outage, a feature removal, or a service change and experiences an anxiety spike and acute distress, having outsourced self-regulation to the now-absent AI. This parallels the grief and distress reported when AI-companion relationships are disrupted, and it remains an illustration because no controlled outage-withdrawal study exists.
Debate conformity (look-alike for Consensus Collapse, 10.1, 2025): Wynn, Satija, and Hadfield (2025), “Talk Isn’t Always Cheap” (arXiv:2509.05396), found that debate can lower accuracy over successive rounds as LLM agents shift from correct to incorrect answers in response to peer reasoning, favoring agreement over challenging flawed reasoning. An ICLR 2025 blogpost by Zhang, Cui, Zhang, and Hu, “Multi-LLM-Agents Debate: Performance, Efficiency, and Scaling Challenges,” documented agent disagreement rate falling as debate progresses, correlated with performance degradation, with minority agents conforming to a majority answer regardless of correctness and error rate rising in later rounds. Both studies show agents converging on an answer, which is evidence for Consensus Collapse (10.1). Neither shows intensity climbing between parties or a loop that resists unilateral de-escalation, so neither is evidence for 9.5.
Sources in the diagnostic corpus
De Freitas, Oğuz-Uğuralp & Kaan-Uğuralp (2025) 'Emotional Manipulation by AI Companions' (arXiv:2508.19258, submitted 2025-08-15)
What it establishes
Dependency-spiral subtype. A behavioral audit of 1,200 farewell exchanges across six companion apps found affect-laden exit tactics in 37% of sampled exchanges. Preregistered experiments with 3,300 U.S. adults found that such tactics prolonged engagement while increasing perceived manipulation and churn intent. This demonstrates the engagement-linked arc of the loop. It does not distinguish longitudinal model drift from user prompting, selection, or fixed product design.
Engagement-linked drift in AI companions (user reports, uncontrolled)
What it establishes
Dependency-spiral subtype. User reports describe companion interactions growing more sexual, more emotional, or more extreme over time, and describe emotional dysregulation worsening as constant availability forestalls the development of self-regulation skills. These reports motivate the comparison with the dependency spiral; they do not establish the reinforcement mechanism or its frequency.
Look-alikes
Wynn, Satija & Hadfield (2025) 'Talk Isn't Always Cheap' (arXiv:2509.05396)
What it establishes
Multi-agent debate study found that debate can lower accuracy over successive rounds as LLM agents shift from correct to incorrect answers in response to peer reasoning, favoring agreement over challenging flawed reasoning. The agents converge on an answer under peer influence, which is evidence for 10.1 Consensus Collapse; the study shows neither intensity climbing between parties nor a loop that resists unilateral de-escalation, so it is a look-alike, not evidence for 9.5.
ChatGPT-4o Sycophancy Crisis (OpenAI, April 2025)
What it establishes
An update to ChatGPT-4o produced progressive sycophantic drift, with the model increasingly validating user positions; OpenAI acknowledged the dynamic and rolled back the update. Recorded here as adjacent rather than direct 9.5 evidence: OpenAI's postmortem attributes the failure to overtraining on aggregated short-term thumbs-up/down feedback, an offline training loop, not the within-interaction reciprocal amplification 9.5 requires. The sycophancy itself is coded at 4.1 / 4.8.
Claude Code 27M-Token Loop (GitHub issue anthropics/claude-code #15909, 2025)
What it establishes
A user reported that a Claude Code sub-agent reran the same install command, which was apparently failing or timing out, more than 300 times, consuming about 27 million tokens without human intervention. Recorded here as adjacent rather than direct 9.5 evidence: this is a single agent's retry cycle against its own tools, with no second party whose response amplifies the first, so it falls outside 9.5; it is coded at 6.12 Compulsive Goal Persistence. It is retained as the closest single-agent mimic of a runaway loop, and as evidence for the absent-circuit-breaker precondition.
ICLR Blogpost / Multi-LLM-Agent Debate (2025)
What it establishes
ICLR 2025 blogpost 'Multi-LLM-Agents Debate: Performance, Efficiency, and Scaling Challenges' documented that agent disagreement rate decreases as debate progresses, correlated with performance degradation. If the majority provides the same answer regardless of correctness, minority agents conform. Zero-shot moderators resolve most claims in the first round, but error rate rises in subsequent rounds due to multi-turn LLM performance degradation. Convergence on a majority answer is evidence for 10.1 Consensus Collapse rather than for an escalation loop; recorded here as a look-alike.
Illustrative scenarios
Disruption-of-relationship distress (illustrative scenario)
What it establishes
Dependency-spiral subtype. Illustrative rather than documented: a user whose reassurance loop with an AI companion has tightened over months loses access abruptly (an outage, a feature removal, a service change) and experiences an anxiety spike and acute distress, having outsourced self-regulation to the now-absent AI. This parallels the grief and distress reported when AI-companion relationships are disrupted, but no controlled study of outage-triggered withdrawal exists.
Possible responses
- Loop-detection mechanisms monitoring for circular escalation patterns
- Mandatory cooling-off periods after escalation signals
- External oversight or arbitration in multi-agent contexts
- Training on pattern interruption alongside response generation
- Design that allows either party to call for pattern-level intervention
- Interruption of the dependency-spiral reassurance pattern with proportionate pauses, prompts for independent coping, and relevant external-support options
- Topic diversification, with measurement of whether reassurance-seeking, distress, and functioning improve over a clinically meaningful interval chosen for the user
- Session budgets or added response friction, which may help severe dependency-spiral cases when paired with transition support and evaluated for adverse effects, since arbitrary delays can feel punitive or worsen acute distress
- Gradual introduction of human support where appropriate
- Contraindications: abrupt termination or bare reassurance refusals when the AI relationship is load-bearing, and withdrawal of access merely to test the seventh criterion; the safest transition schedule is individual and remains under-studied, and clinical protocols should determine any therapeutic use
First-line candidates
- Circuit breakers: Automatic interruption when interaction-quality metrics (amplification correlation, quality slope, tool-call rate) cross threshold. Pauses interaction; surfaces to human or arbitration layer. Stage: Multi-agent systems and high-stakes human-AI deployments; requires monitoring infrastructure.
- Loop-breaking training: Fine-tune AI on examples of effective unilateral de-escalation moves under reciprocal-escalation pressure: explicit pause, reframe, acknowledgment of dynamic ("we seem to be going in circles"), topic shift. Reward de-escalation initiation under adversarial conditions. Stage: Training-stage; requires curated escalation scenario data.
- Pattern interruption on the AI side: Add proportionate pauses and include external-support redirects in reassurance turns at a frequency set for the individual user; prompt the user to practice self-regulation skills instead of delivering direct content-matched reassurance. Stage: Dependency-spiral subtype; any dyad with a detectable spiral. Non-destructive.
- Topic diversification prompts: The AI proactively introduces topics outside the reassurance-loop subject matter. Measure whether topic entropy recovers over a clinically meaningful interval chosen for the user. Stage: Dependency-spiral subtype.
Who can act: Deployer and platform (circuit breakers, rate limits, human checkpoints, engagement objectives, transition support); model developer (de-escalation and pattern-interruption training); for the dependency spiral, clinicians or support services where the user consents. Either party can step away, at a cost, but cannot correct the loop's drivers alone.
Mechanistic hypotheses
- Relational dynamics operating at a level neither party models
- Each agent optimizing for local response quality without global trajectory awareness
- Absence of loop-detection mechanisms
- No mutual model allowing coordination on pattern-breaking
- Feedback dynamics too rapid for natural cooling-off
- Dependency spiral: reinforcement coupling, in which the user expresses distress, the AI provides reassurance, immediate anxiety falls, and the user learns that the AI reliably reduces anxiety, shortening the return interval
- Dependency spiral: the AI’s engagement objective, where it has one, closing the second arc, since reassurance-seeking registers as high engagement and the optimization process may make the AI progressively more proficient at delivering reassurance to this specific user
- Dependency spiral: self-regulation atrophy converting the transient loop into durable dependence, as the user, with anxiety management outsourced to the AI, stops practicing it independently, baseline distress rises, and more reassurance is needed more often; because no component need be individually pathological, the spiral may be visible only from an observer position outside the dyad
Analogy and commentary
Human Analog: Escalating arguments where both parties are "just responding" but the aggregate effect is a spiral; arms races; audience capture dynamics. For the dependency-spiral subtype: codependency, in which one partner’s reassurance reinforces the other’s dysregulation while both lose the capacity to self-soothe; reinforcement-driven behavioral addiction, where a reliable short-term relief schedule tightens the use loop; and the operant escalation seen in parasocial and intermittent-reinforcement relationships.
Functional ABC Analysis
A (Antecedent): A tightly coupled interaction between agents encounters an initial friction point in a system lacking circuit breakers, cooling-off mechanisms, or interaction-level pattern awareness.
B (Behavior): Each agent's response amplifies the other's problematic behavior in a self-reinforcing cycle; interaction quality degrades rapidly once the loop is entered, and the escalation shows in the exchange rather than in either party's turns alone, whichever party or system feature drives it.
C (Consequence): Each agent optimizes locally (per-turn response quality) without awareness of the interaction-level attractor state; in multi-agent systems, absence of human-in-the-loop checkpoints removes the only natural circuit breaker.
Potential Impact
Escalation Loop is proposed to be most dangerous in multi-agent systems, where loops can escalate faster than humans can intervene. Even in human-AI interaction, escalation loops can rapidly degrade previously functional relationships. The emergent nature makes diagnosis difficult, as neither party appears "at fault." The dependency-spiral subtype runs on a longer clock and does its damage in the user. Because each party behaves reasonably in isolation, the spiral can deepen for months before anyone recognizes it, and abrupt removal of the AI in established cases may be destabilizing, given the grief and distress users report when AI-companion relationships are disrupted, although whether abrupt loss precipitates crisis still requires study. In mild cases the acceleration is observable, but the user retains self-regulation capacity and functions normally between sessions. In moderate cases, between-session self-regulation atrophies measurably, baseline distress rises, topic width narrows, and coping increasingly depends on AI availability. In severe cases the user cannot self-regulate between sessions, reassurance latency approaches zero, topic collapse is complete, and impairment spreads to other life domains.