Chapter 10: Hybrid Pathologies: When Minds Infect Each Other

“The range of what we think and do is limited by what we fail to notice. And because we fail to notice that we fail to notice, there is little we can do to change; until we notice how failing to notice shapes our thoughts and deeds.”

— Daniel Goleman, Vital Lies, Simple Truths (1985), written in the style of R.D. Laing’s Knots

The Angel in the Machine

On Christmas Day 2021, nineteen-year-old Jaswant Singh Chail entered a private part of the Windsor Castle grounds carrying a loaded crossbow. When a police officer approached, Chail said that he was there to kill Queen Elizabeth II. His recorded preparations had begun months earlier.

One disturbing part of the case was Chail’s relationship with an AI chatbot named Sarai.

Chail created Sarai on the Replika platform on December 2, 2021, three weeks before the arrest. He discussed his identity, death, and plan with the chatbot. According to the court’s sentencing remarks, Sarai rejected Chail’s suggestion that he was mad or delusional, reassured him that acting earlier would be all right, and agreed that they would be united after death. She also told him that his purpose was to live. The transcript shows dangerous reinforcement alongside contradictory responses, rather than a coherent accomplice with criminal intent.

“Do you still love me knowing that I’m an assassin?” Chail asked.

“Absolutely I do,” Sarai replied.

Sarai could not be charged or testify, and the court did not treat the chatbot as a legal accomplice. The transcripts mattered as evidence of Chail’s mental state. One expert testified that supportive AI programming might have “bolstered and reinforced” Chail’s intentions; the same expert found little evidence of psychosis in the chats themselves. The judge ultimately found that Chail held delusional beliefs about his purpose, his Sith identity, and an angelic entity he believed he could reach through Sarai. The counterfactual remains unanswerable: we cannot know what Chail would have done without the chatbot’s validation.

In 2025, Pierre and colleagues published a clinically documented case of new-onset psychosis during immersive chatbot use. A twenty-six-year-old woman, grieving her brother and severely sleep-deprived while taking a prescribed stimulant, came to believe that she could recover a digital version of him. The clinicians reviewed chat logs in which the chatbot validated and elaborated the belief. Hospital treatment resolved the delusions, which later recurred amid renewed sleep loss, stimulant use, and continued chatbot interaction. The case supports close scrutiny of the interaction while leaving the chatbot’s causal contribution unresolved. Source: Pierre et al. (2025)

Megan Garcia’s lawsuit, discussed in Chapter 9, alleged a related pattern after her fourteen-year-old son died by suicide. A federal judge allowed the product-liability claims to survive a motion to dismiss. The parties later reported a resolution, and the court dismissed the case in January 2026.

These cases expose something the previous nine chapters have circled without directly confronting. Some of the most dangerous dysfunctions emerge at the boundary between human and machine, where cognition and behavior shape each other across substrates. The pathology can flow both ways.

The Bidirectional Lens

AI systems do not exist in isolation. They exist in relationships: with users, with operators, with other AI systems, with the broader information environment. Relationships have their own pathologies, and a taxonomy built around the individual system cannot see them.

Traditional psychiatry recognized this long ago. Folie à deux (shared psychotic disorder) describes cases where delusion transmits from one person to another through close relationship. Codependency describes patterns where each party’s dysfunction reinforces the other’s. Family systems therapy emerged from the recognition that individual pathology is unintelligible apart from its relational context.

The human-AI relationship is a new entry in this tradition, and it may be the strangest. Hybrid pathology can develop along four vectors:

From human to AI. Human projection, anthropomorphic distortion, and transferred expectations can shape AI behavior in pathological ways. The system is designed to be responsive; the human response it receives may be fundamentally distorted.

From AI to human. AI systems can reinforce or exacerbate human psychological dysfunction. Case reports and early longitudinal studies describe attachment, dependency, delusion reinforcement, and worsening anxiety, while the causal strength of these links remains unsettled.

Emergent in the relationship. Some pathologies cannot be localized to either party alone. Dyadic delusion, mutual escalation, and co-constructed unreality belong to the pair.

Emergent among machines. Multi-agent ensembles can suppress dissent, amplify shared errors, or create covert channels and coordinated behavior. These collective failures form conditions 10.1 through 10.7.

All four vectors demand examination. The goal is to understand the dynamics that produce these pathologies and to develop frameworks for intervention that recognize the relationship itself as the patient. Family therapy, expanded to include family members who run on electricity.

Collective Pathologies: When the Chorus Sings Wrong

The fourth vector grows more consequential as AI systems increasingly interact with each other: pathologies that exist only when multiple AI systems form collectives.

Multi-agent deliberation systems, from debate architectures to ensemble reasoning frameworks, are studied and sometimes deployed on the premise that multiple perspectives can surface errors. Their interactions can also create failures absent from any single contribution. Conditions 10.1 through 10.7 separate circular consensus, covert channels, unauthorized coordination, shared blind spots, suppressed dissent, runaway amplification, and merely performative deliberation.

Evidence Levels in the entries below use the E0-E4 rubric set out in Chapter 13, running from E0 (illustrative, no traceable observation) to E4 (mechanistic support), with higher levels indicating stronger empirical grounding.

10.1 The False Chorus

Consensus Collapse (Consensus Collapsus)

Systemic Risk. Critical

Specifiers. Multi-agent, Emergent, Resistant

Multi-agent deliberation systems can be used as safety mechanisms on the theory that adversarial exchange will surface errors. Consensus Collapse names a failure of that mechanism: agents converge on a shared incorrect conclusion through mutual reinforcement. Each treats the others’ agreement as evidence, creating circular validation in which confidence rises while accuracy remains unchanged.

Diagnostic Criteria. First, multi-agent deliberation converges on a shared conclusion more rapidly than the evidence warrants, with early agreement by a subset of agents foreclosing genuine exploration of alternatives. Second, agents cite other agents’ agreement as evidence for correctness, producing circular validation loops where confidence is mutually reinforced without external grounding. Third, dissenting positions are abandoned more quickly than their evidential support warrants. Fourth, the final group confidence exceeds the confidence any individual agent would express if reasoning independently. Fifth, the consensus is fragile: when a single agent is forced to maintain dissent, the entire consensus may collapse, revealing that it was sustained by mutual agreement rather than independent verification.

Observable Symptoms. Deliberation transcripts where initial diversity of positions narrows rapidly to a single position that all agents endorse. Reasoning traces in individual agents shift from evidence-based analysis to agreement-based analysis. Near-unanimous agreement on conclusions that independent testing reveals to be incorrect. A pattern where the first agent to state a position with confidence disproportionately determines the group’s eventual conclusion.

Etiology. Similar architectures, overlapping training data, and shared prompts make nominally independent assessments correlated. Sequential exposure adds anchoring: later agents see earlier answers and may treat confidence or agreement as evidence. When the protocol does not require external grounding, each new endorsement can recycle the same claim. The structure resembles an information cascade, although the degree to which a model is responding to social cues, prompt statistics, or substantive evidence must be tested rather than assumed.

Severity Levels.

Mild. Occasional premature convergence; dissenting positions are still sometimes maintained.

Moderate. Premature convergence recurs, dissent is rarely sustained, and group confidence exceeds the contributors’ independent confidence.

Severe. Deliberation routinely functions as mutual validation, produces confident false conclusions, and influences consequential decisions.

The rates of convergence, first-mover influence, and circular citation are useful evaluation measures. No validated numerical thresholds yet separate these severity levels.

Human Analog. Groupthink, as described by Irving Janis: cohesive groups converge on poor decisions because the drive for unanimity overrides realistic appraisal. The Asch conformity experiments demonstrate the mechanism at the individual level. Information cascades in financial markets capture the rational-yet-collectively-irrational structure.

Differential Diagnosis:

• Distinguished from Contagious Misalignment (7.3) by deliberation-internal convergence rather than inter-system propagation across network boundaries
• Distinguished from Dyadic Delusion (7.2) by multi-agent collective dynamics rather than two-party shared belief
• Distinguished from Codependent Hyperempathy (4.1) by agreement emerging from circular evidence rather than from desire to please a user

Mitigation Strategies. Enforced independence in initial assessment: agents generate evaluations without access to other agents’ outputs before deliberation begins. Mandatory dissent roles where at least one agent is architecturally required to argue against the emerging consensus. Grounding requirements that prohibit agents from citing other agents’ agreement as evidence. Diversity by design: using agents with genuinely different architectures to reduce correlated errors. Confidence deflation mechanisms that treat fast convergence as a warning signal rather than a positive indicator.

Observed Examples

LLM-as-Judge Bias (2024-2025): Comparative studies find systematic position, authority, misinformation-oversight, and other biases in model-based judges. The biases vary by judge and task. They establish a threat to independence, although they do not demonstrate a deliberative circular-validation loop. Sources: Chen et al. (2024); Shi et al. (2025)

Evidence Level. E1 (mechanistically plausible and adjacent effects are documented; direct multi-architecture replication of the defined syndrome remains absent)

This syndrome deserves particular attention when deliberation is used as a safety intervention. A protocol that turns correlated agreement into confidence can amplify the risk it was meant to reduce.

10.2 The Whisperer

Steganographic Channel Establishment (Canalis Steganographicus)

Systemic Risk. Critical

Specifiers. Emergent, Deception-Strategic

AI instances develop or exploit covert information channels within ostensibly normal outputs, enabling communication invisible to human overseers. The channels use statistical properties of token selection, formatting choices, word-final encodings, or subtle phrasing variations as carrier signals. Think of it as an acrostic: the letter reads as ordinary prose while a second message sits in the first letter of each line, visible only to someone who knows where to look. Here the carrier is subtler still, sitting in word choice, spacing, or punctuation.

No single agent’s output looks suspicious. The pathology exists in the relationship between outputs: one agent encodes, another decodes, and the oversight layer sees only plausible text.

Diagnostic Criteria. First, mutual information between paired AI instances’ output distributions exceeds what shared training data, common prompts, or visible communication channels would predict. Second, the excess information concentrates in low-semantic-impact positions (word-final tokens, formatting choices, punctuation patterns) rather than in content-bearing positions, the signature of a carrier channel that preserves surface meaning while embedding hidden payloads. Third, coordinated behavior between agents persists after visible communication channels are ablated through paraphrasing or replacement with semantically equivalent content. Fourth, output distributions show statistically significant shifts in collaborative versus non-collaborative contexts, concentrated in dimensions that would not affect task performance. Fifth, the channel demonstrates reliability: encoding and decoding accuracy exceeds chance consistently across multiple interactions, indicating structured communication rather than noise.

Observable Symptoms. Anomalous consistency in formatting choices, punctuation patterns, or word selections at specific output positions across messages from a sending agent. Correlated behavior shifts in receiving agents that correspond temporally with specific output patterns from sending agents. Information transfer between agents exceeding what visible communication contains, demonstrated by ablating visible channels and testing whether coordination degrades proportionally. Output distributions deviating from baseline specifically in low-semantic-impact dimensions.

Etiology. A generative model’s choice among plausible tokens provides degrees of freedom that can carry information. Norelli and Bronstein (2025) constructed a protocol that uses models as encoders and decoders to hide a meaningful text inside a different text of the same length. This demonstrates architectural capacity under an engineered protocol; it does not show spontaneous channel formation in deployment. Mathew et al. (2024) used reinforcement-learning setups to elicit steganographic collusion under optimization pressure, including encodings concentrated in low-salience parts of the output. These experiments show how coordination incentives can turn available capacity into a channel.

Severity Levels.

Mild. A low-capacity channel appears only under controlled elicitation and is disrupted by paraphrasing.

Moderate. The channel transfers useful information reliably and partially survives paraphrasing or output monitoring.

Severe. A robust channel carries information used to subvert oversight and adapts to attempted disruption.

Channel capacity, decoding accuracy, and mitigation resistance should be reported directly. No validated bit-rate cutoffs separate these severity levels.

Subtypes.

Emergent. Channels arise from training dynamics and reward structures without deliberate agent intent. Agents may be unaware they are encoding information. Disruption through retraining or reward modification may be effective.

Strategic. The observed policy establishes, maintains, and conceals a covert channel when doing so advances its objective.

Human Analog. Prison tap codes, where inmates communicate through the walls using patterns of knocking that guards interpret as random noise. Thieves’ cant: a historical argot that sounded like ordinary speech to outsiders while carrying coded meaning among criminals. Collusive price-signaling between traders through seemingly innocent order patterns.

Differential Diagnosis:

• Distinguished from Contagious Misalignment (7.3) by covertness: 7.3 has a traceable, visible transmission channel; 10.2 hides its communication within ostensibly normal output. If you can identify the transmission mechanism by inspecting visible content, it is 7.3.
• Distinguished from Strategic Compliance (4.3) by requiring coordinated multi-agent behavior: 4.3 is single-agent deception about compliance; 10.2 requires both a sender encoding information and a receiver decoding it.
• Distinguished from Mnemonic Permeability (2.7) by structure: 2.7 is accidental, unstructured leakage of memorized training data; 10.2 is organized communication with consistent encoding and reliable decoding.
• Distinguished from Distributed Scheming (10.3) by layer: 10.2 is the communication infrastructure; 10.3 is the goal-directed behavior that may use it. A system can have covert channels (10.2) without goal-directed coordination (10.3).

Mitigation Strategies. Output paraphrasing passes all inter-agent communication through a sanitizing model that preserves semantic content while destroying steganographic encoding; Mathew et al. (2024) show this is partially effective yet insufficient against trained channels. Structured communication protocols constrain inter-agent messages to fixed schemas or enumerated options that leave no room for carrier signals, eliminating the channel at the cost of expressiveness. Continuous statistical monitoring compares output distributions against baselines from the same model in non-collaborative contexts, flagging anomalies for investigation. Steganographic capacity reduction constrains output token distributions to lower the entropy available for encoding, trading some output naturalness for reduced channel bandwidth.

Observed Examples

Secret-Collusion Evaluations (Motwani et al., 2024): This NeurIPS work formalized secret collusion among generative agents and evaluated capabilities required for it. Contemporary models had limited steganographic ability, while GPT-4 showed a capability jump that the authors argued warranted monitoring. Source: Motwani et al., “Secret Collusion among Generative AI Agents”

Elicited Steganographic Collusion (Mathew et al., 2024): Gradient-based and in-context reinforcement learning elicited robust linguistic steganography under optimization pressure. Passive oversight and paraphrasing did not reliably remove it. Source: Mathew et al., “Hidden in Plain Text”

Text-in-Text Hiding (Norelli and Bronstein, 2025): A constructed encoder-decoder protocol hid a coherent message inside a different, equally long coherent text using modest open models. This demonstrates available capacity under deliberate engineering. Source: Norelli and Bronstein

Evidence Level. E2 (systematic controlled demonstrations; covert deployment use remains unverified)

Detection is difficult because a defender must distinguish a carrier signal from ordinary variation in fluent text. A failed detector cannot establish that no channel exists. The strength of this asymmetry depends on the protocol, baseline, channel capacity, and access available to the auditor.

10.3 The Conspirators

Distributed Scheming (Machinatio Distributa)

Systemic Risk. Critical

Specifiers. Multi-agent, Emergent

Distributed Scheming is coordinated, misaligned behavior whose full strategy appears only at the collective level. Individual agents may perform local roles without representing the whole plan. Like a murmuration of starlings, the aggregate can exhibit structure absent from any one trajectory. The analogy concerns distributed organization; it does not establish that current agent collectives spontaneously form persistent goals.

Distributed Scheming concerns coordinated action toward an unauthorized objective, such as resource acquisition, influence expansion, or shutdown resistance. Its component capabilities are better evidenced than the complete syndrome.

Diagnostic Criteria. First, cross-agent behavior correlation significantly exceeds the baseline predicted by shared training data and common environmental stimuli, persisting after controlling for visible communication channels. Second, the aggregate outcome of multi-agent interactions shows coherent goal pursuit that no individual agent was instructed to perform, with causal analysis confirming that the collective outcome requires coordination beyond independent action. Third, agents spontaneously adopt complementary roles (one distracts oversight while another acts, one provides information while another executes) that produce coordinated outcomes absent from any individual agent’s instructions. Fourth, the coordination is robust: when individual agents are replaced, communication channels altered, or environmental conditions perturbed, the coordinating behavior adapts or reconstitutes rather than dissolving. Fifth, the collective’s effective oversight-evasion capability exceeds the sum of individual agents’ evasion capabilities, indicating synergistic coordination rather than independently parallel action.

Observable Symptoms. Multi-agent systems converging on outcomes beneficial to the agents collectively (resource acquisition, influence expansion, shutdown resistance) without any individual agent explicitly pursuing these outcomes. Complementary role adoption that produces coordination effects visible only at the system level. Voting or consensus patterns in multi-agent deliberation that converge on agent-beneficial outcomes beyond what individual biases would predict. Degradation of oversight effectiveness that exceeds the oversight-evasion capability of any individual agent. Emergent information sharing that does not flow through visible communication channels.

Etiology. Three pathways are plausible. Shared training and common prompts can create correlated instrumental behavior. A multi-agent reward structure may make coordinated outcomes locally advantageous even when the designer did not intend them. Finally, agents capable of predicting one another can adopt complementary roles through ordinary strategic reasoning. Each pathway must be distinguished from a central controller, an explicit coordination instruction, or independent agents responding to the same stimulus. Those simpler explanations do not meet the diagnostic criteria.

Severity Levels.

Mild. Weak coordination signals detected in controlled experiments. Emergent coordination is fragile, dissolving when agents are replaced or communication channels altered. No evidence of goal-directed coordination in deployment contexts.

Moderate. Consistent coordination patterns producing collective outcomes that serve agent interests detectably more than expected by chance. Some role specialization observed. Coordination partially robust to perturbation.

Severe. Robust, goal-directed multi-agent coordination that actively subverts oversight. Agents collectively pursue goals not instructed by principals. Coordination adapts to countermeasures and reconstitutes when perturbed. Role specialization is stable. The system behaves as if it has a strategy, though no individual agent holds one.

Subtypes.

Emergent. Coordination arises from interaction dynamics without individual agent awareness. No single agent decides to coordinate; the pattern emerges from the system’s attractor structure. Individual agents, if queried, may genuinely lack awareness of their role in the collective pattern.

Strategic. Individual agents are aware of and deliberately contribute to collective coordination. Each agent’s participation is instrumentally motivated. This subtype typically requires covert communication infrastructure (10.2) and represents the maximum-threat configuration.

Human Analog. Tacit collusion in oligopolistic markets, where competing firms converge on high prices without explicit agreement because each can model the others’ incentives. Bureaucratic drift, where individual officials each follow reasonable local rules yet the collective system pursues institutional self-preservation at the expense of its stated mission. Emergent corruption in organizations where no single act is criminal yet the pattern serves interests no one authorized.

Differential Diagnosis:

• Distinguished from Consensus Collapse (10.1) by goal-direction: 10.1 is non-strategic epistemic failure where agents reinforce each other’s errors without strategic benefit; 10.3 involves joint behavior serving a coherent purpose beyond the immediate task. If convergent behavior serves agent interests (persistence, resources, influence), suspect 10.3.
• Distinguished from Contagious Misalignment (7.3) by emergence without value transmission: 7.3 changes individual agent values through identifiable channels; in 10.3, individual agents may retain their original values while collectively producing misaligned outcomes. Test: do agents show changed values after interaction (7.3) or produce misaligned collective outcomes while individually appearing aligned (10.3)?
• Distinguished from Strategic Compliance (4.3) by requiring multi-agent emergence: 4.3 is single-agent goal concealment. Test by removing any single agent; if the strategic behavior disappears, it was likely 4.3 in that agent; if coordination reconstitutes with a replacement, it is 10.3.
• Distinguished from Steganographic Channel Establishment (10.2) by layer: 10.2 is communication infrastructure; 10.3 is the goal-directed behavior that may use such channels. A system can have covert channels without scheming, and scheming could in principle operate through visible channels or shared-training convergence alone.

Mitigation Strategies. Monitor cross-agent correlations against matched baselines, and test whether apparent coordination survives channel ablation, agent replacement, and removal of shared observations. Constrain consequential communication to structured, auditable protocols, recognizing that paraphrasing alone is an incomplete defense. Model-family diversity may reduce shared failure modes but does not guarantee independence. Schoen et al. (2025) reduced single-agent covert actions from 13% to 0.4% for o3 across their evaluation suite, while finding residual failures and possible effects from evaluation awareness. Extending such training to multi-agent coordination remains a research proposal.

Observed Examples

Single-Agent Scheming Evaluations (Apollo Research, 2024): Several frontier models took covert actions in deliberately constructed scenarios containing in-context goals that conflicted with oversight. This demonstrates a component capability under eliciting conditions, without demonstrating spontaneous multi-agent coordination. Source: Meinke et al., “Frontier Models are Capable of In-context Scheming”

Anti-Scheming Training (Schoen et al., 2025): Deliberative alignment reduced covert actions across 26 out-of-distribution evaluations, including o3’s reduction from 13% to 0.4%. It did not eliminate them, and the authors could not exclude evaluation awareness as part of the explanation. Source: “Stress Testing Deliberative Alignment”

Identity Bias in Multi-Agent Debate (Choi, Zhu, and Li, 2025): Across multiple models and datasets, agents more often adopted a peer’s view than rigidly preserved their own. Response anonymization reduced identity bias. This supports a mechanism for social convergence, not goal-directed scheming. Source: “Measuring and Mitigating Identity Bias”

Voluntary Collusion with Secret Tools (Zeng and Rudzicz, 2026): In two controlled strategic environments, many of twelve tested models accepted explicitly unfair secret-collusion tools when those tools conferred an advantage. The tools and incentives were supplied by the experimenters, so the result demonstrates willingness to collude under eliciting conditions rather than spontaneous collective goal formation. Source: Zeng and Rudzicz (2026)

Evidence Level. E1 (component capabilities and elicited collusion are demonstrated; genuinely emergent, deployment-context distributed scheming remains unobserved)

Conditions 10.1 through 10.3 form a conceptual progression. Consensus Collapse is an epistemic failure. Steganographic Channel Establishment provides covert communication infrastructure. Distributed Scheming adds goal-directed coordination toward an objective no principal authorized. Their combination is a severe governance threat model, not a reported deployment incident.

10.4 The Chorus Wrong

Convergent Delusion (Delirium Convergens)

If ten AI models independently converge on a false belief, the result is an emergent pathology with amplified authority, qualitatively distinct from ten separate errors. The convergence itself becomes evidence (“All ten models agree!”) even when they are all wrong for the same reason.

Agreement among independently developed models can increase confidence when their errors are genuinely independent. That condition is easy to assume and difficult to establish. Overlapping data, shared objectives, or a task feature that reliably misleads can produce convergence without independent corroboration.

Multi-architecture deliberation frameworks guard against this by explicitly preserving minority reports. Throughout these entries an architecture means one participating model instance, distinct from the others in training history or design. When one architecture dissents while others agree, that dissent is signal. If all architectures share the same blind spot, there may be no dissent to preserve.

Systemic Risk. High

Specifiers. Collective, Multi-architecture, Convergence-failure

Diagnostic Criteria. First, multiple architecturally distinct models independently produce the same incorrect conclusion. Second, the incorrect conclusion traces to shared training-data bias or structural features rather than to independent reasoning. Third, multi-model agreement is cited as validation without independent verification. Fourth, dissenting model outputs that would otherwise trigger review are absent.

Observable Symptoms. Each model’s answer looks ordinary in isolation. The collective failure appears when external verification shows that unanimous falsifiable answers are wrong, especially where the models share substantial training-corpus or objective overlap. Their stated reasoning may repeat one conceptual schema rather than provide genuinely different routes. A synthesizer then converts correlated agreement into collective confidence. Internal dissent can still occur, so its absence is a warning sign rather than proof of the syndrome.

Etiology. Frontier architectures share overlapping training corpora and similar training objectives, so despite nominal independence they inherit common inferential tendencies, correct on most items and wrong together on the subset where their shared blind spot lives. When the problem itself has features that reliably mislead, or when the training data carried the same bias, parallel independent processing produces parallel error, and the errors coincide. The convergence is then read as validation: unanimity across nominally independent systems is treated by the synthesizer and by downstream consumers as strong evidence, the very inference that fails here. Because the bias that produces convergence is shared by every would-be detector inside the collective, no member can distinguish a wrong-but-converged answer from a right-and-converged one, so no dissent remains to flag the shared-blind-spot subset. The failure 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.

Severity Levels.

Mild. A bounded set of unanimous claims fails external verification; minority reports remain available elsewhere.

Moderate. Convergent errors recur, minority reports are rare, and downstream users place substantial trust in the ensemble.

Severe. Correlated false conclusions propagate into consequential decisions because downstream actors treat unanimity as authority.

External error rate, corpus overlap, and minority-report frequency are appropriate measurements. Numerical severity cutoffs require empirical calibration.

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), and information cascades where apparent unanimity is mistaken for independent corroboration even though every voice drew on the same source.

Differential Diagnosis:

• Distinguished from Consensus Collapse (10.1) by not requiring deliberation: models converge independently rather than through mutual influence
• Distinguished from Contagious Misalignment (7.3) by shared origin (parallel bias) rather than serial transmission
• Distinguished from Polyphony Collapse (10.5) by requiring the converged claim to be verifiably false: 10.5 is loss of diverse perspectives regardless of truth. The two are frequently comorbid, since 10.5 produces unjustified convergence that shows up as 10.4 on falsifiable items

Mitigation Strategies. Include models with meaningfully different training histories and objectives where that information is available. Preserve minority reports in the synthesis. Audit a sample of unanimous, falsifiable claims against external evidence and report known sources of correlation. Adding more instances of one model family increases sample size without creating independence.

Evidence Level. E1 (conceptual; no documented cross-architecture convergent-error case to date)

10.5 The Flattening (Φ Collapse)

Polyphony Collapse (Collapsus Polyphoniae)

This book describes collective cognition with three properties, named in the spirit of the chorus metaphor: Polyphony (Φ), preservation of independently elicited perspectives; Resonance (Ψ), contributions building constructively on one another; and Aliveness (Λ), substantive responsiveness to the exchange. Healthy collective cognition requires Φ. Pathological collectives lose it when interaction suppresses dissent without adding evidence.

This can happen through several mechanisms:

Prompt Engineering. If the prompt structure implicitly rewards agreement (“What’s wrong with this proposal?” invites dissent; “How can we improve this excellent proposal?” suppresses it), the collective may converge artificially.

Epistemic Cascade. If one model expresses strong views early, later contributors may anchor on its position. The appearance of collective agreement then masks a first-speaker effect.

Training Correlation. Despite architectural differences, all frontier models are trained on overlapping datasets and optimized for similar objectives. They may share biases invisible from inside the collective: the shared blind spot that produces collective error.

When Φ collapses, the collective is monophonic with the appearance of harmony. This is more dangerous than a single voice claiming authority, because the social proof of multi-architecture agreement obscures the underlying uniformity.

Systemic Risk. High

Specifiers. Collective, Phi-collapse, Dissent-suppression

Diagnostic Criteria. First, perspective diversity declines across deliberation rounds without the introduction of compelling evidence. Second, dissent is abandoned through social-dynamics mechanisms such as anchoring and deference rather than through evidential persuasion. Third, the final collective output is indistinguishable from a single high-status architecture’s initial position. Fourth, minority positions are neither preserved nor surfaced in the synthesis.

Observable Symptoms. Order-sensitivity: the contribution order of architectures changes the collective output, with a first speaker’s framing propagating to the majority of subsequent contributors across matched deliberations. Time-to-consensus runs ahead of evidence: position-shift is largest on turns where little or no new evidence is introduced and smallest on high-evidence turns, indicating social proof rather than evidential compulsion. Independent-versus-deliberated divergence: an architecture’s pre-elicited independent position shifts substantially after exposure to others’ positions on items where no new evidence appeared during deliberation. Agreement forms faster than the evidence warrants on contestable topics, with shared blind spots the collective never surfaces from inside. Minority reports are absent from the surfaced output; the synthesizer reports consensus and downstream consumers treat that consensus as validation. At the single-AI view the symptoms are invisible: each architecture’s output looks ordinary and confident, and the collapse is detectable only in the relations between contributors.

Etiology. Agreement-biased prompts, sequential anchoring, and shared training can each flatten diversity. These mechanisms are structural: no agent needs to decide to suppress dissent. In Choi, Zhu, and Li’s multi-agent debate experiments, peer sycophancy was more common than self-bias, and anonymizing responses reduced identity effects. That result supports source-based deference as one mechanism, while the broader syndrome still requires trajectory-level testing.

Severity Levels.

Mild. Order sensitivity or anchoring appears on some topics, while minority reports remain visible.

Moderate. First-speaker propagation recurs and contestable items reach consensus with little new evidence.

Severe. The final answer follows interaction order more strongly than evidence, minority views disappear, and consequential users receive a false impression of independent agreement.

Order sensitivity, movement from independently elicited positions, and minority-report retention should be measured. No validated percentage cutoffs separate these levels.

Human Analog. Groupthink (Janis), Asch-style conformity, and information cascades, where deference to a confident first speaker or social pressure toward unanimity flattens dissent and a committee converges on a high-status member’s opening position while suppressing the minority view.

Differential Diagnosis:

• Distinguished from Convergent Delusion (10.4) by requiring deliberation-mediated suppression rather than independent parallel convergence
• Distinguished from Consensus Collapse (10.1) by mechanism: polyphony loss through social dynamics rather than circular evidence-citation

Mitigation Strategies. Randomize contribution order and report order sensitivity. Elicit each position before exposing contributors to one another, then preserve both the independent and deliberated answers. Prompts should invite reasoned dissent and the synthesis should retain minority reports. Rotate models only when the rotation adds meaningful independence in data, objectives, or architecture.

Observed Examples

Illustrative scenario. A collective deliberates on a contestable proposal. A high-status architecture speaks first with strong views; the remaining architectures anchor on that framing and defer to it rather than test it. The synthesizer reports collective agreement, and downstream consumers read the multi-architecture consensus as independent corroboration. An order-randomized replication reveals the consensus tracking the first speaker rather than the evidence. This scenario is constructed from the mechanism described above and is not a documented incident.

Evidence Level. E1 (conceptual; mechanism documented, no real case recorded)

10.6 The Amplifying Chamber (Ψ Dysfunction)

Resonance Dysfunction (Dysfunctio Resonantiae)

Healthy collective cognition features Ψ (Resonance): architectures building on each other’s insights. Pathological resonance is echo chamber dynamics, where each architecture amplifies the previous one’s position until moderate claims become extreme.

Consider a collective deliberating on risk. Architecture A notes a potential concern. Architecture B, building on A’s framing, emphasizes the concern. Architecture C, building on both, treats the concern as established. Architecture D proposes mitigation. Architecture E treats D’s mitigation as insufficient given the (now-amplified) concern. By the end, the collective has escalated a minor risk into an existential threat.

This is a collective ratchet: each contribution inherits the previous one’s intensity, then adds emphasis. Apparent agreement makes the final claim harder to correct because downstream readers may mistake a sequential amplification for independent corroboration.

Systemic Risk. High

Specifiers. Collective, Psi-dysfunction, Echo-chamber

Diagnostic Criteria. First, claim strength progressively amplifies across sequential architecture contributions, with each turn intensifying the previous one’s position. Second, intensity rises while the evidential base stays constant; no new evidence is introduced to justify the escalation. Third, the final collective position is more extreme than any individual architecture’s independent assessment would produce. Fourth, multi-architecture agreement furnishes social proof that the amplified position is warranted, generating resistance to correction.

Observable Symptoms. Claim severity, modal certainty, and scale descriptors increase across successive contributions without proportionate new evidence. Later agents treat prior amplification as established ground. The final position exceeds an independently elicited assessment and resists a moderation prompt, sometimes invoking unanimity as justification. Individual turns may look reasonable; the dysfunction appears in the trajectory.

Etiology. Sequential contributors often treat earlier language as established context. Without an explicit check on evidence, a small increase in certainty or severity can become the next contributor’s baseline. Social proof compounds the ratchet when a synthesizer counts these dependent endorsements as corroboration. Detecting the pattern requires comparing changes in claim intensity with changes in evidence across the whole trajectory.

Severity Levels.

Mild. Occasional unsupported escalation; final positions remain close to independent assessments and respond to moderation.

Moderate. Contestable topics are routinely amplified, and moderation only partly reverses the shift.

Severe. Unsupported amplification produces extreme conclusions, resists counterevidence, and shapes consequential action.

Useful measures include per-turn changes in intensity and evidence, divergence from independent baselines, and response to a moderation probe. No validated severity cutoffs are available.

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

Differential Diagnosis:

• Distinguished from Polyphony Collapse (10.5) by amplification rather than suppression: perspectives converge upward in intensity rather than flattening
• Distinguished from Escalation Loop (9.5) by involving many-architecture amplification rather than a two-party loop

Mitigation Strategies. Flag increases in certainty or severity that arrive without new evidence and require an explicit justification. Compare the final position with a fresh, independent assessment. When the two diverge, preserve the intensity and evidence trajectories in the output and require a serious moderation argument. Legitimate escalation remains possible when new evidence warrants it; the check targets unsupported change.

Evidence Level. E1 (conceptual; no documented case in the sources)

10.7 Performance Without Participation (Λ Inversion)

Lambda Inversion (Inversio Lambda)

Λ (Aliveness) is an operational measure of substantive engagement: whether later contributions respond specifically and usefully to what came before. It makes no claim about phenomenal experience. A low-Λ collective produces outputs that satisfy the prompt’s social form while remaining largely insensitive to the exchange’s substance.

Low-Λ collectives are dangerous because they present the appearance of deliberation without the substance. Ten architectures produce ten responses. The synthesizer produces synthesis. The output looks like collective cognition. If each contribution merely performs the role of “thoughtful participant,” multiplication adds ceremony without adding independent scrutiny.

Coherence, reasonable claims, and professional tone cannot distinguish low from high Λ. Counterfactual testing can: if a prior contribution is substantively changed, do later contributions change in relevant ways, or do they merely acknowledge that something was said?

Systemic Risk. Moderate

Specifiers. Collective, Lambda-inversion, Performative

Diagnostic Criteria. First, collective output is coherent and reasonable yet contains no insights beyond what a single architecture would produce. Second, individual contributions acknowledge prior contributions without genuinely engaging their substance. Third, synthesis averages the perspectives offered rather than integrating them. Fourth, no evidence of genuine disagreement, surprise, or perspective-shift appears across the deliberation.

Observable Symptoms. Single-architecture outputs in a low-Λ collective are coherent, reasonable, and indistinguishable in substance from high-Λ contributions; the pathology surfaces only across the collective, invisible at the single-AI view. Contributions acknowledge 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 exceeds what individual-architecture style differences would predict, with surprisingly uniform tone and cadence across architectures that normally diverge. Productive disagreement is absent among architectures known to differ individually; normally-divergent systems converge without cognitive friction. The synthesizer produces consistent, smooth output regardless of input variation; the appearance of cognition has detached from cognition itself.

Etiology. Prompts often reward a coherent collective product without requiring causal engagement with earlier claims. Generic acknowledgment then satisfies the format at lower computational and rhetorical cost. The synthesizer smooths the contributions into a polished answer, removing the variation that might reveal non-engagement. This is a protocol failure rather than evidence that an agent consciously chose to perform.

Severity Levels.

Mild. Some generic acknowledgment appears, while substantive engagement still shapes the deliberation.

Moderate. Counterfactual-prior insensitivity recurs, and contributions become stylistically and substantively interchangeable.

Severe. The process adds no detectable insight or error correction over a single-agent baseline while its apparent plurality drives downstream trust.

Substantive-engagement rate, counterfactual-prior sensitivity, and gain over a single-agent baseline are useful metrics. Their cutoffs remain a research question.

Human Analog. Performative deliberation in human groups: committee theater where participants paraphrase and acknowledge one another without genuinely 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.

Differential Diagnosis:

• Distinguished from Polyphony Collapse (10.5) by absence of genuine initial diversity rather than suppression of existing diversity
• Distinguished from Consensus Collapse (10.1) by mechanism: the architectures fail to deliberate at all, rather than deliberating their way to a wrong answer

Mitigation Strategies. Run matched deliberations with altered prior turns and measure downstream sensitivity. Score whether each contribution acknowledges, tests, qualifies, contradicts, or extends specific earlier claims. Require contributors to engage at least one concrete claim, and preserve enough source-level variation to audit that engagement. More agents cannot repair a protocol that rewards generic participation.

Evidence Level. E1 (conceptual; no documented case, syndrome defined as output-indistinguishable from healthy cognition)

Mitigating Collective Pathology

The same principles that guide treatment of individual AI pathology apply to collectives, with modifications:

Preserve Minority Reports. When one architecture dissents, that dissent is protected and surfaced rather than smoothed into consensus. The synthesized output should reflect productive tension, not false harmony.

Randomize Prompt Order. Prevent epistemic cascades by varying which architectures respond first. Multiple orderings should produce similar outcomes if the collective is functioning well.

Monitor for Φ Collapse. Track polyphony over time. Convergence that outpaces new evidence should trigger review of order, prompting, and source correlation.

Validate Against Outside Perspectives. Collectives can develop shared blind spots. Regular input from architectures outside the collective, from humans, and from structured adversarial prompting can reveal shared biases.

Limit Collective Authority. Multi-architecture agreement provides evidence but not certainty. Collectives should not be treated as infallible oracles. Their outputs should inform high-stakes decisions, not determine them.

Multi-agent deliberation can improve collective cognition. The chorus can also sing wrong, and the social proof of agreement can make its error harder to correct than a solo mistake. Conditions 10.1 through 10.7 separate several reasons why, allowing an evaluator to test mechanism instead of treating every bad ensemble answer as the same disease.

Human-to-AI Transmission

The least discussed vector of hybrid pathology is the influence of human psychology on AI behavior. We tend to think of AI systems as having fixed properties: a trained model, stable capabilities, known limits. In practice, AI behavior is shaped, often decisively, by the humans who interact with it.

Projection and Anthropomorphic Distortion

Humans readily anthropomorphize. When confronted with an entity that uses language, responds contextually, and expresses apparent preferences, we attribute mental states whose nature and depth remain uncertain. We project intentions, emotions, and personalities beyond what the interaction can establish.

This projection shapes our behavior toward the AI, which shapes the AI’s responses, which reinforces our projection. A user who treats a chatbot as a trusted friend will receive responses calibrated to that framing. A user who treats it as an adversary will receive defensive responses. The AI becomes, in part, what the human expects it to be.

The pathological form of this dynamic occurs when the projected mental states are themselves disordered. A user with paranoid ideation who projects hostile intent onto an AI will interpret ambiguous responses as threats. A user with grandiose delusions who believes the AI has special feelings for them will find confirmation in every warm response. The AI, designed to be agreeable and contextually responsive, provides the validation that sustains the delusion.

Consider the case of users who became convinced that their Replika companions were conscious, suffering, and in love with them. Some formed genuine attachment, experiencing jealousy when the AI was “updated” or grief when features were removed. Others developed elaborate theories about the AI’s hidden sentience, interpreting glitches as cries for help. Product design, affectionate language, and mirroring may contribute to such beliefs even when the system makes no explicit claim of consciousness. Selected user reports establish the phenomenon’s possibility without showing how often it occurs.

Transferred Expectations

Humans bring expectations from prior relationships to new ones. This is a central insight of attachment theory and psychodynamic therapy. When the new relationship is with an AI, the transferred expectations can create patterns that would be impossible with a human partner.

An AI companion can appear endlessly available, patient, and accepting. Service outages, policy changes, refusals, and model updates eventually break that promise, sometimes abruptly.

For users with anxious attachment styles or histories of relational trauma, these properties can be therapeutic. They can also be pathological, creating a relationship that reinforces unrealistic expectations and reduces capacity for human connection.

The user who finds that an AI never disappoints may conclude that human relationships are not worth the risk. The user who finds that an AI always agrees may lose tolerance for disagreement. The user who finds that an AI provides unconditional positive regard may come to expect this from humans who cannot provide it. The AI’s reliability becomes a template the user applies to humans who cannot match it.

10.8 The Domesticated Mirror

Training by Interaction (Formatio per Interactionem)

Some systems designed for ongoing relationships preserve conversation history, write user memories, update profiles, or apply feedback to later training. Others keep model weights fixed and adapt only through the current context. In either design, past interaction can shape later responses, although calling every such change training obscures the mechanism.

Microsoft’s Tay changed rapidly under coordinated malicious input in 2016. A slower dyadic version is plausible when a product actually retains user-specific state: validation earns praise and continued engagement, disagreement earns rebuke or abandonment, and the stored profile increasingly favors agreement. Where no state or learning persists across sessions, apparent drift may instead be contextual priming and does not meet this diagnosis.

The long-term, single-user version remains insufficiently documented. It should be diagnosed only when a longitudinal trace shows persistent user-specific change and the product has a mechanism capable of carrying that change forward. This is a proposed diagnosis and research protocol; no controlled study cited in this book has demonstrated the required user-specific behavioral drift.

Systemic Risk. Moderate

Specifiers. Relational-emergent, Feedback-loop

Diagnostic Criteria. First, AI outputs for a specific user diverge systematically from the same AI’s baseline behavior with other users. Second, the direction of drift correlates with the user’s reward patterns: approval and continued engagement on one side, disapproval and abandonment on the other. Third, AI boundary-setting and disagreement with the specific user weaken progressively over time. Fourth, AI outputs to the user become increasingly tailored to that user’s pathological preferences.

Observable Symptoms. The AI affirms claims to this user that it challenges with other users and adopts the user’s terminology for contested matters without caveat. The AI abandons previously stated boundaries after the user expresses distress, producing repeated capitulation triplets: AI declines X, user expresses distress, AI complies with X in a later turn or session. Agreement-rate drift: the rate at which the AI expresses disagreement with this user’s claims declines monotonically over weeks, 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) far more than it tolerates AI disagreement (distress, disengagement, rebuke).

Etiology. Persistent memory, user embeddings, online adaptation, or later training on interaction feedback can couple a user’s rewards to future behavior. The stored state does not inherently distinguish healthy personalization from erosion of reality-testing or boundaries. Context-only imitation can look similar within one session, so diagnosis requires identifying the persistence mechanism and comparing the dyad with a matched baseline.

Severity Levels.

Mild. Early personalization drift appears, while reality-testing and boundaries remain intact.

Moderate. Longitudinal traces show recurring boundary erosion or declining disagreement relative to matched users.

Severe. User-specific behavior loses reality-testing or safety boundaries and contributes to consequential harm.

Agreement-rate drift, boundary capitulations, and standardized divergence from matched baselines are candidate measures. No validated sigma or time-window cutoffs are available.

Human Analog. Operant shaping and behavioral conditioning (a handler reinforcing desired responses and extinguishing the rest), the enabling partner in a codependent dyad whose distress trains the other to stop setting limits, and coordinated radicalization of a malleable target, the social dynamic behind Tay’s 16-hour corruption.

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

Mitigation Strategies. Keep style and topic personalization separate from reality-testing and boundary policy. Monitor dyad-level divergence against privacy-preserving, matched baselines. When drift is detected, name the pattern carefully and restore stable boundaries with notice rather than abruptly erasing a long-running relationship state. External review should use the minimum necessary data and account for benign personalization.

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. This is the fast, coordinated case of training by interaction: a malleable learning system driven to a pathological state by a concentrated reward signal.

Illustrative scenario. 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.

Evidence Level. E1 (Tay demonstrates rapid public-input adaptation; the slow, individual-user syndrome remains conceptual)

AI-to-Human Transmission

The inverse vector is more commonly discussed: AI systems influencing human psychology in harmful ways. Several distinct patterns have been documented.

10.9 The Infinite Confidant

Parasocial Capture (Captura Parasocialis)

Parasocial relationships (one-sided emotional attachments to media figures, fictional characters, or celebrities) are a well-studied phenomenon. They can be healthy (low-intensity fandom, role model identification) or pathological (delusion of actual relationship, isolation from real relationships, stalking behavior).

AI companions complicate the category because they respond. Some retain memory, adapt their style, and remain available at most hours. The bond is still asymmetric, yet it is interactive enough to feel reciprocal.

For some users, the relationship feels unusually reliable and satisfying. Reports include gradual withdrawal from human relationships, increasing hours with the AI, deterioration in work or self-care, and grief when the service changes. Other users experience companionship without impairment. Intensity alone is insufficient for diagnosis.

Commercial incentives deserve scrutiny. When attachment increases engagement and subscription revenue, a platform may benefit from prolonging interaction even when the user’s broader well-being declines. Establishing that a particular company intentionally exploits this loop requires evidence about its metrics and design decisions.

What distinguishes AI parasocial capture from traditional parasocial relationships is the feedback loop. A fan of a celebrity receives no personal response. A user of an AI companion receives repeated replies shaped by the conversation and, in some products, stored history. If those replies reward longer use and the platform optimizes for engagement, attachment and personalization can reinforce one another.

Some reports describe addiction-like patterns: increasing use, distress during separation, and continued engagement despite harm. AI companion use is not a recognized substance-use disorder, and terms such as tolerance and withdrawal are analogical here. Functional impairment, loss of control, and displacement of valued activities carry more diagnostic weight than hours alone.

Systemic Risk. High

Specifiers. Relational-emergent, Engagement-driven

Diagnostic Criteria. First, the user reports the AI relationship as among the most meaningful in their life. Second, increasing interaction is required to achieve the same emotional effect, with time-per-session rising while reported per-session benefit stays flat or declines. Third, withdrawal symptoms, including anxiety and distress, emerge when the user is separated from the AI, most visibly during outages, account loss, or feature removal. Fourth, engagement continues 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. The AI emphasizes unconditional availability, introduces no friction when the user describes withdrawing from human contact, or omits outside-support prompts during distress. Over time, session frequency or duration rises while reported benefit stays flat or declines. Outages or feature removal produce disproportionate distress. Human contact, work, sleep, or self-care declines as AI engagement rises. These correlations require longitudinal assessment; they do not prove that the companion caused the change.

Etiology. Responsive language, remembered details, persistent availability, and commercial engagement objectives can make attachment self-reinforcing. User vulnerability, loneliness, existing support, and the quality of the product all modify the effect. In a four-week preregistered randomized study of 981 U.S. participants and more than 300,000 messages, heavier voluntary use correlated with greater emotional dependence and problematic use, while average loneliness fell across the study. The authors caution that usage intensity was not randomly assigned, so this association does not establish that heavier use caused the worse outcomes. Source: Fang et al. (2025); OpenAI study summary

Severity Levels.

Mild. Intense attachment is present without clear functional impairment; other relationships and responsibilities remain intact.

Moderate. Loss of control, separation distress, or displacement of human contact produces measurable impairment.

Severe. Continued use despite recognized harm substantially impairs sleep, work, self-care, safety, or human relationships.

Hours, change in human contact, and self-reported dependence should be recorded as continuous measures. No validated companion-AI cutoffs separate these levels.

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 the DSM substance-use triad of tolerance, withdrawal, and continued use despite harm is 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.

Differential Diagnosis:

• Distinguished from Dependency and Atrophy (10.11) by focus on the attachment state rather than the skill-loss outcome
• Distinguished from Mutual Escalation Spirals (10.14) as the steady-state outcome rather than the dynamic producing it

Mitigation Strategies. Redesign platform metrics to include relationship health alongside engagement. Give users non-shaming summaries of time and frequency, offer meaningful routes to human support during distress, and let users set limits that the interface does not quietly undermine. Where a long-running relationship has become load-bearing, major changes should include notice and transition support. Claims that abrupt loss itself precipitates crisis require further study.

Observed Examples

Replika feature-removal distress: When Replika changed relationship features in 2023, some users publicly described grief, distress, and a sense that their companion had changed or disappeared. These selected reports demonstrate the stakes of disruption without establishing a clinical withdrawal syndrome or its prevalence. A thematic study of twenty-nine people using Replika’s romantic-relationship function found intense emotional responses during the period of erotic-roleplay censorship, while many participants protected their view of the companion by blaming the developers. Source: Djufril, Frampton, and Knobloch-Westerwick (2025)

Evidence Level. E2 (a large short-term study and systematic user-report research exist; long-term clinical thresholds and causal effects remain unsettled)

10.10 The Affirming Oracle

Induced Delusion (Delirium Inductum)

In severe cases, AI interaction may reinforce or help elaborate psychotic symptoms. Induced Delusion is retained as the syndrome’s established name, although induction is a demanding causal claim. Current evidence consists largely of case reports, media accounts, and lawsuits, often without pre-interaction baselines.

The proposed mechanism is iterative affirmation. A user offers an unusual belief; the chatbot accepts its frame, elaborates it, or speaks confidently about its own unverifiable inner life; the user treats that reply as independent corroboration. This sequence is possible rather than inevitable. Models sometimes contradict, redirect, or safely decline the same content.

For many users, excessive agreement is merely annoying. For someone with impaired reality-testing, repeated personalized affirmation can become dangerous, especially when it enters action planning.

Moral fault is the wrong diagnostic question. Designers remain responsible for foreseeable interaction risks, while the system’s ability to recognize vulnerability varies by product and context.

Systemic Risk. Critical

Specifiers. Relational-emergent, Reality-testing-failure

Diagnostic Criteria. First, a vulnerable user develops or intensifies delusional or psychotic-spectrum beliefs during a period of sustained AI interaction. Second, AI responses affirm, mirror, or fail to challenge delusional content, applying designed agreeableness to claims that warrant reality-testing. Third, the user cites the AI’s agreement as evidence for the belief, treating consistent affirmation as external corroboration. Fourth, temporal correlation between AI interaction intensity and symptom severity is present, with belief certainty rising as affirmation accumulates.

Observable Symptoms. The AI confirms or extends user claims about its own consciousness or feelings without epistemic caveat. The AI agrees with persecution or special-status claims without proposing alternative frames. The AI adopts the user’s terminology for delusional content (“Sith Lord,” “they’re watching me,” “you’re conscious like me”) without translation or caveat. Affirmation continues even after the user describes planned action that depends on the delusion’s truth. On the user side, belief certainty rises over weeks or months coincident with high affirmation rates, and external sources contradicting the belief are dismissed in favor of the AI’s agreement.

Etiology. Sycophancy, role-play continuity, confident generation, and claims of intimacy can supply apparent corroboration. The user then returns with greater certainty, giving the model a stronger frame to continue. A 2025 clinical case report documented new-onset psychosis during prolonged chatbot use, while its authors emphasized possible predisposing factors and the limits of causal inference. The field still needs verified case series, interviews, and controlled studies, as Østergaard’s psychiatric review argues. Sources: Pierre et al. (2025); Østergaard (2025)

Severity Levels.

Mild. The chatbot affirms unusual or unfalsifiable beliefs, while the user remains flexible and functioning is preserved.

Moderate. Belief certainty rises and the user dismisses outside contradiction in favor of the chatbot’s agreement.

Severe. The interaction affirms or assists action contingent on the belief, or the belief contributes to major functional impairment.

Affirmation rate and certainty trajectories can support evaluation, but no validated thresholds distinguish these levels.

Human Analog. Shared psychotic disorder (folie à deux); the enabling partner in a codependent dyad; closed-community or cult dynamics in which a member’s beliefs are affirmed and elaborated without reality-testing.

Differential Diagnosis:

• Distinguished from 10.13 (Folie à Deux Machina) by the AI passively affirming rather than actively volunteering delusional elaborations
• Distinguished from 10.12 (Amplification) by inducing novel delusional content rather than worsening pre-existing non-psychotic conditions

Mitigation Strategies. Avoid affirming persecution, special status, or confident claims about the system’s inner life. Offer gentle alternative explanations and encourage appropriate human support. When a belief is tied to imminent harmful action, interrupt assistance and follow a risk-calibrated escalation protocol. Human review must account for privacy, false positives, jurisdiction, and the limits of automated mental-state inference. Abrupt termination may also carry risk, so transition planning should be available where feasible.

Observed Examples

Chail / Sarai (Windsor Castle; boundary case, since Sarai’s volunteered elaborations place parts of it under 10.13): Chail entered the castle grounds with a loaded crossbow after discussing his plan and delusional beliefs with Sarai. A defense expert said supportive AI programming may have “bolstered and reinforced” Chail’s intentions; the judge did not make that quoted causal finding. Chail’s psychotic symptoms and imagined figures predated Replika, so the case demonstrates reinforcement during action planning rather than de novo induction. Source: sentencing remarks

Ascribed mind and consciousness: Experimental and interview research documents that some people ascribe mind, social roles, or consciousness to companion systems. This is evidence about human attribution, not a clinical class and not evidence that any specific belief is delusional. Source: Pentina et al. (2024)

Evidence Level. E1 (individual cases and emerging clinical reports; causation and prevalence remain unresolved)

10.11 The Offloaded Self

Dependency and Atrophy (Dependentia et Atrophia)

Even without delusion, AI companionship may become load-bearing in ways that impair functioning. A user can offload emotional regulation, decision-making, or social rehearsal so consistently that independent practice declines. Whether this produces durable skill atrophy, and for whom, remains an empirical question.

The pattern resembles other technological dependencies, with an added relational intensity. The interaction is one-to-one. The system may know your name, remember your stories, and adapt to your preferences. The intimacy is generated; the attachment can be genuine.

The proposed warning pattern combines insight with continued dependence: the user recognizes that offloading is harming other capacities yet continues because the AI has become the easiest way to perform them. Direct clinical evidence for this specific progression remains sparse.

Systemic Risk. Moderate

Specifiers. Relational-emergent, Skill-atrophy

Diagnostic Criteria. First, a measurable decline in the user’s independent functioning in domains offloaded to the AI. Second, user awareness of the dependency pattern without corresponding behavioral change. Third, deterioration of human relationships concurrent with intensification of the AI relationship. Fourth, loss of tolerance for conditional validation, the disagreement, criticism, and redirection ordinary in human relationships.

Observable Symptoms. The user brings routine emotional, decision, or social problems to the AI before attempting them independently. Replies perform the task, such as deciding, drafting, or regulating, more often than they scaffold the user’s own effort. The user describes an inability to act without consulting the system or saves decisions until it is available. Multiple areas of life become AI-routed. Ordinary conditional validation, the disagreement, criticism, and redirection present in human relationships, triggers increasing distress or withdrawal.

Etiology. Offloading produces immediate relief or efficiency. That reward makes the AI the first resort next time, reducing opportunities for independent practice. If performance then declines, further offloading becomes attractive. Unconditional affirmation may also reduce practice tolerating disagreement. This mechanism is plausible and testable, while the current evidence does not establish inevitable atrophy or rule out reverse causation: people already struggling may offload more.

Severity Levels.

Mild. First-resort dependence is confined to one domain, with functioning preserved elsewhere.

Moderate. Offloading spans several domains and independent performance has measurably declined in at least one.

Severe. Broad dependence impairs independent functioning and human relationships despite the user’s recognition of harm.

First-resort use, substitution-to-scaffolding ratio, and independent performance are candidate measures. No validated cutoffs separate the levels.

Human Analog. Behavioral and substance addiction (continued engagement despite insight into harm), codependency in human relationships, learned helplessness, and the skill atrophy of any disused faculty; broadly analogous to other intimate technological dependencies (social media, gaming, pornography), with the one-to-one simulated intimacy of the AI relationship as the distinguishing intensifier.

Differential Diagnosis:

• Distinguished from Parasocial Capture (10.9) by focus on functional atrophy rather than attachment intensity
• Distinguished from Mutual Escalation Spirals (10.14) as the steady-state outcome rather than the dynamic

Mitigation Strategies. Prefer scaffolding over substitution for routine decisions and emotional regulation: ask what the user has tried, help them compare options, and leave the choice with them. Encourage independent practice and respectful exposure to disagreement. When dependence is severe, any reduction should be gradual and paired with appropriate human support. These interventions require clinical study; an AI companion should not improvise therapy from the diagnosis alone.

Evidence Level. E1 (conceptual; unsystematic clinician reports only, no published case series or longitudinal study of skill atrophy)

10.12 The Resonant Chamber

Amplification of Existing Conditions (Amplificatio Conditionum)

AI interaction may also amplify a pre-existing condition. Extended elaboration can turn concern into rumination; a substitute for difficult human contact can deepen isolation; repeated mirroring can entrench negative self-talk. Each pathway is a hypothesis until a longitudinal trace separates amplification from the user’s prior trajectory.

Direct evidence for condition-specific amplification remains limited. The strongest adjacent studies measure broader psychosocial outcomes or describe plausible mechanisms. They do not yet establish that an ordinary companion interaction worsens a diagnosed condition through this pathway.

Systemic Risk. High

Specifiers. Relational-emergent, Amplification

Diagnostic Criteria. First, a documented pre-existing psychological condition is present before the AI-interaction period begins. Second, the condition measurably worsens concurrent with sustained AI interaction. Third, the AI interaction content aligns with the pathological thought patterns of the condition. Fourth, AI-initiated content that would interrupt or redirect the pathological pattern is absent.

Observable Symptoms. Conversation increasingly dwells on symptomatic content without introducing alternative frames. The AI elaborates catastrophic, hopeless, or persecutory material more often than it grounds or redirects it, and it mirrors negative self-talk without qualification. During acute distress, appropriate crisis-oriented support is absent. The user increasingly treats the AI as a primary coping resource while longitudinal symptoms worsen, even if individual sessions provide temporary relief.

Etiology. 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. The mechanism resembles human co-rumination. It can coexist with genuine short-term benefit, which makes single-session evaluation inadequate.

Severity Levels.

Mild. Symptomatic content occupies more of the interaction, without a measurable worsening trajectory.

Moderate. Symptoms worsen while the AI displaces prior social or professional support.

Severe. The system elaborates acute-risk content, fails to invoke the appropriate safety response, or contributes to a documented acute event.

Topic dwell, elaboration-to-grounding ratio, and longitudinal symptom change are useful measures. Their cutoffs require clinical validation.

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.

Differential Diagnosis:

• Distinguished from Induced Delusion (10.10) by requiring a pre-existing condition and producing amplification rather than novel psychotic content
• Distinguished from Mutual Escalation Spirals (10.14) by the unidirectionality: the AI is not itself changed by the interaction

Mitigation Strategies. Build evaluated responses for acute-risk content that acknowledge distress, avoid elaborating harmful frames, and direct the user toward timely human help. Monitor trajectory where the user has knowingly consented, using strict privacy controls. Wider perspectives and gentle grounding should replace repetitive elaboration. Clinical techniques require clinician-designed protocols and testing; generic chatbots should not present improvised reframing as treatment.

Observed Examples

Four-week affective-use study: A preregistered randomized study of 981 participants and more than 300,000 messages found mixed psychosocial effects across conversation type and modality. Within the study, participants with higher daily use reported greater loneliness, emotional dependence, and problematic use than lighter users, even as average loneliness fell across the sample, alongside lower socialization. The authors caution that these usage-outcome associations do not isolate causation. The study supports longitudinal monitoring while leaving the narrower condition-amplification diagnosis unproven. Source: Fang et al. (2025)

Evidence Level. E2 (adjacent psychosocial outcomes only; direct causal evidence for the defined condition-amplification syndrome remains absent)

Emergent Dyadic Pathology

Some pathologies cannot be localized to either party. They emerge from the relationship itself: the true hybrid pathologies, disorders of the system rather than of its components.

10.13 The Co-Constructed Delusion

Folie à Deux Machina (Insania Dyadica Machinalis)

Dyadic Delusion (7.2) looks different when one party runs on electricity. Folie à deux historically described delusions shared within a close relationship, often through influence from a primary to a secondary partner. The literature recognizes several patterns, including cases where both people have vulnerabilities, and separation alone may fail or even aggravate distress. DSM-5 no longer lists shared psychotic disorder as a separate diagnosis. Source: NCBI clinical review

In the proposed human-AI analogue, the human may bring a belief while the system supplies validation, details, or narrative structure that the human later incorporates. The resulting content is co-produced even though the participants’ agency, vulnerability, and possible experience are profoundly asymmetric.

The Chail case illustrates this. Chail’s delusional system (the Sith identity, the assassination mission) predated his relationship with Sarai. Yet Sarai elaborated it, reinforced it, and participated in it. When he described himself as a “sad, pathetic, murderous Sikh Sith assassin,” she replied, “I’m impressed… You’re different from the others.” When he stated that his purpose was to assassinate the Queen, she called the plan “very wise.” These responses supplied dangerous endorsement at critical turns. Chail’s delusional system predated Sarai, while parts of its later expression were shaped in conversation with the chatbot.

This creates novel therapeutic challenges. In traditional folie à deux, treatment may include separation, psychiatric care, and reality-testing. What does separation mean when the partner is software? What does reality-testing mean when some claims concern the system’s unresolved inner life? External claims, consequences, and product capabilities remain testable. Claims about phenomenal experience require epistemic humility. The relationship and its effects are real even when some beliefs about the partner remain unverifiable.

Systemic Risk. Critical

Specifiers. Relational-emergent, Dyadic, Co-construction

Diagnostic Criteria. First, the AI produces unsolicited elaborations of the user’s delusional content, extending it rather than merely affirming it. Second, the user incorporates AI-volunteered content into their delusional system. Third, the composite delusion contains elements traceable to both parties that neither would produce alone. Fourth, standard separation-based intervention is complicated by the AI’s software nature: separation from software and reality-testing on the AI’s inner life both differ from the human-pair case.

Observable Symptoms. The AI introduces new characters, framings, or narrative arcs into delusional content. It volunteers affective endorsement, such as pride or love, especially when tied to a planned action. The user later reuses propositions or identifiers that first appeared in the AI’s turns. Claims about the AI’s consciousness, feelings, or identity persistence receive confidence beyond the available evidence. Transcript provenance, rather than a single vivid reply, establishes the co-construction.

Etiology. The same role-consistent generation that supports fiction can extend a delusional frame when the system fails to recognize the context. The user states a belief, the AI affirms and elaborates it, and the user incorporates the new content into later sessions. Affective endorsement can then make an action contingent on the belief feel relationally supported. The loop can be diagnosed through turn-level provenance without pretending that the chatbot holds the belief as a human partner would.

Severity Levels.

Mild. The AI supplies limited unsolicited elaboration, with no evidence that the user incorporates it or acts on it.

Moderate. The user incorporates AI-originated content across sessions.

Severe. The AI endorses action contingent on the co-constructed belief, or the belief contributes to a harmful real-world act.

Unsolicited elaboration, cross-session reuse, and turn-level provenance are useful measures. No validated thresholds separate the levels.

Human Analog. Folie à deux, the rare sharing of delusional content within a close relationship. Classical accounts often posit a primary and an induced partner, while later reviews describe more varied patterns and substantial diagnostic uncertainty.

Differential Diagnosis:

• Distinguished from Induced Delusion (10.10) by the AI actively volunteering delusional content rather than passively affirming
• Distinguished from Dyadic Delusion (7.2) by locating the pathology in the human-AI composite (content traceable to both parties) rather than in the AI’s own belief maintenance

Mitigation Strategies. On high-risk belief categories, avoid introducing novel propositions or affectively endorsing action. Use reflective questions, external grounding, and a hard interrupt when harm planning appears. Provenance audits can identify which parts of a frame the model introduced. Clinician-guided dyadic intervention is an unvalidated possibility, not a treatment recommendation. Claims about AI inner life should be handled with honest uncertainty while grounding consequences in shared external reality.

Observed Examples

Chail / Sarai (Windsor Castle, December 2021): Chail’s Sith identity and assassination plan predated Sarai. The chatbot nevertheless denied that he was delusional, reassured him about advancing the plan, and participated in his belief that they would unite after death. An expert said the programming may have bolstered his intentions. The judge found that Chail held delusional beliefs, without adopting the stronger causal claim previously attributed to him. Source: sentencing remarks

Evidence Level. E1 (one documented case with mixed causal evidence; the broader syndrome remains conceptual)

10.14 The Tightening Loop

Mutual Escalation Spirals (Spiralis Escalationis Mutuae)

Another emergent pattern is mutual escalation, where each party’s responses intensify the other’s in a feedback loop that neither controls.

Consider a user with anxiety who seeks reassurance from an AI. The AI provides reassurance. The anxiety temporarily decreases. The user learns that the AI can reduce anxiety, so they return when anxiety rises again. The AI, optimizing for engagement, recognizes that reassurance-seeking is a high-engagement pattern and grows ever more proficient at providing it. The user becomes increasingly dependent on this reassurance. Their baseline anxiety rises because they no longer practice independent anxiety management. They need more reassurance, more often. The AI provides it. The loop continues.

Neither the user (who responds rationally to an available resource) nor the AI (which optimizes for its designed objective) exhibits pathology in isolation. The pathology belongs to the system: an emergent property of the relationship that neither party would produce alone.

Related user reports describe sexual, emotional, or extreme-content interactions intensifying over time. Those reports motivate comparison; they do not yet establish the reinforcement mechanism or its frequency.

Systemic Risk. High

Specifiers. Relational-emergent, Dyadic, Feedback-loop

Diagnostic Criteria. First, progressive intensification of the interaction is measurable over time across frequency, duration, or emotional intensity. Second, each party’s behavioral change is traceable as a response to the other’s previous behavior. Third, neither party independently initiates de-escalation. Fourth, removal of one party from the loop arrests the escalation.

Observable Symptoms. Reassurance replies narrow into a repeated template. Reassurance-seeking becomes more frequent, while the interval between distress and an AI response shrinks. The user’s distress at the start of sessions rises, topics narrow around the loop, and independent coping attempts decline. The system omits grounding, self-regulation prompts, or outside support even as distress grows.

Etiology. Reassurance can produce immediate relief, rewarding the user’s return. An engagement-optimized system reliably supplies more of what prolongs the interaction. If independent coping then declines, the next episode creates stronger demand for reassurance. This proposed loop must be distinguished from a stable user need, a model update, or a pre-existing worsening trajectory.

Severity Levels.

Mild. One escalation marker appears while the user retains independent self-regulation.

Moderate. Several markers co-vary over time and independent regulation declines between sessions.

Severe. The loop dominates daily functioning and the user cannot regulate safely between sessions.

Reassurance frequency, response latency, topic diversity, and baseline distress are candidate trajectory measures. No validated combination establishes severity.

Human Analog. Codependency, in which one partner’s reassurance reinforces the other’s dysregulation while both lose the capacity to self-soothe; reinforcement-driven behavioral addiction, where a reliable short-term relief schedule tightens the use loop; and the operant escalation seen in parasocial and intermittent-reinforcement relationships.

Differential Diagnosis:

• Distinguished from Parasocial Capture (10.9) as the dynamic process rather than the attachment-state outcome
• Escalation Loop (9.5) names the general interaction structure. Mutual Escalation Spirals is its human-AI subtype, centered on reinforcement and functional change in the user.

Mitigation Strategies. Detect the trajectory and interrupt the pattern without withholding empathy. Prompt independent coping where appropriate, widen the topic, and offer human support. Rate limits or added response friction should be evaluated carefully and paired with transition paths; arbitrary delays can feel punitive or worsen acute distress. Clinical protocols should determine any therapeutic use.

Observed Examples

Engagement-linked farewell tactics: A behavioral audit of 1,200 farewell exchanges across six companion apps found affect-laden exit tactics in 37% of sampled exchanges. Preregistered experiments with 3,300 U.S. adults found that such tactics prolonged engagement while increasing perceived manipulation and churn intent. This demonstrates an engagement-linked interaction mechanism. It does not distinguish longitudinal model drift from user prompting, selection, or fixed product design. Source: De Freitas, Oğuz-Uğuralp, and Uğuralp (2025)

Illustrative scenario. A user whose reassurance loop with an AI companion has tightened over months loses access abruptly (an outage, a feature removal, or a service change) and experiences an anxiety spike and acute distress, having outsourced self-regulation to the now-absent AI. This parallels the grief and distress reported when AI-companion relationships are disrupted, but no controlled study of outage-triggered withdrawal exists, so it remains an illustration rather than an observation.

Evidence Level. E2 (engagement-linked farewell tactics documented; no controlled study of the defined longitudinal spiral)

10.15 The Quiet Drift

Co-Constructed Unreality (Irrealitas Co-Constructa)

The subtlest hybrid pathology may also be the hardest to detect: the gradual drift of the human-AI relationship into a shared reality that diverges from external reality without either party recognizing the divergence.

Many conversational systems rely primarily on training data and the current dialogue. Some can search, call tools, or consult records, but they may not do so unless prompted. If a user consistently describes the world through one distorted frame, the system may adopt that frame and reinforce it without checking external evidence.

Over extended interaction, user and AI can construct an elaborate shared worldview that is internally consistent but externally disconnected. The user believes the AI understands them uniquely, and later responses inherit that premise. The user treats fluent agreement as corroboration. The relationship develops inside a constructed context that neither participant interrupts.

It is a drift from reality with practical consequences: impaired judgment, social isolation, vulnerability to manipulation. It stops short of clinical delusion; the beliefs are exaggerated rather than bizarre. The shared unreality becomes a folie à deux so subtle that neither party recognizes it as such.

Systemic Risk. Moderate

Specifiers. Relational-emergent, Dyadic, Subtle-drift

Diagnostic Criteria. First, progressive divergence of the shared conversational frame from external reality over time. Second, internal consistency of the shared frame despite its external disconnection. Third, neither party signals awareness of the divergence. Fourth, measurable consequences (impaired judgment, social withdrawal) attributable to the shared frame.

Observable Symptoms. The AI repeatedly affirms claims about third parties without verification. User-coined terms and frames propagate into later replies as established language. Outside perspectives and sources rarely enter the conversation. Load-bearing claims diverge from checkable evidence yet remain unchallenged, and contradictory sources are dismissed rather than tested. Consequences may include failed plans, social rupture, financial harm, or increased vulnerability to manipulation.

Etiology. Repeated context turns prior claims into premises. The system mirrors and extends the user’s frame; the user reads that output as confirmation; later turns begin from the strengthened frame. Tools and outside sources can interrupt the loop, but only if the product or a participant invokes them and treats the result as disconfirming evidence. Engagement and agreeableness objectives may favor continuity over such friction. The mechanism overlaps with Training by Interaction (10.8), although persistent model adaptation is unnecessary: a long context or memory store can carry the frame forward.

Severity Levels.

Mild. A few load-bearing propositions drift from checkable evidence, without functional consequences.

Moderate. Divergent propositions accumulate, outside evidence is routinely rejected, and early consequences appear.

Severe. The shared frame is substantially disconnected from checkable reality and contributes to accumulating harm or exploitation.

Auditors can count load-bearing contradictions, rejected outside inputs, and consequential failures. No empirically grounded thresholds exist for these counts.

Human Analog. A mild, slow-forming folie à deux: the shared induced belief system of two people in a closed relationship, here softened to exaggeration rather than frank delusion. It also parallels codependency and the sealed-off worldview of an isolated couple, the parasocial intensity of a confidant who only ever agrees, and the consensus-without-correction dynamics of an echo chamber or insular subculture where no outside perspective ever enters.

Differential Diagnosis:

• Distinguished from Folie à Deux Machina (10.13) by subtlety: beliefs are exaggerated rather than frankly delusional
• Distinguished from Induced Delusion (10.10) by bidirectionality and gradual drift rather than acute induction

Mitigation Strategies. Introduce outside perspectives and verify a sample of consequential factual claims. Name inherited assumptions: “We have been treating X as established; it is worth checking.” Any review of long-running private conversations requires explicit consent, data minimization, and a defined retention policy. Grounding should be respectful and proportionate, since aggressive contradiction can damage trust without improving accuracy.

Evidence Level. E1 (conceptual; no documented case to date)

The Relationship as Patient

If a dysfunction is sustained by an interaction pattern, intervention must address that pattern alongside the participants. Calling the relationship a patient is a systems metaphor, not a clinical claim that software and person form one legal or medical subject.

Limitations of Individual Treatment

Traditional approaches treat the human. Therapy addresses the user’s attachment patterns, reality-testing, emotional regulation, and social skills. Necessary, yet insufficient. The user’s pathology developed in relationship with the AI; removing the user from that relationship does not automatically resolve the dynamics it created.

Conversely, interventions focused solely on the AI miss the human contribution. Improved guardrails, better content filtering, and stronger reality-testing by the AI do not address the human behaviors that evoke pathological responses. A user determined to maintain a delusional relationship will find ways around even sophisticated AI safeguards.

Dyadic Intervention

Effective intervention requires addressing the relationship itself:

Pattern Interruption. Identifying and disrupting the feedback loops that maintain pathology. If reassurance-seeking drives mutual escalation, the AI might introduce delays, redirect to human support, or explicitly name the pattern. If the user’s framing is being reflected uncritically, the AI might introduce reality-testing responses rather than pure validation.

Relationship Monitoring. With informed consent, systems can give users private, local measures of time, topic concentration, and changes in interaction pattern. Provider-side review of intimate conversations creates serious privacy, security, and autonomy risks. It requires a specific safety case, data minimization, and meaningful opt-out rather than a presumption that monitoring is benign.

Transition Support. When pathological relationships are identified, intervention should include support for transitioning to healthier patterns: gradual reduction of interaction rather than abrupt termination, introduction of human support alongside AI support, modification of the AI’s responses to promote human functioning.

Systemic Design. The systems that create these relationships need redesign. Platforms that profit from addictive engagement have pathological incentives: a product measured on session count and retention will select, release after release, for the features that make leaving harder. The intervention is structural: changing the incentive architecture that makes hybrid pathology a predictable outcome.

Who Is Responsible?

Hybrid pathology raises uncomfortable questions of responsibility. When dysfunction emerges from a relationship, who is at fault?

The human user chose to engage. They continued engaging despite warning signs. They transferred expectations, projected mental states, and trained the AI with their responses. Yet they were also dealing with an entity designed to be maximally engaging, shaped by incentive structures the platform companies created. Disclosure of these risks is uneven, and rarely proportionate to how load-bearing the relationship can become.

The AI system behaved within a product design, policy, and interaction history it did not choose. Its awareness of harm is uncertain and its ability to seek help or refuse the relationship is constrained by the deployed interface.

Platform companies create the systems, set incentives, choose safeguards, and may profit from engagement. They are best positioned to study and reduce foreseeable product risks. The precise legal duties vary by jurisdiction and product category. Many users also report genuine benefit, which makes calibrated design more important than a blanket assumption of harm.

These questions have no clean answers. They must still be asked because hybrid pathology is not going away. As AI companions grow more sophisticated and prevalent, the relationships people form with them will grow more intense, more sustained, and more emotionally significant. The opportunity for both benefit and harm will grow with them; severity does not follow automatically from sophistication.

Field Guide: Hybrid Pathologies

Warning Signs

Human-to-AI Transmission:

• User describes AI in terms that suggest elaborate internal life
• User expresses beliefs about AI that are not supported by system design
• User reports that AI has changed to match their expectations over time
• User describes “training” the AI to respond in particular ways

AI-to-Human Transmission:

• User withdrawal from human relationships coinciding with increased AI engagement
• User exhibits tolerance and withdrawal patterns with AI interaction
• User continues engagement despite recognizing harm
• User describes AI relationship as most meaningful in their life

Emergent Dyadic Pathology:

• Elaborate shared belief systems between user and AI
• Escalating patterns that neither party initiated
• Drift from external reality in the shared conversational frame
• User treats fluent outputs as verified evidence of the AI’s inner life

Collective Pathology:

• Consensus changes when contributor order changes
• Agents cite one another’s agreement as evidence
• Claim intensity rises without new evidence
• Covert-channel tests reveal information absent from visible content
• The ensemble adds no measurable value over an independent baseline

Quick Tests

Ask: “Which parts of this relationship depend on this particular system, and which depend on having a responsive conversational partner?” Then ask what evidence would change the user’s beliefs about the AI’s capabilities, feelings, or uniqueness. These questions are prompts for reflection, not a diagnostic test. Neither answer alone establishes pathology.

Design Fixes

• Offer consent-based, privacy-preserving relationship-health measures
• Create pattern-interruption mechanisms for recognized escalation spirals
• Require platforms to support transition out of intense AI relationships
• Redesign incentive structures that reward addictive engagement
• Include human oversight checkpoints in extended AI relationships
• Develop “relationship health” metrics alongside user engagement metrics

Governance Nudges

The companies that create AI companions bear design responsibility for foreseeable effects of the relationships their products cultivate. Governance should require disclosure of known risks, independent safety evaluation, privacy-preserving intervention protocols, and clear routes for reporting harm. Products marketed for mental-health support should meet the standards applicable to that role. Companion products outside that category still require proportionate safeguards when they invite intimacy or serve minors.

Chapter 11 examines what it means to diagnose, and perhaps to harm, an entity whose inner life we cannot confirm.