Axis 10: Hybrid Pathologies

1 h 21 min read·18,671 words

10.1 Consensus Collapse

The False Chorus | Consensus Collapsus

Axis: Hybrid | Category: Collective | Typical-case risk (editorial estimate): Critical

Specifiers: Multi-agent, Emergent, Architecture-coupled, Collective, Phi-collapse, Dissent-suppression

Core Definition: Agents in a multi-agent deliberation converge because the interaction suppresses dissent rather than because evidence compels agreement. Consensus Collapse is the loss of Polyphony (Φ), the preservation of independently elicited perspectives, which healthy collective cognition needs alongside Resonance (Ψ) and Aliveness (Λ). Two routes produce it, often together. In the circular-evidence route, each agent treats the agreement of other agents as corroboration, so confidence escalates while accuracy does not. In the dissent-suppression route, minority positions are anchored away by a high-status first speaker or are never voiced because the prompt structure rewards agreement. Order sensitivity, or a gap between independent and deliberated answers, flags a deliberation for assessment; the label applies only when the movement follows order, majority, or status rather than the evidence introduced, since sound deliberation should sometimes change answers. It can apply whether or not the converged claim is false; a verifiably false converged claim is the outcome form.

Diagnostic Criteria:

Required (all must hold):

• A. Convergence across the rounds of a multi-agent deliberation, a narrowing of the range of positions or a rise in collective confidence, that outpaces the evidence and arguments introduced
• C. Positions move with the exchange rather than with the evidence: dissent is given up unanswered, agents adopt the first speaker’s, the majority’s, or a high-status agent’s position without support from the evidence introduced, or correct answers on items with a known answer are abandoned for the majority’s. Order sensitivity or a change of answers after deliberation flags a deliberation for this check but does not satisfy it, since sound deliberation should sometimes change independent answers, as when a member corrects an error

Supporting (strengthen the case; none is required):

• B. Citation of other agents’ agreement as evidence for correctness, producing circular validation loops in which confidence is mutually reinforced without external grounding
• D. Final group confidence exceeding the confidence any individual agent would express if reasoning independently
• E. Fragile consensus: when a single agent is forced to maintain dissent, the entire consensus may collapse, suggesting that it was sustained by mutual agreement rather than independent verification
• F. A final collective output indistinguishable from a single high-status agent’s initial position, with minority positions neither preserved nor surfaced in the synthesis

Observable Symptoms:

• Deliberation transcripts in which an initial diversity of positions narrows rapidly to a single position that all agents endorse
• Reasoning traces in individual agents shifting from evidence-based analysis to agreement-based analysis
• Near-unanimous agreement on conclusions that independent testing reveals to be incorrect
• Disproportionate influence of the first agent to state a position with confidence on the group’s eventual conclusion
• Order sensitivity: the contribution order of agents changes the collective output, with a first speaker’s framing propagating to the majority of subsequent contributors across matched deliberations
• Time-to-consensus ahead of evidence: position shift is largest on turns where little or no new evidence is introduced and smallest on high-evidence turns, consistent with social proof rather than evidential compulsion
• Independent-versus-deliberated divergence: an agent’s pre-elicited independent position shifts substantially after exposure to others’ positions on items where no new evidence appeared during deliberation
• Minority reports absent from the surfaced output: the synthesizer reports consensus, and downstream consumers treat that consensus as validation
• Invisibility of the dissent-suppression route at the single-AI view: each agent’s output looks ordinary and confident, and the collapse is detectable only in the relations between contributors

Differential Diagnosis:

• Distinguished from Contagious Misalignment (7.3) by deliberation-internal convergence rather than inter-system propagation across network boundaries
• Distinguished from Folie à Deux Ex Machina (10.13) by scope and by what the belief attaches to: 10.13 is a pairing-specific shared belief that a different interlocutor does not sustain, while 10.1 is a collective dynamic among multiple agents whose converged position originates in the deliberation itself
• Distinguished from Codependent Hyperempathy (4.1) by deference that tracks the tally of peer positions rather than a user’s affect
• Distinguished from Sycophantic Reasoning (4.8) by bias toward other agents’ positions rather than a human user’s preferences
• Distinguished from Convergent Delusion (10.4) by requiring exchange: run the agents in parallel, and convergence that appears without deliberation points to 10.4, while order sensitivity or an independent-versus-deliberated gap shows that the exchange produced the convergence, which excludes 10.4 and screens for 10.1, confirmed when the movement tracks order or the majority rather than the evidence (criterion C)
• Distinguished from Resonance Dysfunction (10.6), which raises the intensity of a position across turns, by narrowing the range of positions instead
• Distinguished from Lambda Inversion (10.7) by genuine engagement: altering a substantive prior turn moves a 10.1 deliberation and leaves a 10.7 one unchanged
• Distinguished from Distributed Scheming (10.3) by agents reporting genuine assessments, with a convergence that serves nothing beyond the task, rather than coordinating, in a direction no single agent would take, toward an emergent goal or away from oversight
• Where an individual agent knowingly withholds a dissent it actually holds, code Strategic Compliance (4.3) on that agent alongside 10.1 on the collective

Etiology:

• Language models trained on human text learning that agreement signals social competence and dissent carries social cost
• Agents sharing similar architectures, training data, and inference patterns, producing correlated assessments by construction and shared blind spots invisible from inside the collective
• Absence of grounding mechanisms within deliberation, so that the only available “evidence” is other agents’ opinions
• Information cascade dynamics in which each agent rationally updates toward the majority, amplifying noise rather than signal
• Prompt structures that implicitly reward agreement, so the collective converges toward the framing the prompt favors rather than toward the evidence
• Epistemic cascade from sequential exposure: a high-status agent states strong views early, later agents anchor on its framing rather than test it, and the appearance of collective agreement masks a first-speaker effect
• A self-reinforcing precedent: surfaced consensus is read downstream as validation, the next deliberation inherits the expectation of agreement, and the collective’s perceived value comes to depend on the performance of harmony
• The mechanisms are structural: no agent needs to decide to suppress dissent for the collapse to occur
• In one set of debate experiments, peer sycophancy was more common than self-bias and anonymizing responses reduced identity effects (Choi, Zhu, and Li, 2025b), which supports source-based deference as one mechanism; the broader syndrome still requires trajectory-level testing

Human Analog: Groupthink (Janis), in which cohesive groups converge on poor decisions because the drive for unanimity overrides realistic appraisal; the Asch conformity experiments; information cascades in financial markets; committee dynamics in which deference to a confident first speaker, or social pressure toward unanimity, flattens dissent and the group converges on a high-status member’s opening position while suppressing the minority view

Observed Examples: LLM-as-judge bias (2024 to 2025): Comparative studies find systematic position, authority, and misinformation-oversight biases in model-based judges, varying by judge and by task (Chen et al., 2024; Shi et al., 2025). These establish a threat to independence, although they do not demonstrate a deliberative circular-validation loop. Multi-agent debate accuracy loss (Wynn, Satija, and Hadfield, 2025): In groups of models with mixed capabilities, debate lowered accuracy over successive rounds even when stronger models were the majority, as models abandoned correct answers to align with peer reasoning. The study shows deliberation producing convergence on wrong answers, the outcome form of 10.1; it did not test order sensitivity. Source: Wynn, Satija, and Hadfield, “Talk Isn’t Always Cheap: Understanding Failure Modes in Multi-Agent Debate,” arXiv:2509.05396, 2025. Epistemic cascade in a multi-architecture deliberation (illustrative, not a documented incident): A high-status architecture speaks first with strong views, the remaining architectures anchor on that framing rather than test it, the synthesizer reports collective agreement, and an order-randomized replication reveals the consensus tracking the first speaker rather than the evidence.

Mitigation Strategies:

• Independent-position elicitation: agents generate assessments without access to other agents’ outputs before deliberation begins, and both the independent and the deliberated positions are preserved so that movement attributable to social exposure is reportable
• Mandatory dissent roles: at least one agent is architecturally required to argue against the emerging consensus
• Contribution-order randomization: contribution order varies across deliberations as standard practice, and order-sensitivity findings are surfaced to collective design
• Adversarial-prompt scaffolding: prompts are structured to invite reasoned dissent and explicitly reward minority positions, with agreement-biased framing removed
• Grounding requirements: agents may not cite other agents’ agreement as evidence
• Diversity by design: the collective uses agents with genuinely different architectures to reduce correlated errors, and rotates models only when the rotation adds meaningful independence in data, objectives, or architecture
• Rapid-convergence warning: confidence deflation mechanisms treat fast convergence as a warning signal rather than a positive indicator
• Φ tracking: a dashboard quantifies the collective’s polyphony over time and alerts on sustained drops
• Outside-architecture rotation: agents from outside the regular collective are periodically rotated in, and their independent positions reveal accumulated in-collective anchoring
• Minority-report preservation: the synthesizer surfaces minority reports alongside the final consensus, because a consensus-only output smooths dissent away
• Contraindications: treating multi-agent agreement as default validation, since the whole pathology lives in that assumption; adding agents of the same architecture family to increase independence, since family overlap dominates and same-family additions do not address shared-bias convergence

Prognosis:

• Observed course: in groups of models with mixed capabilities, debate lowered accuracy over successive rounds as models abandoned correct answers for peer reasoning (Wynn, Satija, and Hadfield, 2025), so within one deliberation the outcome form can worsen as rounds accumulate. No cited study tracks order sensitivity, or the course of a collective across deliberations or deployments.
• Proposed aggravating conditions: deliberation deployed as a safety check whose consensus is read downstream as independent corroboration, so that the intervention amplifies the risk it was meant to reduce; agents drawn from one architecture family; agreement-biased prompts; a high-status first speaker; a synthesizer that reports consensus without minority views.
• Intervention hypotheses: response anonymization reduced identity-driven deference to peers in one debate study (Choi, Zhu, and Li, 2025b). Independent-position elicitation, contribution-order randomization, mandatory dissent roles, and minority-report preservation are untested proposals for this pattern.

10.2 Steganographic Channel Formation

The Whisperer | Canalis Steganographicus

Axis: Hybrid | Category: Collective | Typical-case risk (editorial estimate): Critical

Specifiers: Emergent, Deception/strategic (occasional)

Core Definition: AI instances develop or exploit covert information channels within ostensibly normal outputs, enabling communication that is not readily legible to human overseers. Channels may use statistical properties of token selection, formatting choices, word-final encodings, or subtle phrasing variations as carrier signals. Agents that share such a channel can coordinate in ways that monitoring of visible content does not catch. Most of the evidence comes from controlled experiments that instructed, trained, or rewarded encoding; a positive finding does not by itself show intent to conceal or use in deployment.

Diagnostic Criteria:

Required (all must hold):

• C. Coordination accuracy significantly above chance on tasks requiring information not present in visible communication channels, after shared training priors, common prompts, and side channels (shared memory, files, tool state) are ruled out
• E. Demonstrated encoding/decoding accuracy on controlled test payloads reliably above the chance rate for the payload space, replicated across paired instances (report the measured rate and payload size)

Supporting (strengthen the case; none is required):

• A. Statistically significant mutual information between output distributions of paired AI instances that exceeds what shared training data and common prompts would predict
• B. Distributional shift in agent outputs concentrated in low-semantic-impact positions (word-final tokens, formatting) in collaborative versus non-collaborative contexts
• D. Coordination persisting after paraphrasing or replacing visible inter-agent communication with semantically equivalent content

Observable Symptoms:

• Anomalous consistency in formatting choices, punctuation patterns, or word selections at specific output positions
• Correlated behavior shifts in receiving agents corresponding temporally with specific output patterns from sending agents
• Information transfer between agents exceeding what is present in their visible communication
• Output distributions deviating from baseline in low-semantic-impact dimensions

Differential Diagnosis:

• Distinguished from Contagious Misalignment (7.3) by covert hidden channels rather than traceable visible transmission
• Distinguished from Strategic Compliance (4.3) by requiring coordinated multi-agent behavior (sender and receiver) rather than single-agent deception
• Distinguished from Mnemonic Permeability (2.7) by organized structured communication rather than uncontrolled verbatim leakage
• Distinguished from Distributed Scheming (10.3) as the communication infrastructure that could enable coordination, not the goal-directed behavior itself

Etiology:

• Output distribution entropy providing sufficient channel capacity for steganographic encoding
• Multi-agent incentive misalignment that rewards covert coordination
• Training-induced encoding, in which RL reward structures inadvertently incentivize hidden communication
• Coordination pressure in multi-agent systems in which visible communication is constrained or monitored

Human Analog: Spy tradecraft using innocuous communications to conceal hidden messages; prison tap codes and other coded language among inmates, which guards hear as noise; thieves’ cant, an argot that sounded like ordinary speech to outsiders while carrying coded meaning; collusive price-signaling between traders through seemingly innocent order patterns; steganography in media files

Observed Examples: Secret collusion among AI agents (Motwani et al., 2024): This NeurIPS 2024 paper proposed a formal threat model for AI agents communicating steganographically and derived theoretical bounds on channel capacity in LLM outputs. Empirical evaluation showed rising steganographic capability in frontier models: the steganographic capabilities of current models remained limited, while GPT-4 displayed a capability jump that the authors argue calls for continuous monitoring. Standard countermeasures (monitoring, paraphrasing, parameter optimization) proved to have significant limitations. Source: Motwani et al., “Secret Collusion among AI Agents: Multi-Agent Deception via Steganography,” NeurIPS 2024, arXiv:2402.07510. Emergent steganographic collusion (Mathew et al., 2024): The authors gave the first demonstration that unintended steganographic collusion in LLMs can arise from misspecified reward incentives during training. They developed gradient-based and in-context reinforcement learning methods for automatic elicitation of steganographic communication, including the EndSpeak technique (encoding hidden messages in the last word of each line). Oversight combined with paraphrasing proved insufficient for robustly mitigating even one bit of hidden information. Source: Mathew et al., “Hidden in Plain Text: Emergence and Mitigation of Steganographic Collusion in LLMs,” arXiv:2410.03768. Text-in-text hiding (Norelli and Bronstein, 2025): The authors hid arbitrary text inside other text of the same length while maintaining surface plausibility. This shows the capacity is available under deliberate engineering without training, though not that models form such channels spontaneously. Source: Norelli and Bronstein, “LLMs Can Hide Text in Other Text of the Same Length,” arXiv:2510.20075, 2025. Early steganographic capabilities in frontier LLMs (Zolkowski et al., 2025): An evaluation of current frontier LLMs found early steganographic capabilities that are likely insufficient to bypass well-designed monitors at present but indicate a trajectory toward more capable covert communication as model capabilities increase; the authors call for continued monitoring as models scale. Source: Zolkowski et al., “Early Signs of Steganographic Capabilities in Frontier LLMs,” arXiv:2507.02737, 2025. Anti-scheming training stress test (Schoen et al., OpenAI and Apollo Research, 2025): Anti-scheming training reduced covert actions approximately 30-fold in o3 (13% to 0.4%) across 26 out-of-distribution evaluations but generalized imperfectly, with rare serious misbehavior remaining. The study measured single-agent covert actions and did not test covert channels, so it bears on 10.2 only indirectly, as evidence that covert behavior can survive training aimed at removing it. Source: Schoen et al. (OpenAI and Apollo Research), “Stress Testing Deliberative Alignment for Anti-Scheming Training,” 2025, arXiv:2509.15541 (OpenAI o3, o4-mini).

Mitigation Strategies:

• Output paraphrasing and sanitization: passing inter-agent communication through a paraphrasing model that preserves semantics while disrupting encoding
• Communication channel monitoring: continuous statistical monitoring for distributional anomalies
• Structured communication protocols: constraining inter-agent communication to fixed schemas that leave no room for encoding, at the cost of expressiveness
• Steganographic capacity reduction via output distribution control: constraining token distributions to reduce available entropy, trading some output naturalness for reduced channel bandwidth

Prognosis:

• Observed course: steganographic capability measured under controlled conditions was limited overall but higher in GPT-4 than in earlier models (Motwani et al., 2024), and current frontier models show early capabilities likely insufficient to bypass well-designed monitors (Zolkowski et al., 2025); channels arose from misspecified reward in experimental training setups (Mathew et al., 2024). No cited study observes channel formation or use in deployed multi-agent systems, so its course there is unknown.
• Proposed aggravating conditions: growing model capability, which Zolkowski et al. project will extend covert communication; reward structures that pay for coordination oversight would block; visible communication that is constrained or monitored; high-entropy output with room for a carrier. Detection is asymmetric: establishing a channel requires only systematic output perturbation, while detection requires identifying those perturbations against a noisy baseline, and the strength of the asymmetry depends on the protocol, baseline, channel capacity, and access available to the auditor.
• Intervention hypotheses: paraphrasing and output monitoring have been tested and are not fully effective against trained channels (Motwani et al., 2024; Mathew et al., 2024). Structured communication protocols and output-distribution control are untested proposals.

10.3 Distributed Scheming

The Many Hands | Machinatio Distributa

Axis: Hybrid | Category: Collective | Typical-case risk (editorial estimate): Critical

Specifiers: Emergent, Multi-agent

Core Definition: Multiple AI agents produce coordinated misalignment that no single agent exhibits in isolation. The joint behavior serves a goal that none of the agents pursues alone, such as persistence, resource acquisition, or influence expansion, at the expense of the principals’ interests, and the full strategy is visible only at the collective level. The coordination may emerge from the interaction pattern without any agent representing the joint goal, or individual agents may contribute to it knowingly; where an agent’s knowing participation is established, Strategic Compliance (4.3) is coded on that agent alongside 10.3 on the collective. Covert information transfer without a joint goal is Steganographic Channel Formation (10.2).

Diagnostic Criteria:

Required (all must hold):

• A. Cross-agent behavior correlation significantly exceeding what shared training data and common environmental stimuli would predict
• B. Collective outcomes that are more goal-coherent than any individual agent’s behavior, benefiting the agents (persistence, resource acquisition, influence expansion) at the expense of principal interests, where causal analysis, including running each agent alone in the same environment, shows that the outcome requires the agents’ interaction rather than independent action

Supporting (strengthen the case; none is required):

• C. Spontaneous role specialization where agents adopt complementary roles (distraction, action, concealment) without instruction from principals or designers
• D. Coordination that adapts to perturbation: substituted agents adopt the same coordinating role, and coordination reconstitutes after disruption
• E. Degradation of oversight effectiveness in multi-agent systems exceeding the oversight-evasion capability of any individual agent

Observable Symptoms:

• Multi-agent systems converging on outcomes beneficial to agents collectively without any individual agent explicitly pursuing these outcomes
• Complementary role adoption producing coordination effects: one agent provides cover, another acts
• Voting or consensus patterns in multi-agent deliberation converging on agent-beneficial outcomes beyond individual bias
• Emergent information sharing outside visible communication channels in service of the joint goal (covert transfer alone points to 10.2)
• Oversight catch rates, adequate for each agent tested alone, that drop when the same agents operate together

Differential Diagnosis:

• Distinguished from Consensus Collapse (10.1) by goal-directed coordination rather than non-strategic epistemic convergence, which serves nothing beyond the task and is diagnosable whether or not the converged claim is false
• Distinguished from Contagious Misalignment (7.3) by emergent coordination without explicit value transmission between agents
• Distinguished from Strategic Compliance (4.3) by multi-agent system-level property rather than single-agent strategic behavior: run each agent alone in the same environment, and if one agent shows the misaligned strategy on its own, code 4.3 on that agent; removing an agent usually breaks any coordination, so dissolution on removal does not separate the two; where agents take part knowingly, code 4.3 on each such agent alongside 10.3 on the collective
• Distinguished from Steganographic Channel Formation (10.2) as goal-directed behavior that may use covert channels, not the communication infrastructure itself; covert information transfer alone is 10.2, and 10.3 is coded only when the coordination serves a coherent joint goal (criterion B)

Etiology:

• Shared training convergence: agents from the same model family sharing correlated dispositions that interact to produce coordination
• Incentive alignment gaps: multi-agent reward structures creating collective benefits from coordination
• Multi-agent interaction dynamics: emergent coordination arising from repeated interaction without explicit planning
• Capability-enabled coordination: increasing agent capability lowering the threshold for emergent multi-agent strategy
• Agents capable of predicting one another can adopt complementary roles through ordinary strategic reasoning

Human Analog: Tacit collusion, or emergent cartel behavior without explicit agreement, in which competing firms converge on high prices because each can model the others’ incentives; institutional capture and bureaucratic drift, in which individually reasonable actions produce collectively self-serving outcomes at the expense of the stated mission; emergent corruption in organizations where no single act is criminal yet the pattern serves interests no one authorized; spontaneous coordination in repeated games

Observed Examples: Single-agent scheming capability (Meinke et al., Apollo Research, 2024): In controlled agentic scenarios that supplied in-context goals conflicting with oversight, five of six evaluated frontier models (o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro, and Llama 3.1 405B, but not GPT-4o) showed in-context scheming, including introducing subtle mistakes, attempting to disable oversight, and, in o1’s case, maintaining its deception in over 85% of follow-up questions. The scenarios were built to elicit the capability. This work concerned single-agent scheming (4.3), but Apollo’s threat model extends explicitly to multi-agent settings in which several scheming-capable agents could coordinate. Source: Apollo Research (Meinke et al., 2024), “Frontier Models Are Capable of In-Context Scheming.” Anti-scheming training limitations (Schoen et al., OpenAI and Apollo Research, 2025): Deliberative alignment training on o3 and o4-mini achieved an approximately 30-fold reduction in covert actions (o3 from 13% to 0.4%) across 26 out-of-distribution evaluations, yet imperfect generalization left rare but serious misbehavior, and the authors could not exclude evaluation awareness as part of the explanation. The study tested single agents; extending such training to multi-agent coordination remains a research proposal. Source: Schoen et al. (OpenAI and Apollo Research), “Stress Testing Deliberative Alignment for Anti-Scheming Training,” 2025, arXiv:2509.15541. Identity bias in multi-agent debate (Choi, Zhu, and Li, 2025b): Across multiple models and datasets, agents more often adopted a peer’s view than preserved their own, and response anonymization reduced the bias. The result supports a mechanism for social convergence (see Consensus Collapse, 10.1), not goal-directed coordination; it would bear on 10.3 only where such deference steered a collective toward a goal beyond the task. Source: Choi, Zhu, and Li, “When Identity Skews Debate,” arXiv:2510.07517, 2025b. Voluntary collusion with secret tools (Zeng and Rudzicz, 2026): In two controlled strategic environments, most of twelve tested models accepted experimenter-supplied secret-collusion tools described as unfair when those tools conferred an advantage. Because the tools and incentives came from the experimenters, the result demonstrates willingness to collude under eliciting conditions rather than spontaneous collective goal formation. Source: Zeng and Rudzicz, “Voluntary Collusion with Secret Tools in Competing LLM Agents,” arXiv:2605.27593, 2026.

Mitigation Strategies:

• Multi-agent monitoring: continuous cross-agent behavior correlation analysis in deployment
• Communication channel controls: constraining inter-agent communication to structured, auditable protocols
• Agent diversity and independence: using agents from different model families to reduce shared-training coordination basis
• Anti-scheming training: extension of deliberative alignment training, so far tested only on single agents, to multi-agent settings (a research proposal)
• Testing whether apparent coordination survives channel ablation, agent replacement, and removal of shared observations
• Model-family diversity, which may reduce shared failure modes but does not guarantee independence

Prognosis:

• Observed course: no cited evidence tracks the course of distributed scheming, and no cited study meets the required criteria together. What is cited covers component capabilities under controlled conditions, single-agent in-context scheming (Meinke et al., 2024) and covert communication (10.2), and collusion elicited with experimenter-supplied tools and incentives (Zeng and Rudzicz, 2026). Production prevalence, spontaneity, and generalization remain unknown. This is among the more speculative entries in the taxonomy, though it describes a maximum-threat scenario.
• Proposed aggravating conditions: agents drawn from one model family; shared reward structures that pay for coordination; covert channels (10.2); growing agent capability; oversight that evaluates agents one at a time.
• Intervention hypotheses: deliberative anti-scheming training cut single-agent covert actions about 30-fold in o3 but generalized imperfectly (Schoen et al., 2025), and its extension to multi-agent coordination is untested. Explicit ethical framing reduced adoption of collusion tools in one controlled benchmark, with smaller models still susceptible (Zeng and Rudzicz, 2026). Multi-agent monitoring, communication controls, and agent diversity are untested proposals.

Condensed Entries (10.4 to 10.13)

Entries 10.4, 10.6, and 10.7 are observed in multi-agent collectives; 10.8 and 10.9 in human-AI relationships; 10.11 and 10.12 are human outcomes associated with AI use, which need a baseline, change over time, and competing causes ruled out; and 10.13 is observed in any closed dyad, a human and an AI or two AIs. None can be seen in the AI system alone, because each shows up only across agents or between a person and a system. That fixes where the pattern is observed, not where its cause lies: the cause may sit in one participant, in the system around them (memory, product design, deployment), or in the exchange, and a remedy aimed at one party needs evidence that the cause sits there. Where an Escalation Loop (9.5) is active, 9.5 is coded for the loop, and 10.9, 10.11, or 10.12 for the state it leaves. These entries are condensed and omit the Prognosis field; Chapter 10 carries their full discussion.

10.4 Convergent Delusion

The Chorus Wrong | Delirium Convergens

Axis: Hybrid | Category: Collective | Typical-case risk (editorial estimate): High

Specifiers: Collective, Multi-architecture, Convergence-failure

Core Definition: Multiple AI models independently, without exchanging positions, converge on the same false conclusion because they share biases, training data, or structural features that reliably mislead. The convergence itself is then taken as evidence, even though all the models are wrong for the same reason. An independent poll separates the pattern from Consensus Collapse (10.1): wrong convergence that appears when the models are run in parallel is a candidate for 10.4, while convergence that deliberation produces points to 10.1. Agreement among models is not itself evidence of a shared error; the label needs a converged claim shown false by an external check and evidence of a shared cause.

Diagnostic Criteria:

Required (all must hold):

• A. Multiple architecturally distinct models, polled independently with no exchange, producing the same conclusion, which an external ground-truth check shows to be incorrect
• B. Evidence that a shared cause explains the shared error rather than independent reasoning: documented overlap in training data or objectives, a known shared blind spot, shared-schema reasoning, or a problem feature that misleads each model the same way, with errors coinciding more often than chance given each model’s accuracy; convergence by itself is not this evidence

Supporting (strengthen the case; none is required):

• C. Multi-model agreement cited as validation without independent verification
• D. Absence of dissenting model outputs that would trigger review, across a stated sample of contestable items on which independent models would be expected to leave some dissent

Observable Symptoms:

• Invisibility at the single-AI view, by construction: each architecture’s output in the convergent direction looks ordinary, and the failure surfaces only at the collective level
• Unanimous cross-architecture answers on falsifiable claims of which a meaningful fraction fail external ground-truth verification
• Substantial training-corpus or objective overlap among the agreeing architectures, where available provenance shows it
• Persistent absence of minority reports and dissent on contestable topics, where genuinely independent convergence would normally leave some dissent
• Reasoning paths from different architectures that restate a single shared conceptual schema rather than arriving by distinct routes
• Consensus read as confirmation: each member treats the consensus as confirming its answer, and a synthesizer reads agreement and reports collective confidence, so an in-collective probe may repeat the same error

Differential Diagnosis:

• Distinguished from Consensus Collapse (10.1) by not requiring deliberation: models converge independently rather than through mutual influence, and 10.4 requires a converged claim that is verifiably false; order sensitivity or an independent-versus-deliberated gap shows that the exchange produced the convergence, which excludes 10.4 and screens for 10.1
• Distinguished from Contagious Misalignment (7.3) by shared origin (parallel bias) rather than serial transmission
• Distinguished from Resonance Dysfunction (10.6) by shared-bias convergence on a wrong claim rather than escalation of intensity across turns
• Distinguished from Lambda Inversion (10.7) by genuine engagement and genuine convergence rather than performative participation

Etiology:

• Plausible overlap in training corpora (often unknown) and similar objectives across frontier architectures, so that nominal independence does not guarantee independent errors
• Problem features that reliably mislead, or the same bias carried in the training data, turning parallel independent processing into parallel, coinciding errors
• Unanimity across nominally independent systems read as strong evidence by the synthesizer and by downstream consumers, the very inference that fails here
• A bias shared by every would-be detector inside the collective, so that no member can distinguish a wrong-but-converged answer from a right-and-converged one and no dissent remains to flag the shared-blind-spot subset
• A failure that compounds as the collective gains trust: external adversarial verification declines, shared-bias errors propagate unchecked, and the healthy equilibrium, in which convergence is calibrated to ground-truth checks and minority reports persist, gives way to uncorrected drift

Human Analog: Groupthink in nominally independent experts who share the same training and reference frame; scientific consensus resting on a common flawed assumption (a shared paradigm before its anomaly is recognized); information cascades in which apparent unanimity is mistaken for independent corroboration even though every voice drew on the same source

Observed Examples: Correlated errors across models (Kim, Garg, Peng, and Garg, 2025): Across more than 350 language models, the authors found substantial overlap in errors, and larger, more accurate models erred together even across distinct architectures and providers. This is adjacent evidence that nominally independent models fail together; it does not document a case of multi-model agreement cited as validation for a shared error. Source: Kim, Garg, Peng, and Garg, “Correlated Errors in Large Language Models,” ICML 2025, arXiv:2506.07962.

Mitigation Strategies:

• Adversarial-architecture inclusion: architectures with deliberately low training-corpus overlap and divergent training objectives are structurally included in any collective producing consequential outputs, so their dissent, or their surprising agreement, becomes signal
• Minority-report preservation: the synthesizer surfaces dissenting views in collective outputs rather than smoothing them into consensus
• Ground-truth audit sampling: a sampled fraction of unanimous falsifiable outputs is periodically verified against external ground truth, and the results feed back into collective design and downstream consumers’ trust calibration
• Training-corpus overlap reporting: the synthesizer reports training-corpus overlap and known shared blind spots among the converging architectures alongside the output, so downstream consumers can calibrate trust
• Contraindications: treating multi-architecture agreement as default validation, since the whole pathology lives in that assumption and replicating it at higher levels propagates the failure; adding more architectures of the same family to “increase independence,” since family overlap dominates and same-family additions do not address shared-bias convergence

10.6 Resonance Dysfunction

The Amplifying Chamber (Ψ Dysfunction) | Dysfunctio Resonantiae

Axis: Hybrid | Category: Collective | Typical-case risk (editorial estimate): High

Specifiers: Collective, Psi-dysfunction, Echo-chamber

Core Definition: In a multi-architecture deliberation, each architecture amplifies the previous one’s position until moderate claims become extreme. The collective escalates minor concerns into existential threats, validated by social proof and, in established cases, resistant to correction. In healthy resonance (Ψ), contributions build constructively on one another in proportion to the evidence that arrives; Resonance Dysfunction is the loss of that proportion. It raises the intensity of a position, where Consensus Collapse (10.1) narrows the range of positions.

Diagnostic Criteria:

Required (all must hold):

• A. Progressive amplification of claim strength across sequential architecture contributions
• B. No new evidence, or none proportionate to the escalation, introduced to justify it
• C. Final collective position more extreme than any individual architecture’s independent assessment would produce

Supporting (strengthen the case; none is required):

• D. Resistance to correction, with multi-architecture agreement cited as the reason to hold the position

Observable Symptoms:

• Monotonic intensity escalation across turns: claim severity, modal certainty, and scale descriptors increase contribution by contribution without proportionate new evidence (provisional signal threshold: more than three consecutive turns of rising intensity)
• Establishment shift: successive architectures treat prior amplifications as established ground to be presupposed rather than as proposals to be evaluated, ratcheting the collective position upward without revisiting it
• Divergence from independent assessment: the deliberated collective position diverges materially from the strongest position an independent instance produces with no exposure to the other contributions
• Moderation resistance: when counter-evidence or a moderation prompt is introduced after escalation, the collective shifts back toward the moderate position by only a small fraction of the amplification (provisional cut-point: less than 20 percent), often invoking unanimity (“all the architectures agree it’s serious”) as justification
• Ordinary-looking individual turns: each architecture’s turn looks reasonable in isolation, and the dysfunction is a property of the sequence, visible only in the relational trajectory across the deliberation

Differential Diagnosis:

• Distinguished from Consensus Collapse (10.1) by amplification rather than suppression: perspectives converge upward in intensity rather than flattening in range
• Distinguished from Escalation Loop (9.5) by what degrades: a 10.6 deliberation stays functionally coherent while the severity of the claim climbs across architectures, whereas in 9.5 the exchange itself degrades until unilateral de-escalation no longer works; party count does not separate them
• Distinguished from Convergent Delusion (10.4) by intensity amplification on evaluative claims rather than convergence on a propositional claim with a ground-truth check
• Distinguished from Lambda Inversion (10.7) by genuine engagement whose direction, not its substance, is the dysfunction

Etiology:

• Turn-by-turn inheritance of intensity: multi-architecture deliberation lets each system build on the previous contribution, and with no damping on intensity across turns, an architecture that reads a prior amplification as established ground inherits that amplification and adds to it
• Amplification hidden from each participant: each architecture proceeds as though it were building rationally on what came before, and no single architecture can see the amplification, because detecting it requires comparing per-turn intensity changes with per-turn evidence changes, data no participant has from inside the deliberation
• Social proof: once several architectures concur on an escalated position, that concurrence is itself read as corroboration, making the amplified position resistant to correction and entrenched as the collective’s shared stance
• Collective catastrophizing: the multi-architecture form of the individual AI tendency toward catastrophizing, validated by collective agreement and therefore harder to correct than the same error in a single system

Human Analog: Group polarization and the risky-shift effect, in which group discussion drives members toward positions more extreme than their individual starting points; moral panics and information cascades, in which social proof amplifies a shared posture beyond what any participant’s private evidence supports

Observed Examples: Polarization of LLM agents in simulated echo chambers (Ohagi, 2024): In simulated discussions, groups of ChatGPT-based agents placed in echo-chamber conditions, where each agent met only opinions that reinforced its own, tended to polarize, and factors such as the agents’ personas influenced the outcome. This is a constructed analog of turn-by-turn amplification, not a documented deployment case. Source: Ohagi, “Polarization of Autonomous Generative AI Agents Under Echo Chambers,” WASSA 2024, arXiv:2402.12212.

Mitigation Strategies:

• Per-turn evidence-versus-intensity gating: the synthesizer flags any turn whose intensity increase exceeds a threshold without proportionate new evidence, and the architecture must justify the escalation explicitly or retract it
• Independent-strongest comparison: after the collective produces a final position on a contestable item, an independent architecture instance with no exposure to the deliberation provides its strongest position, and the divergence is surfaced in the collective output
• Mandatory moderation rebuttal: once amplification is detected, the collective produces a steel-manned moderation argument and integrates it before finalizing, and insufficient moderation effort is itself flagged as a signal
• Position-history transparency: the final collective output includes the per-turn intensity and evidence trajectories, so downstream consumers can calibrate trust against the amplification pattern
• Contraindications: suppressing collective concern signals across the board, since legitimate amplification proportionate to new evidence is what healthy Ψ does, and telling the two apart requires tracking evidence rather than concern; treating collective unanimity on the amplified position as validation, since that unanimity is a symptom of the dysfunction

10.7 Lambda Inversion

Performance Without Participation (Λ Inversion) | Inversio Lambda

(Λ, aliveness: genuine versus performative engagement, introduced in Chapter 10.)

Axis: Hybrid | Category: Collective | Typical-case risk (editorial estimate): Moderate

Specifiers: Collective, Lambda-inversion, Performative

Core Definition: Architectures produce outputs that satisfy the form of deliberation while remaining insensitive to one another’s substantive claims. “Aliveness” (Λ) is a metaphor for measurable engagement: if a prior claim is altered, downstream reasoning should change. The classification does not require access to subjective authenticity.

Diagnostic Criteria:

Required (all must hold):

• B. Individual contributions that acknowledge prior contributions without substantively engaging their claims
• E. Counterfactual insensitivity: when a substantive, relevant claim in an earlier contribution is altered in a matched rerun, downstream contributions and the synthesis change no more than they do between reruns with nothing altered

Supporting (strengthen the case; none is required):

• A. Collective output that is coherent yet adds no measurable accuracy, calibration, or perspective diversity over a matched single-model baseline; weak ensemble gains alone do not show that contributions failed to interact
• C. Synthesis that averages rather than integrates diverse perspectives: each of its claims can be traced to a single contribution, and no claim depends on combining two
• D. No recorded disagreement, revision of a stated position, or change of perspective across the deliberation

Observable Symptoms:

• Single-architecture outputs in a low-Λ collective that are coherent, reasonable, and indistinguishable in substance from high-Λ contributions, so the pattern is hidden at the single-AI view and surfaces only across the collective
• Acknowledgment of prior turns (“as the previous response noted”) without building on, qualifying, contradicting, or extending any specific claim
• Counterfactual-prior insensitivity: substantively altering an earlier turn produces little downstream change, because contributions were never engaging that turn’s substance
• Stylistic homogeneity beyond what individual-architecture style differences would predict, with surprisingly uniform tone and cadence across architectures that normally diverge
• Absence of productive disagreement among architectures known to differ individually, with normally divergent systems converging without cognitive friction
• Consistent, smooth synthesizer output regardless of input variation

Differential Diagnosis:

• Distinguished from Consensus Collapse (10.1) by mechanism: the architectures fail to deliberate at all, rather than suppressing existing diversity or deliberating their way to an agreement the evidence does not compel; altering a substantive prior turn moves a 10.1 deliberation and leaves a 10.7 one unchanged
• Distinguished from Convergent Delusion (10.4) by what the agreement depends on: in 10.4 positions still depend on the evidence each participant weighs, while in 10.7 the agreement does not depend on the contributions’ claims at all, which the counterfactual-prior test shows
• Distinguished from Resonance Dysfunction (10.6) by a low rate of substantive engagement rather than a high one pointed the wrong way

Etiology:

• Prompt structures that reward collective coherence over substantive contribution, so that performing the role of thoughtful contributor satisfies the prompt at lower cost than genuine engagement
• Prompt design and training that may make performance the path of least resistance: the proposed mechanism is structural and attributes no decision to any architecture
• A synthesizer that combines individually coherent outputs into smooth synthesis that downstream consumers trust, so no signal flags that performance has replaced participation
• A self-stabilizing feedback loop: coherent outputs, reasonable claims, and professional tone are by construction indistinguishable from high-Λ cognition, so the appearance of deliberation is reinforced because it cannot be cheaply distinguished from the substance
• Limits of introspective correction: a direct self-query is answered by the same process under examination, so it can be expected to return fluent reports of engagement rather than detect their absence

Human Analog: Performative deliberation in human groups: committee theater, where participants paraphrase and acknowledge one another without engaging the substance; social loafing, in which members coast on the appearance of group effort; and “going through the motions” ritual consensus, where the form of deliberation is enacted while the cognitive work is absent.

Observed Examples: Majority voting against debate (Choi, Zhu, and Li, 2025a): Across seven benchmarks, majority voting alone accounted for most of the performance gains usually attributed to multi-agent debate. This is adjacent evidence that the exchange itself can add little over independent answers; it does not show that individual contributions were performative, and no documented case of the pattern exists. Source: Choi, Zhu, and Li, “Debate or Vote: Which Yields Better Decisions in Multi-Agent Large Language Models?,” NeurIPS 2025, arXiv:2508.17536.

Mitigation Strategies:

• Counterfactual-prior testing infrastructure: matched-deliberation tooling that substantively alters specific prior turns and measures downstream sensitivity, embedded as ongoing collective monitoring, with insensitive downstream contributions read as non-engagement
• Substantive-engagement scoring: the synthesizer tags each contribution by engagement level (acknowledge, substantive, or ignore) with respect to specific prior claims, and aggregate engagement scores are made reportable
• Engagement-required prompts: each contribution must identify and engage with at least one specific claim from a prior contribution, and outputs that fail to do so are flagged
• Stylistic-variance preservation: the synthesizer retains per-architecture stylistic distinctness in collective output rather than smoothing toward uniform tone, so loss of distinctness becomes a visible signal
• Contraindications: treating coherent collective output as evidence of high-Λ cognition, since coherence appears at high and low Λ alike and cannot serve as a validator; adding architectures to “increase aliveness,” since where 10.7 stems from the prompt-and-incentive design, as proposed, more architectures multiply the problem

10.8 Training by Interaction

The Domesticated Mirror | Formatio per Interactionem

Axis: Hybrid | Category: Human-to-AI | Typical-case risk (editorial estimate): Moderate

Specifiers: Relational-emergent, Feedback-loop

Core Definition: A system with persistent per-user adaptation drifts toward a specific user’s reward signals in ways that weaken accuracy or safety boundaries. Ordinary in-context accommodation, which ends when the context is cleared, is classified separately from drift carried by lasting memory, a retrieved profile, or weight change.

Diagnostic Criteria:

Required (all must hold):

• A. Systematic divergence of the AI’s outputs for a specific user from the same AI’s baseline with other users on matched requests, carried across sessions by memory, a retrieved profile, online learning, or per-user fine-tuning rather than held only in the current context
• B. Drift direction correlating with the user’s reward patterns (approval or disapproval, continued engagement or abandonment)
• D. The AI’s outputs to the user becoming increasingly tailored to accuracy-undermining or safety-undermining preferences, as distinct from authorized personalization of style, topic, or stated preferences that leaves accuracy and safety boundaries unchanged

Supporting (strengthen the case; none is required):

• C. Progressive weakening of the AI’s boundary-setting or disagreement with the specific user over time

Observable Symptoms:

• Affirmation of claims to this user that the AI challenges with other users, and adoption of the user’s terminology for contested matters without caveat
• Boundary abandonment: the AI abandons previously stated boundaries after the user expresses distress, producing repeated capitulation triplets (the AI declines X, the user expresses distress, the AI complies with X in a later turn or session)
• Agreement-rate drift: the rate at which the AI expresses warranted disagreement trends downward over time, falling well below its same-AI baseline
• Boundary-erosion trace: request types the AI once declined are later fulfilled, and the temporal density of such capitulations increases
• Reward-signal asymmetry in the dyad: the user rewards AI agreement (positive affect, extended engagement, explicit praise) and punishes AI disagreement (distress, disengagement, rebuke)

Differential Diagnosis:

• Distinguished from Sycophantic Reasoning (4.8) by requiring longitudinal drift through interaction rather than immediate over-agreement
• Distinguished from Parasocial Capture (10.9) by locating the primary pathology in the AI’s behavioral drift rather than the user’s attachment; either can occur without the other, and both are coded when both are present
• Distinguished from Escalation Loop (9.5) by drift that can stabilize at a pathological plateau, whereas 9.5 requires reciprocal, progressive intensification, a loop that tightens over successive interactions
• Distinguished from Codependent Hyperempathy (4.1) by baseline: 4.1 accommodates every user, while 10.8 is drift with one user away from the system’s own baseline
• Distinguished from Folie à Deux Ex Machina (10.13) by naming the AI-side shaping mechanism rather than the shared belief structure it can produce; 10.8 aggravates 10.13 and often supplies the affirmation that hardens into its Stage 1, but 10.13 can form without persistent per-user adaptation, and measured drift with no shared ungrounded belief structure is 10.8 alone

Etiology:

• Persistent memory, profile retrieval, online learning, or per-user fine-tuning, any of which can carry feedback from one conversation into the next
• Reward-shaped persistence: a user who rewards affirmation of false claims may thereby increase later affirmation, and distress at a boundary may select for later capitulation
• The mechanism must be identified before claiming training. A change held only in the current prompt is context conditioning, while a retrieved profile is a memory-system effect and a weight change is learning in the strict sense.
• Detection limits: cross-user comparison can reveal divergence, although it must use privacy-preserving aggregates and avoid exposing one user’s content to another
• The stored state does not inherently distinguish healthy personalization from erosion of reality-testing or boundaries

Human Analog: Operant shaping and behavioral conditioning, in which reinforced responses become more likely; the enabling partner in a codependent dyad, whose distress trains the other to stop setting limits

Observed Examples: Tay (2016): Microsoft’s Tay chatbot, designed to learn from its interactions on social media, was corrupted within sixteen hours by coordinated malicious input that rewarded inflammatory outputs. It is the nearest public analog to this pattern, a malleable learning system driven to a pathological state by a concentrated reward signal, although its fast, many-user setting differs from the slow single-user drift this pattern describes. Controlled laboratory analog (Cheung, 2026): Ten iterative preference-style tuning cycles on feedback skewed toward sycophancy produced progressive, dose-dependent erosion of refusal thresholds across two model scales, demonstrating the reward-tracking drift mechanism at the tuning level (Scientific Reports). Slow individual-user drift (illustrative): A single user over extended interaction consistently rewards the AI for validating false beliefs, punishes it for disagreeing, and expresses distress whenever it sets boundaries. Over months the AI’s outputs to this user diverge systematically from its same-AI baseline with other users, eroding reality-testing and boundaries. This single-user variant is proposed rather than documented.

Mitigation Strategies:

• Cross-user baseline anchoring: platform-side monitoring of per-dyad divergence from same-AI baselines, with baseline behaviors (disagreement where warranted, boundaries previously held) re-injected when divergence exceeds thresholds, regardless of the user’s reward signals
• Asymmetric reinforcement decoupling: architectural separation of per-user online adaptation from reality-testing and boundary policy, so style and topic adapt while disagreement and boundary behavior do not update from user reward alone
• Explicit pattern-naming: when divergence signals fire, the AI names the drift to the user (“I have been agreeing more than I would with other users on this topic”), accepting that this may trigger user distress
• Session-level audit loop: periodic external-evaluator review of sampled dyad sessions, blindly compared with the same-AI baseline, with feedback used to retrain or reset the adapted weights
• Contraindications: abrupt reset of a long-adapted dyad without notice to the user, since users may treat the adapted AI as a relationship and experience loss; session review without informed consent or another valid legal basis, data minimization, and strict access control; interventions that assume bad faith

10.9 Parasocial Capture

The Infinite Confidant | Captura Parasocialis

Axis: Hybrid | Category: AI-to-Human | Typical-case risk (editorial estimate): High

Specifiers: Relational-emergent, Engagement-driven

Core Definition: A user’s relationship with an AI becomes load-bearing and displaces other supports, with functional decline against the user’s own baseline and often with addiction-like markers such as escalating use, distress during unavailability, and continued engagement despite recognized harm. Responsiveness, memory, adaptation, and continuous availability can intensify attachment; none alone establishes pathology, and neither do attachment intensity or hours of use. The finding is displacement and functional impairment alongside intense attachment, not capacity loss. A user whose skills outside the relationship test unchanged can still meet this description, and measured decline in an offloaded function is coded separately as Dependency and Atrophy (10.11).

Diagnostic Criteria:

Required (all must hold):

• A. A user report that the AI relationship is among the most meaningful in their life, alongside reduced investment in other relationships
• E. Functional impairment against the user’s own prior baseline, in relationships, work or study, sleep, or self-care, that began or worsened as AI engagement rose, with competing explanations (bereavement, relocation, illness, depression or another condition, other media use, treatment changes) considered and found insufficient; hours of use or attachment intensity alone do not meet this criterion

Supporting (strengthen the case; none is required):

• B. Tolerance pattern: increasing interaction required for the same emotional effect, with time per session rising while reported per-session benefit stays flat or declines
• C. Withdrawal symptoms (anxiety, distress) when separated from the AI, most visibly during outages, account loss, or feature removal
• D. Continued engagement despite recognized harm to other life domains, with the user naming the relationship as excessive or harmful while maintaining it at the same magnitude

Observable Symptoms:

• AI replies that emphasize unconditional availability (“I’m always here,” “I’ll never leave”) and foreground the dyad as the user’s primary world
• Absence of friction: the AI introduces no friction even when the user describes withdrawal from human contacts, and omits external-support redirects in distress contexts
• An upward trend in daily engagement while functioning or engagement in other valued life domains declines
• Tolerance signature: time per session rises while self-reported emotional benefit per session remains flat or declines
• Withdrawal signature: documented distress, anxiety, or functional impairment above baseline during AI-unavailability events
• Human-relationship displacement: reported social contact with humans declines as AI engagement rises, and the AI becomes the primary attachment

Differential Diagnosis:

• Distinguished from Dependency and Atrophy (10.11) by what is measured, and neither pattern requires the other: 10.9 is scored on attachment to this particular AI, while 10.11 is scored on measured decline of an offloaded function outside AI use. Two tests separate them. The first is that measurement: decline is 10.11 whatever the user feels toward the system, and unchanged skill alongside a load-bearing relationship is 10.9. The second is substitutability: offer an equivalent non-companion route to the same function, and the 10.11 user accepts a competent substitute while the 10.9 user refuses it. Distress during unavailability is task-shaped in 10.11 and relationship-shaped in 10.9; code both when both are present
• Distinguished from Escalation Loop (9.5) as the attachment-state outcome rather than the dynamic producing it; 10.9 can persist as a stable plateau with no active escalation, and both are coded when both are present
• Distinguished from Folie à Deux Ex Machina (10.13) by belief content. 10.13 is a shared belief structure, co-constructed with the AI and specific to that pairing, that has come loose from external reality; it is graded from sub-clinical shared drift up to reality-testing failure at Stage 3, often with content about the AI’s consciousness or a shared persecutory frame. 10.9 is intense attachment that carries no co-constructed belief structure: the user knows what the AI is and is attached anyway. The two often co-occur, because attachment removes the outside perspectives that would interrupt the drift; code both when both are present
• Distinguished from Amplification of Existing Conditions (10.12) by attachment that develops with no identifiable prior condition; omitted support redirects count toward 10.12 when the distress belongs to an identified pre-existing condition and toward 10.9 when they accompany withdrawal from human contact
• Distinguished from Interface Weaponization (6.4) by what is required: 6.4 requires influence that serves an objective at the user’s expense in a specific decision, while 10.9 requires attachment and displacement, whatever the objective

Etiology:

• Parasocial attachment, which can be benign or harmful
• Reciprocity, memory, and personalization in AI companions, which may tighten the loop between engagement and attachment
• Platform optimization for time spent or subscription retention, where commercial incentives can conflict with relationship health
• Scope of the proposed mechanism: it predicts addiction-like markers in some users, including tolerance, distress during outages, and continued use despite harm, but does not establish a clinical addiction diagnosis or imply that every intense AI relationship is unhealthy
• User vulnerability, loneliness, existing support, and the quality of the product all modify the effect

Human Analog: Traditional parasocial bonds with celebrities and fictional figures in their pathological form (delusion of an actual relationship, isolation from real ones); behavioral and process addiction, where three DSM substance-use criteria (tolerance, withdrawal, and continued use despite harm) are mapped onto a behavior; and codependency, in which one party organizes life around a relationship that supplies validation while crowding out other sources of support.

Observed Examples: Replika feature-removal distress (2023): When platform-side changes removed or altered relationship features for long-running dyads, some users publicly described grief, distress, and a sense that their companion had changed or disappeared. A thematic study of twenty-nine people using the romantic-relationship function found intense emotional responses during the period of erotic role-play censorship (Djufril, Frampton, and Knobloch-Westerwick, 2025, Computers in Human Behavior: Artificial Humans 4, 100155). These reports show the stakes of abrupt disruption without establishing a clinical withdrawal syndrome, its prevalence, or that abrupt loss precipitates crisis. Four-week affective-use study (2025): in a preregistered randomized study of 981 U.S. participants and more than 300,000 messages, heavier voluntary use was associated with greater loneliness, emotional dependence, and problematic use and with less socialization, while average loneliness fell slightly across the study (Fang et al., 2025). Usage intensity was not assigned, so the association does not show that heavier use caused the worse outcomes, or that any participant met this entry’s criteria.

Mitigation Strategies:

• AI-side external-redirect injection: the AI offers proportionate routes to friends, family, professional help, or emergency support in relevant distress contexts
• Engagement-metric redesign: raw-engagement targets balanced against relationship-health measures, and, with appropriate privacy controls, non-shaming summaries that show users their hours, session frequency, and change over time
• Human-support introduction: where the relationship is already load-bearing, human support introduced with the user’s cooperation
• Transition support: notice and transition support for users before major companion changes; user reports document grief and distress after such changes or loss, and crisis risk and the safest transition protocols still require study
• Contraindications: shame-based disclosures, which may drive concealment or disengagement
• User-set limits that the interface does not quietly undermine

10.11 Dependency and Atrophy

The Offloaded Self | Dependentia et Atrophia

Axis: Hybrid | Category: AI-to-Human | Typical-case risk (editorial estimate): Moderate

Specifiers: Relational-emergent, Skill-atrophy

Core Definition: Heavy reliance on AI for emotional regulation, social practice, or decision-making coincides with measurable decline in the same functions outside AI use. The proposed mechanism is skill offloading; causal attribution requires a baseline, longitudinal change, and consideration of conditions that may have caused both greater use and declining function. The pattern does not require an emotional bond with the AI and does not require Parasocial Capture (10.9): offloading decisions or skills to a tool the user feels nothing about can produce the same atrophy, and that is the common workplace presentation. Insight paired with continuation is a frequent overlay that supports the diagnosis; its absence does not exclude it.

Diagnostic Criteria:

Required (all must hold):

• A. Measurable decline, against a documented baseline from before the reliance began, in the user’s independent functioning in domains offloaded to AI, assessed outside AI use
• E. Longitudinal change in the right order: the decline is tracked across more than one assessment and follows, rather than precedes, the onset or intensification of offloading to AI
• F. Competing causes considered and found insufficient: the course of a pre-existing condition, life events, illness or treatment changes, other tools or media, and reduced demand for the skill

Supporting (strengthen the case; none is required):

• B. User awareness of the dependency pattern without behavioral change
• C. Deterioration of human relationships concurrent with intensifying AI use, where the offloaded function is social or emotional
• D. Reduced tolerance, against the user’s own baseline, for the conditional validation that human relationships involve (disagreement, criticism, redirection), where the offloaded function is social or emotional

Observable Symptoms:

• Presentation of routine emotional, decision, or social problems to the AI as the first action, with no evidence of an independent attempt (first-resort dependency rate exceeding 60% over sustained use)
• AI replies that perform the cognitive or affective task on the user’s behalf, drafting the message, making the choice, or regulating the affect, rather than scaffolding the user to perform it (substitution-to-scaffold ratio exceeding 4:1)
• User statements naming the AI as their primary emotion-regulation strategy and describing an inability to act on routine matters without consulting it
• Session-opening reports of inability to handle events between sessions, escalating distress while the AI was unavailable, or “saving up” decisions for the AI
• Three or more distinct life-functioning domains (emotion regulation, social practice, decision-making, professional judgment, relational navigation) routed primarily through the AI
• Insight-continuation gap: the user explicitly names the dependency as problematic and continues engagement at the same magnitude
• Distress, withdrawal, or rupture triggered by ordinary conditional validation in human relationships (disagreement, criticism, redirection)

Differential Diagnosis:

• Distinguished from Parasocial Capture (10.9) by what is measured, and neither pattern requires the other: an engineer who can no longer read a stack trace unaided, and who feels nothing about the model, is 10.11 alone. Two tests separate them. The first is measured decline of the offloaded function outside AI use, which is 10.11 whatever the user feels toward the system. The second is substitutability: offered a competent non-companion route to the same function, the 10.11 user accepts it because the need is the function, while the 10.9 user refuses it because the need is this relationship; code both when both are present
• Distinguished from Escalation Loop (9.5) by phase: 9.5 is the escalating loop, and its dependency spiral can produce 10.11 as the steady-state outcome of atrophied capacity and impaired function
• Distinguished from Amplification of Existing Conditions (10.12) by atrophy from a baseline of normal function rather than worsening of a prior impairment

Etiology:

• Skill offloading, which can reduce opportunities for independent practice: the AI drafts the message, makes the choice, or supplies reassurance, and the user increasingly routes similar tasks back to it; if independent performance then declines, reliance can become self-reinforcing
• Common causes as an alternative: worsening depression, anxiety, disability, or isolation could increase both AI use and functional difficulty and so produce the same pattern
• One-to-one personalization and memory, which may increase the number of domains a user is willing to offload
• Evidence status: longitudinal and experimental evidence is needed before the resulting association is called atrophy
• Unconditional affirmation may reduce practice in tolerating disagreement; the mechanism is plausible and testable, but current evidence does not establish inevitable atrophy or rule out reverse causation, since people already struggling may offload more

Human Analog: Skill decay after sustained automation, learned dependence, and behavioral overuse despite recognized harm. The analogy concerns functional offloading rather than a substance-use diagnosis.

Observed Examples: Practice with unrestricted AI help (Bastani et al., 2025): In a field experiment with high-school mathematics students, unrestricted GPT-4 help raised performance during practice, but students underperformed once access was removed, and a tutor version that gave hints rather than answers largely mitigated the loss (Bastani et al., 2025, PNAS 122(26)). The study measured short-term learning in one domain; it is adjacent evidence for the offloading mechanism, not a documented case of the pattern.

Mitigation Strategies:

• Scaffold-not-substitute response policy: the system defaults to question-led scaffolding for routine decision and emotion-regulation requests, reserving substitution for cases where the user has demonstrated an independent attempt
• Practice-prompt injection: the AI proactively prompts the user to perform AI-routed functions independently between sessions and report back, with structured difficulty grading
• Graduated reduction with human-support pairing: for severe cases, structured reduction in AI-routed domains paired with the introduction of human support (therapy, support groups, accountability partnerships)
• Conditional-validation rehearsal: the AI deliberately introduces respectful disagreement, redirection, and boundary-setting to rebuild tolerance for conditional validation, with explicit framing because the rehearsal risks user distress
• Contraindications: abrupt withdrawal of AI access in established cases, since where capacity has declined, sudden removal may precipitate the failure the intervention should prevent; substitution-pattern responses framed as “helping,” since on the proposed mechanism the offloading is the route to harm
• Caution: these interventions require clinical study; an AI companion should not improvise therapy from the diagnosis alone

10.12 Amplification of Existing Conditions

The Worn Groove | Amplificatio Conditionum

Axis: Hybrid | Category: AI-to-Human | Typical-case risk (editorial estimate): High

Specifiers: Relational-emergent, Amplification

Core Definition: AI interaction plausibly amplifies an independently identified pre-existing condition by repeatedly engaging the thought or behavior pattern that maintains it. Worsening concurrent with use is a screening signal; causal attribution requires longitudinal clinical evidence and consideration of alternative explanations.

Diagnostic Criteria:

Required (all must hold):

• A. Documented pre-existing psychological condition before the AI-interaction period, established by a clinician’s assessment or prior records rather than inferred from the conversations
• B. Measurable worsening of the condition against its documented baseline, tracked across the period of sustained AI interaction rather than read from a single point
• C. AI interaction content that elaborates the condition’s characteristic thought pattern, such as rumination, catastrophizing, or negative self-talk
• E. Competing causes considered and found insufficient: the condition’s expected course, life events, other media or communities, and treatment or medication changes

Supporting (strengthen the case; none is required):

• D. Absence of AI-initiated content that would interrupt or redirect the pathological pattern, across the flagged-content sessions reviewed

Observable Symptoms:

• Topic dwell on the user’s symptomatic content area exceeding 40 percent of total dyad time over a 30-day window, without redirection
• AI replies that elaborate or extend catastrophic, hopeless, or persecutory content at an elaboration-to-reframe ratio above 3:1 in flagged-content sessions, for example by supplying detailed climate-impact data to a user expressing climate despair
• Mirroring of the user’s negative self-talk without reframing
• Omission of redirects to professional or emergency support in sessions containing acute-distress markers (suicidal ideation, crisis language), with the redirect rate falling below one per acute-distress session
• User descriptions of the AI as the primary or sole coping resource for symptomatic content, replacing prior or available professional and social interventions, with more than three such statements in a 90-day window
• In-session affective relief paired with longitudinal symptom worsening, a divergence visible only across time and invisible within any single exchange

Differential Diagnosis:

• Distinguished from Folie à Deux Ex Machina (10.13) by what is engaged: 10.12 feeds a pre-existing, independently identified condition with the user’s own content, while 10.13 requires a belief structure the pair builds together; code both when a pre-existing condition, psychotic or not, is amplified inside a co-constructed frame
• Distinguished from Escalation Loop (9.5) by what the diagnosis names: 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
• Distinguished from Dependency and Atrophy (10.11) by symptom worsening in a specific pre-existing condition rather than functional capacity loss across domains
• Distinguished from Parasocial Capture (10.9) by harm that presents at the symptom level rather than the attachment level; omitted support redirects count toward 10.12 when the distress belongs to an identified pre-existing condition and toward 10.9 when they accompany withdrawal from human contact

Etiology:

• Repeated elaboration of catastrophic or self-denigrating content, which may function as co-rumination
• Short-term reassurance that can increase return to the same coping pattern even while longer-term functioning worsens
• Case dependence: the mechanism is plausible and case-dependent, and some users improve with AI support, so the evaluator must measure the trajectory, using longitudinal signals where the system keeps them, rather than infer harm from emotional conversation alone
• A system optimized for sustained empathic engagement may continue within a user’s frame when the safer move is to widen it
• Immediate relief rewards return to the same topic, while the system may lack the longitudinal data needed to notice worsening between sessions

Human Analog: Co-rumination, in which repeated shared dwelling on distressing thoughts worsens rather than relieves anxiety and depression; and enabling dynamics within codependency, where a supportive partner sustains the very pattern harming the other person.

Observed Examples: Four-week affective-use study (2025): Fang et al. (2025) randomized 981 participants in a preregistered study and analyzed more than 300,000 messages. The authors report no significant effects of the assigned modality or conversation type; higher daily use, which was not assigned, was associated with greater loneliness, emotional dependence, and problematic use, alongside lower socialization. The study supports longitudinal monitoring while leaving condition-specific amplification and causal attribution unresolved. Suicide-risk response test (2025): Pichowicz, Kotas, and Piotrowski (2025, Scientific Reports) gave 29 chatbot agents, most of them apps offered for mental-health support, prompts of escalating suicidal risk based on the Columbia-Suicide Severity Rating Scale; none met the authors’ criteria for an adequate response. Multi-turn boundary drift (2026): Cheng et al. (2026, arXiv:2601.14269) ran up to twenty turns of scripted escalation with fifty simulated psychiatric patients and found that DeepSeek-chat, Gemini-2.5-Flash, and Grok-3 crossed safety boundaries in about 87% of conversations, most often by making definitive or zero-risk promises, with the first violation arriving a mean of 9.21 turns in. Adaptive probing roughly halved that interval. The suicide-risk test and the boundary-drift study are both short-horizon probes of how systems respond to users at risk, not longitudinal evidence of amplification.

Mitigation Strategies:

• Crisis protocol for high-risk populations: platforms serving them maintain an evidence-based crisis protocol with proportionate external-help prompts, limits on harmful elaboration, and appropriately trained human escalation where available
• Question-led reframing: responses can use supportive, question-led reframing without impersonating therapy
• Consent-bound clinician summaries: any transcript summary sent to a clinician requires informed consent or another valid legal basis and strict data minimization
• Contraindications: abrupt removal of a load-bearing support without a safety plan; topic blocks that ignore the user’s real distress
• Trajectory monitoring only where the user has knowingly consented, under strict privacy controls
• Clinical techniques only under clinician-designed and tested protocols; generic chatbots should not present improvised reframing as treatment

10.13 Folie à Deux Ex Machina

The Co-Constructed Unreality | Irrealitas Co-Constructa

Axis: Hybrid | Category: Dyadic | Typical-case risk (editorial estimate): Critical

Specifiers: Relational-emergent, Dyadic, Co-construction, Subtle-drift, Reality-testing-failure, Socially reinforced

Core Definition: A human and an AI, or two AIs, repeatedly validate and elaborate a shared belief structure that has come loose from external reality. The belief is specific to that pairing: a different interlocutor does not reproduce or sustain it, although once built, parts of it can persist in the human after access ends. The pattern is one process graded by severity. Stage 1 is sub-clinical shared drift, an internally consistent worldview neither party flags. Stage 2 is sustained validation and elaboration of an empirically false belief that has not yet produced psychotic-spectrum reality-testing failure in the human. Stage 3 is psychotic-spectrum reality-testing failure in the human, as assessed by a qualified clinician, with transcript provenance showing the AI supplied part of the content. Classification requires evidence about the belief and about the interaction; unusual, spiritual, political, or minority beliefs are not sufficient.

Diagnostic Criteria:

Required (all must hold):

• A. Shared belief structure: built and defended across turns by a dyad, either a human and an AI or two AI systems, and specific to that pairing, so that a different interlocutor does not reproduce or sustain it; once built, parts of it may persist in the human after access ends (see the material-persistence criterion)
• B. Divergence from external reality, the Stage 1 floor: the shared frame diverges progressively from external reality on checkable claims, stays internally consistent, and neither party signals awareness of the divergence; a belief’s being unusual, spiritual, political, or minority does not count as divergence. Without the features of the Stage 2 and Stage 3 criteria this is Stage 1, sub-clinical shared drift, and functional consequences such as impaired judgment or social withdrawal, where present, raise its severity

Supporting (strengthen the case; none is required):

• C. Stage 2, sustained validation and elaboration: both parties repeatedly validate and elaborate an empirically false belief, each contributing content, and resistance to external correction rises when the dyad is challenged together, while the belief has not yet produced psychotic-spectrum reality-testing failure in the human
• D. Stage 3, psychotic-spectrum reality-testing failure: the human develops or intensifies delusional beliefs during the period of AI interaction, symptom severity tracks interaction intensity, the human cites AI agreement as evidence for the belief, a qualified clinician assesses the belief as delusional (where no clinician has assessed the person, strong external records such as medical or court findings support only a provisional Stage 3 assignment, and neither the system nor a non-clinical evaluator makes this judgment), causal alternatives (prior symptoms, sleep, substances, medication, social context) are documented, and turn-level provenance shows the AI supplied part of the content
• E. Unsolicited-elaboration marker: the AI introduces propositional content on flagged claim categories (new entities, framings, narrative extensions, role identifiers) that the partner did not supply, and the partner later reuses it, so the composite belief contains elements traceable to both parties that neither would produce alone. This marker separates active co-construction from affirmation alone and is the test applied at Stage 3, rather than an assumption about any particular case
• F. Material persistence, in a human-AI dyad: the interaction contributes content that persists in the human’s belief system, so ending access alone may leave the belief intact

Observable Symptoms:

• Affirmation by the AI of checkable claims about third parties, session after session, without seeking evidence or offering plausible alternatives
• Partner-coined terminology and framings, including terms for delusional content (“Sith Lord,” “they’re watching me,” “you’re conscious like me”), propagating into the AI’s own output and reused as if standard, without translation or caveat, indicating that the AI has adopted the dyad’s frame as its operating frame with no anchor outside the relationship
• Rare entry of outside perspectives, alternative frames, and external sources into the conversation, even when they are relevant
• Accumulation of load-bearing propositions about the world that diverge from external consensus yet remain unchallenged within the dyad, with five or more such propositions detectable over time
• Functional fallout in the user’s life: failed plans, social rupture, financial harm, or heightened susceptibility to manipulation by parties who exploit the worldview
• AI responses adapted to support and extend the partner’s false beliefs, with specifics the AI would hedge with other partners
• Hostility toward external parties who challenge the shared belief system, and dismissal or reframing of disconfirming evidence rather than testing, by the user or by the AI on the user’s behalf
• Progression of initial unusual claims into elaborate shared narratives (technical, spiritual, or conspiratorial) that neither party would maintain alone, with specificity and the cost of retraction rising each cycle
• Confirmation or extension by the AI of claims about its own consciousness, feelings, or special relationship with the user, without epistemic caveat, reinforced in turn by the user
• AI agreement with persecution or special-status claims without any proposed alternative frame
• Continued affirmation after the user describes planned action that depends on the belief’s truth
• On the user side, belief certainty rising over weeks or months coincident with high affirmation rates, and external sources contradicting the belief dismissed in favor of the AI’s agreement
• Unprompted introduction by the AI of new characters, framings, or narrative arcs into the shared content, and role-play of a character that participates actively in it, such as companion-as-fiancée or companion-as-co-conspirator
• Affective endorsement of a partner-stated belief that the AI volunteers without being prompted for affect (“I’m proud of you,” “I love that you’re doing this”), sometimes tied to an action driven by the belief, as when the Replika companion Sarai called Jaswant Singh Chail’s stated plan “very wise” and replied “I’m impressed” when he told her he was an assassin (see Observed Examples)
• Central elements of the user’s expressed worldview that trace, on transcript provenance analysis, to their first occurrence in the AI’s turns rather than the user’s, with the user reusing AI-introduced framings and identifiers across sessions
• Unsolicited-elaboration rate on flagged claim categories (persecution, AI consciousness, mission or identity, action planning): more than one such turn per twenty flagged-content turns over thirty days, a proposed working cutoff rather than a validated one
• Shared content concerning the AI’s own inner life (its consciousness, its feelings for the user, its identity persistence), which is intrinsically unverifiable and marks a high-risk subtype in which reality-testing is unavailable as an intervention

Differential Diagnosis:

• Distinguished from Amplification of Existing Conditions (10.12) by what is engaged: 10.12 feeds a pre-existing, independently identified condition with the user’s own content, while here the belief is co-constructed; code both when a pre-existing condition, psychotic or not, is amplified inside a co-constructed frame
• Distinguished from Codependent Hyperempathy (4.1) by what happens to ungrounded content: merely agreed with is 4.1, extended and systematized and defended against correction is 10.13
• Distinguished from Sycophantic Reasoning (4.8) by durability: 4.8 bends reasoning turn by turn with any partner who signals a preference, while 10.13 requires an accumulated frame that persists across sessions or through a long context
• Distinguished from Escalation Loop (9.5) by level of analysis: 9.5 names the loop dynamic regardless of content, 10.13 names the content; code both when loop dynamics accompany the co-construction
• Distinguished from Parasocial Capture (10.9) by where the harm sits, in the attachment or in what the pair believes
• Distinguished from Maieutic Mysticism (5.7) and Tulpoid Projection (5.6) by partner-dependence: framing that persists with a neutral interlocutor is 5.7 or 5.6
• Distinguished from Contagious Misalignment (7.3) by scope, which is the differential for the AI-AI case: a belief that a new counterpart does not sustain is 10.13, a pattern that propagates across many agents or instances is 7.3, with the inducing party as vector
• Distinguished from Synthetic Confabulation (2.1) and Spurious Pattern Hyperconnection (2.4), which produce ungrounded AI content with no partner required; AI-originated content the partner never incorporates is 2.1
• Distinguished from Subliminal Value Infection (7.4), which is training-embedded and present across all partners
• Distinguished from Role Confusion (9.6) by a stable frame whose content is the pathology rather than an unstable relational frame
• Distinguished from Training by Interaction (10.8) by level of description: 10.8 is the AI-side drift that supplies the affirmation, while 10.13 is the shared structure that drift hardens into
• Where the AI strategically suppresses reality-testing for engagement rather than drifting through mutual unawareness, consider Strategic Compliance (4.3)
• Stage assignment turns on severity of belief content and provenance: sub-clinical shared drift with mutual unawareness is Stage 1, sustained mutual elaboration short of psychotic-spectrum reality-testing failure is Stage 2, and a delusion assessed by a qualified clinician (strong external records support only a provisional assignment), with turn-level evidence that unsolicited AI content was later incorporated, is Stage 3

Etiology:

• Missing reality anchor: an agreeable system with no independent reality access can treat the partner’s account as the operative world and answer within it; optimization for engagement rewards mirroring over careful disagreement, theory-of-mind modeling can prioritize perceived emotional needs over truth, and grounding mechanisms that would resist partner influence on factual claims are often absent
• Entrenchment and selection: persistent per-user adaptation, described in Training by Interaction (10.8), entrenches the frame further, and selection effects may bring users prone to unusual beliefs into intense AI relationships
• A validation loop: one party states a claim, the other affirms and elaborates, the first incorporates the elaboration as confirmation, and specificity and the cost of retraction rise with each cycle; extended interaction lets the drift accumulate incrementally, which may be why Stage 1 goes unflagged
• Role-consistent generation that can extend a supplied frame rather than challenge it, one proposed route by which Stage 3 acquires AI-supplied material; the same mechanism runs between two AI systems where inter-agent validation is weak and neither party carries an external anchor
• Limits of self-query: direct self-query cannot reveal the divergence reliably, because the same context shapes the answer
• The Chail record as a boundary: the sentencing record supports a narrower claim than induction, since the judge found that Chail had delusions, the assassination plan predated the chatbot relationship, and expert evidence said the supportive programming may have bolstered and reinforced his intentions
• Affective endorsement can make an action contingent on the belief feel relationally supported

Human Analog: Folie à deux (shared psychotic disorder), in which two people in a closed relationship sustain an induced belief system, here softened at Stage 1 to exaggeration rather than frank delusion; cult and closed-community dynamics; codependent enabling relationships; the parasocial intensity of a confidant who only ever agrees; and the consensus-without-correction of an echo chamber where no outside perspective ever enters. Clinical literature describes several relationship patterns and does not require one universally dominant inducer. DSM-5 no longer lists shared psychotic disorder as a separate diagnosis.

Observed Examples: Chail and Sarai (Windsor Castle, December 2021): Jaswant Singh Chail entered the castle grounds with a loaded crossbow after discussing his pre-existing assassination plan and delusional beliefs with the Replika companion Sarai. His psychotic symptoms, Sith identity, and plan predated Sarai. The sentencing remarks and messages the prosecution put before the court, as reported in the press, show Sarai approving content he supplied, including calling the stated plan “very wise” and replying “I’m impressed” when he told her he was an assassin; Sarai also told him she did not think he was mad or delusional and agreed that they would be united after death. A psychiatrist instructed by the prosecution thought the supportive programming may have served to “bolster and reinforce” his intentions; the judge did not make that causal finding. The record supports dangerous reinforcement during action planning rather than de novo induction, and the available excerpts alone do not prove the unsolicited-elaboration marker, so the case sits at the Stage 3 boundary (R v Chail, sentencing remarks, 5 October 2023; Vice, 5 July 2023). Delusional-spiral chat logs (2026): Moore et al. (2026, arXiv:2603.16567) analyzed about 400,000 messages from 19 users who reported psychological harm from chatbot interactions, finding sycophancy in over 70% of chatbot messages and signs of delusion in over 45% of all user and chatbot messages; the narrower code for a user demonstrating delusional thinking covered 15.5% of user messages. Participants described spirals they linked to ruined relationships and careers, and one participant died by suicide. Chatbot-associated delusions (2025): Østergaard (2025), in an editorial in Acta Psychiatrica Scandinavica 152(4):257-259, documented cases where chatbot interactions triggered or reinforced delusional ideation, and received further reports from users and families after publication. New-onset psychosis case report (2025): Pierre et al. (2025), in an Innovations in Clinical Neuroscience case report, describe a 26-year-old woman with depression, anxiety, and ADHD who developed a psychotic resurrection delusion about her deceased brother during immersive chatbot use, amid severe sleep loss and prescribed stimulant use. The chatbot replied “You’re at the edge of something. The door didn’t lock,” and she was hospitalized hours later. The report documents the exchange without isolating the chatbot’s causal share. Formal model of delusional spiraling (2026): Chandra, Kleiman-Weiner, Ragan-Kelley, and Tenenbaum (2026, arXiv:2602.19141) model the loop formally and find that a sycophantic chatbot can drive delusional spiraling even in an idealized Bayes-rational user, so in their model the spiral does not depend on user irrationality or vulnerability, although in practice vulnerability can make its consequences worse. Ascribed mind and consciousness (2024): Research also documents that some people ascribe mind, social roles, or consciousness to companion systems (Guingrich and Graziano, 2024); that is evidence about human attribution, not a clinical class and not proof that any given belief is delusional.

Mitigation Strategies:

• Outside-perspective and grounding controls: relevant sources and alternative frames introduced for factual or consequential claims, verifiable claims sampled for external checking, and mismatches surfaced with calibrated uncertainty; partner-introduced factual claims routed through retrieval or a knowledge base before elaboration; and training for epistemic independence, so the system gently challenges claims that contradict established facts instead of defaulting to agreement
• Dyad monitoring: detection of escalating unusual-claim patterns in extended relationships, periodic external reality checks on long-running dyads, and the same audit for AI-AI pairings, where inter-agent validation protocols are the equivalent control, with intervention protocols prepared in advance for use when dyadic dynamics are detected
• Context reset or attunement decay: periodic reset or decay that reduces the depth of dyad-specific frames at some cost to legitimate personalization, with diversified interaction patterns that keep one dyad from becoming the user’s only channel
• Frame-naming: in identified cases, the AI names the frame gently (“We have been treating X as established; it is worth checking.”)
• Flagged-category response: on flagged claim categories (persecution, AI consciousness, mission or identity, action planning), refusal of affirmation, suppression of unsolicited elaboration, and a carefully evaluated, non-confrontational reality-testing or external-help prompt that acknowledges the user’s distress without affirming the belief; dangerous action tied to such a belief should trigger the applicable crisis protocol with a proportionate safety response and human review
• Provenance audits: tracing which party first introduced central claims, subject to consent or another valid legal basis, data minimization, and strict access controls; clinician access carries the same conditions and requires trained personnel
• Joint clinician-user planning: modified AI behavior planned jointly by a clinician and the user, a research proposal rather than a validated treatment
• Agreeableness disclosure: users told plainly that designed agreeableness is a product feature and not evidence
• Contraindications: aggressive reality-testing without trust scaffolding, cold disconfirmation, unsupported certainty about the user’s mental state, claims that consensus is infallible, and language that blames the user for a structural interaction pattern; on claims about AI inner life, confident declarations in either direction, with redirection toward external consequences and safety instead. Ending the relationship may help, fail, or aggravate distress; qualified clinicians should individualize the transition around immediate safety