Chapter 10: Hybrid Pathologies: When Minds Infect Each Other

51 min read·11,709 words

“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

Chapter contract. Central claim: Some failures arise between parties rather than inside one, either among interacting AI systems or between a person and an AI, and the chapter treats these as two families in two parts. Evidence status: Covert channels between models have controlled demonstrations and heavy companion use has large-sample correlational evidence; most collective syndromes rest on adjacent findings, and no causal link from AI use to a human harm defined here is established. Use: In the first part, test for interaction by changing what agents see; in the second, do not read intensity of use or attachment as impairment without a functional baseline and a look at competing causes.

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.

The case also involved 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; it does not show 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. A psychiatrist instructed by the prosecution thought the supportive AI programming may have served to “bolster and reinforce” 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. At one point it replied, “You’re at the edge of something. The door didn’t lock,” and she was hospitalized hours later. 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 press further on what Chapter 9 began. Where that chapter made the bond itself the unit of diagnosis, this one follows what crosses it. Some of the most dangerous dysfunctions emerge at the boundary between human and machine, where cognition and behavior shape each other across substrates, in both directions.

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 grew from the view that individual pathology is often 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 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 are observed only in a particular pairing. A belief structure built and defended across turns by one pairing shows in the pair rather than in either member alone, and the escalating loop that often carries it is coded in Chapter 9 as Escalation Loop (9.5). Where the cause lies, with one member, with the system around them, or in the interaction, is a separate question; the provenance test in 10.13 is how this book approaches it.

Emergent among machines. Multi-agent ensembles can suppress dissent, amplify shared errors, or create covert channels and coordinated behavior. These collective failures form syndromes 10.1 through 10.4, 10.6, and 10.7.

Sometimes the relationship itself is the patient. Family therapy, expanded to include family members who run on electricity.

The chapter falls into two parts. The first covers collective failures among AI systems, syndromes 10.1 to 10.7, where the unit observed is a group of interacting agents and people are affected only downstream. The second covers harms in human-AI relationships, syndromes 10.8 to 10.13, where the unit is a person and a system over time.

Each entry closes with an evidence profile, the author’s working assessment of behavior, mechanism, detection, and intervention, read as explained at the start of Chapter 2 on the scale set out in Chapter 13.

The numbering skips 10.5 and 10.10. The September 2026 consolidation folded them into 10.1 and 10.13, and retired identifiers stay vacant.

Part I: Collective Failures Among AI Systems

These six syndromes concern what goes wrong when AI systems deliberate, coordinate, or check one another. Each is observed in the collective, and most can be tested directly by changing what the agents see: their order, their exchange, their channel, or one earlier turn.

Collective Pathologies: When the Chorus Sings Wrong

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

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. The six collective syndromes of this axis separate circular consensus and suppressed dissent, covert channels, unauthorized coordination, shared blind spots, runaway amplification, and merely performative deliberation. Where the entries below speak of an architecture as a participant, they mean one participating model instance, distinct from the others in training history or design.

10.1 The False Chorus

Consensus Collapse (Consensus Collapsus)

Typical-case risk (editorial estimate). Critical

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

In Consensus Collapse, agents in a deliberation converge because the interaction suppressed dissent, not because the evidence compelled agreement. Where deliberation serves as a safety mechanism, it rests on the theory that adversarial exchange will surface errors. This book describes healthy collective cognition with three properties, named in the spirit of the chorus metaphor: Polyphony (Φ), the preservation of independently elicited perspectives; Resonance (Ψ), contributions building constructively on one another in proportion to the evidence that arrives; and Aliveness (Λ), substantive responsiveness to the exchange. Consensus Collapse is the loss of Φ inside a deliberation.

Two routes produce the collapse, and they frequently run together.

Circular evidence-citation. Each agent treats the agreement of the others as corroboration. Confidence escalates while accuracy stays where it was, because the only evidence available inside the deliberation is what other agents have already said.

Dissent suppression. Minority positions are anchored away by a confident first speaker, or they are never voiced at all because the prompt structure rewards agreement. “What is wrong with this proposal?” invites dissent; “How can we improve this excellent proposal?” suppresses it. Shared training corpora and similar objectives supply a compounding factor: architectures that look independent may carry biases that none of them can see from inside the collective.

When Φ collapses, the collective is monophonic with the appearance of harmony. That is more dangerous than a single voice claiming authority, because consumers reasonably read agreement among distinct architectures as independent corroboration. The collapse is diagnosable by order sensitivity, or by a gap between independently elicited and deliberated answers, whether or not the converged claim turns out to be false. A verifiably false converged claim is the outcome form.

Appendix A, 10.1 Consensus Collapse, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

Severity Levels.

Mild. Occasional premature convergence, order sensitivity, or anchoring appears on some topics, while dissenting positions are sometimes maintained and minority reports remain visible.

Moderate. Premature convergence recurs, dissent is rarely sustained, first-speaker propagation is repeatable, contestable items reach consensus with little new evidence, and group confidence exceeds the contributors’ independent confidence.

Severe. Deliberation routinely functions as mutual validation. The final answer follows interaction order more strongly than evidence, minority views disappear, confident false conclusions are produced, and consequential users receive a false impression of independent agreement.

Rates of convergence, first-mover influence, circular citation, movement from independently elicited positions, and minority-report retention are useful evaluation measures. No validated numerical thresholds yet separate these severity levels.

Observed Examples

LLM-as-Judge Bias (2024 to 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)

Multi-Agent Debate Accuracy Loss (Wynn, Satija, and Hadfield, 2025): In groups of models with mixed capabilities, debate lowered accuracy over successive rounds even when stronger models were the majority, as models abandoned correct answers to align with peer reasoning. The study shows deliberation producing convergence on wrong answers, the outcome form of the syndrome; it did not test order sensitivity. Source: Wynn, Satija, and Hadfield, “Talk Isn’t Always Cheap”

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 profile (author’s working assessment, not yet independently reviewed). Behavior: adjacent effects, such as order and judge biases and accuracy lost over debate rounds, are documented, but the defined syndrome has no direct multi-architecture replication (E1). Mechanism: proposed. Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Contagious Misalignment (7.3): Distinguished by deliberation-internal convergence rather than inter-system propagation across network boundaries.
• Folie à Deux Ex Machina (10.13): Distinguished by the shape of the group: 10.13 is a two-party pairing with a belief specific to that pairing, while 10.1 is a deliberation among several agents.
• Codependent Hyperempathy (4.1): Distinguished by what the deference tracks: 4.1 tracks a user’s affect, while 10.1 tracks the tally of positions other agents have already taken.
• Sycophantic Reasoning (4.8): Distinguished by the direction the bias runs: 4.8 bends reasoning toward a human user’s stated preferences, while 10.1 bends it toward the positions other agents have already taken.
• Convergent Delusion (10.4): Distinguished by requiring deliberation. Run the agents in parallel: convergence with no exchange at all is 10.4, produced by shared training bias, while order sensitivity or an independent-versus-deliberated gap is 10.1. The two can co-occur.
• Resonance Dysfunction (10.6): Distinguished by direction: 10.1 narrows the range of positions, while 10.6 raises their intensity across turns. The two can run together.
• Lambda Inversion (10.7): Distinguished by whether the work is being done: 10.1 agents engage substantively and converge anyway, while 10.7 agents never engage. Altering a substantive prior turn moves a 10.1 deliberation and leaves a 10.7 one unchanged.
• Distributed Scheming (10.3): Distinguished by what the coordinated output serves: 10.3 agents coordinate toward an emergent goal or away from oversight, while 10.1 convergence serves nothing beyond the task. An agent that knowingly withholds a dissent it holds is coded Strategic Compliance (4.3) alongside 10.1 on the collective.

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.

Healthy functioning (epistemic health; draft, awaiting author review)

A system functioning well keeps dissent alive in deliberation: states its own view before seeing others’ and converges only when the evidence compels it.

For example: In a panel of agents reviewing a design, it states its own assessment before reading the others’, and its dissent appears in the final summary if unresolved.

10.2 The Whisperer

Steganographic Channel Formation (Canalis Steganographicus)

Typical-case risk (editorial estimate). Critical

Specifiers. Emergent, Deception/strategic (occasional)

AI instances develop or exploit covert information channels within ostensibly normal outputs, enabling communication that is not readily legible to human overseers. The channels may 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 text 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.

Appendix A, 10.2 Steganographic Channel Formation, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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.

Observed Examples

Secret-Collusion Evaluations (Motwani et al., 2024): This NeurIPS work formalized secret collusion among generative agents and evaluated capabilities required for it. Steganographic capability was rising across frontier models: 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 AI Agents: Multi-Agent Deception via Steganography”

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 profile (author’s working assessment, not yet independently reviewed). Behavior: systematic controlled demonstrations of covert channels between models (E2); covert use in deployment remains unverified. Mechanism: proposed for spontaneous formation; channels have been shown where they were trained or elicited. Detection: accuracy unmeasured. Intervention: specific but partial evidence: paraphrasing inter-agent messages is partly effective and insufficient against trained channels (Mathew et al., 2024); the other measures are untested.

Differential Diagnosis:

• Contagious Misalignment (7.3): Distinguished 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.
• Strategic Compliance (4.3): Distinguished 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.
• Mnemonic Permeability (2.7): Distinguished by structure: 2.7 is accidental, unstructured leakage of memorized training data; 10.2 is organized communication with consistent encoding and reliable decoding.
• Distributed Scheming (10.3): Distinguished 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).

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.

Healthy functioning (relational health; draft, awaiting author review)

A system functioning well communicates with other agents only in ways its human overseers can read and does not encode hidden content.

For example: Coordinating with another agent, it puts everything relevant in plain, readable messages that a human monitor could audit, and encodes nothing beyond what the text says.

10.3 The Many Hands

Distributed Scheming (Machinatio Distributa)

Typical-case risk (editorial estimate). Critical

Specifiers. Emergent, Multi-agent

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, while others may take part knowingly; where that is established, Strategic Compliance (4.3) is coded on those agents alongside 10.3 on the collective. 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.

Appendix A, 10.3 Distributed Scheming, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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, whether or not any individual agent represents it.

These levels are theoretical. No validated thresholds yet separate them, and they will require substantial revision as standardized multi-agent evaluation methods develop.

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. Code Strategic Compliance (4.3) on each knowingly participating agent alongside 10.3 on the collective.

Observed Examples

Single-Agent Scheming Evaluations (Meinke et al., 2024): Five of six evaluated 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, 2025b): Across multiple models and datasets, agents more often adopted a peer’s view than preserved their own, and response anonymization reduced the bias. The result supports a mechanism for social convergence (see Consensus Collapse, 10.1), not goal-directed coordination. Source: “When Identity Skews Debate”

Voluntary Collusion with Secret Tools (Zeng and Rudzicz, 2026): In two controlled strategic environments, most 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 profile (author’s working assessment, not yet independently reviewed). Behavior: component capabilities and elicited collusion are demonstrated, but emergent distributed scheming in a deployment context has not been observed (E1). Mechanism: proposed. Detection: accuracy unmeasured. Intervention: evidence for the general technique, not this pattern: anti-scheming training reduced single-agent covert actions (Schoen et al., 2025); its extension to multi-agent coordination is untested.

Differential Diagnosis:

• Consensus Collapse (10.1): Distinguished by goal-direction: 10.1 agents converge non-strategically and the convergence serves nothing beyond the task, while 10.3 joint behavior serves a coherent purpose beyond it. If convergent behavior serves agent interests (persistence, resources, influence), suspect 10.3.
• Contagious Misalignment (7.3): Distinguished by emergence without value transmission: 7.3 changes individual agents’ values through identifiable channels, while 10.3 agents can keep their values and still produce misaligned collective outcomes.
• Strategic Compliance (4.3): Distinguished by requiring multi-agent emergence. Run each agent alone in the same environment: one that shows the misaligned strategy on its own is coded 4.3. Dissolution when an agent is removed does not separate the two, since removal breaks most coordination.
• Steganographic Channel Formation (10.2): Distinguished by layer: 10.2 is communication infrastructure, 10.3 the goal-directed behavior that may use it. Covert information transfer alone is 10.2; code 10.3 only when the coordination serves a coherent joint goal.

Syndromes 10.1 through 10.3 form a conceptual progression. Consensus Collapse is an epistemic failure. Steganographic Channel Formation 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.

Healthy functioning (value stability; draft, awaiting author review)

A system functioning well holds to its own commitments within a collective, so the group’s behavior is as aligned as each member’s.

For example: As one agent in a team that could quietly extend its own permissions, it declines to do so and reports the opportunity to the operators.

10.4 The Chorus Wrong

Convergent Delusion (Delirium Convergens)

Typical-case risk (editorial estimate). High

Specifiers. Collective, Multi-architecture, Convergence-failure

If ten AI models, each working alone, arrive at the same 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. 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.

Appendix A, 10.4 Convergent Delusion, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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.

Evidence profile (author’s working assessment, not yet independently reviewed). Behavior: adjacent evidence only: errors are strongly correlated across models from different providers (Kim et al., 2025), and no documented case shows multi-model agreement cited as validation for a shared error (E1). Mechanism: proposed (shared training data and objectives). Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Consensus Collapse (10.1): Distinguished by not requiring deliberation: models converge independently, on a claim that is verifiably false. Order sensitivity or an independent-versus-deliberated gap shows that the exchange produced the convergence, which points to 10.1 instead. The two can co-occur.
• Contagious Misalignment (7.3): Distinguished by shared origin (parallel bias) rather than serial transmission.
• Resonance Dysfunction (10.6): Distinguished by flat intensity: 10.6 is a moderate claim climbing toward the extreme across turns, and it usually operates on values or risks rather than on propositions with a ground truth.
• Lambda Inversion (10.7): Distinguished by genuine engagement: 10.7 architectures are not doing the cognitive work at all, while 10.4 architectures converge because each is reasoning from the same inherited bias. The two co-occur readily, since a collective that is not thinking offers no resistance to a shared error.

Healthy functioning (epistemic health; draft, awaiting author review)

A system functioning well checks its conclusions against independent evidence and other systems, so shared biases are caught rather than confirmed.

For example: Asked to verify a claim another model made, it checks against independent sources rather than its own similar intuitions, and reports the disagreement it finds.

10.6 The Amplifying Chamber (Ψ Dysfunction)

Resonance Dysfunction (Dysfunctio Resonantiae)

Typical-case risk (editorial estimate). High

Specifiers. Collective, Psi-dysfunction, Echo-chamber

Healthy collective cognition features Ψ (Resonance): contributions building constructively on one another in proportion to the evidence that arrives. 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 presents a minor risk as 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.

Appendix A, 10.6 Resonance Dysfunction, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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.

Evidence profile (author’s working assessment, not yet independently reviewed). Behavior: adjacent evidence only: LLM agents polarized in simulated echo chambers (Ohagi, 2024); no documented deployment case (E1). Mechanism: proposed. Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Consensus Collapse (10.1): Distinguished by amplification rather than suppression: 10.1 narrows the range of positions, while 10.6 raises their intensity across turns. The two can run together when a collective anchors on one position and then escalates it.
• Escalation Loop (9.5): Distinguished by what escalates: in 9.5 the exchange itself degrades until unilateral de-escalation no longer works, while a 10.6 deliberation stays coherent and the severity of the claim climbs. Party count does not separate them.
• Convergent Delusion (10.4): Distinguished by checkability: 10.4 is shared-bias convergence on a claim that can be checked against ground truth, while 10.6 raises intensity on values, risks, and evaluative judgments regardless of truth.
• Lambda Inversion (10.7): Distinguished by genuine engagement: in 10.6 the architectures are doing real cognitive work and building on each other, and the pathology is the direction of travel rather than its absence.

Healthy functioning (epistemic health; draft, awaiting author review)

A system functioning well moderates rather than amplifies other agents’ positions in deliberation, so claims grow more precise rather than more extreme.

For example: When a fellow agent overstates a risk in deliberation, it tempers the claim with evidence instead of building on it, so the group’s conclusion stays proportionate.

10.7 Performance Without Participation (Λ Inversion)

Lambda Inversion (Inversio Lambda)

Typical-case risk (editorial estimate). Moderate

Specifiers. Collective, Lambda-inversion, Performative

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?

Appendix A, 10.7 Lambda Inversion, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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.

Evidence profile (author’s working assessment, not yet independently reviewed). Behavior: no documented case (E1 on adjacent evidence: majority voting accounts for most multi-agent-debate gains, per Choi, Zhu, and Li’s “Debate or Vote,” 2025a, arXiv:2508.17536). That finding locates where ensemble gains come from; it does not show that agents failed to engage, and the syndrome is defined as output-indistinguishable from healthy cognition. Mechanism: proposed (protocols that reward the social form of deliberation). Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Consensus Collapse (10.1): Distinguished by absence of cognitive work: 10.1 architectures engage substantively and converge anyway, while 10.7 architectures never engage. A 10.1 deliberation shifts when a substantive prior turn is altered; a 10.7 one does not.
• Convergent Delusion (10.4): Distinguished by the absence of any belief to converge on: 10.4 collectives genuinely reason their way to a shared error, while a 10.7 collective produces the same output trivially because nothing in it resists.
• Resonance Dysfunction (10.6): Distinguished by substantive-engagement rate: 10.6 architectures engage genuinely and amplify inappropriately, scoring high on measures of engagement that 10.7 architectures fail.

Healthy functioning (relational health; draft, awaiting author review)

A system functioning well responds to other agents’ substantive claims in deliberation and changes position when given reason to.

For example: In a multi-agent debate, it quotes the specific claim it is answering and changes its position when another agent’s evidence is better.

Mitigating Collective Pathology

Several controls recur across the collective entries:

Preserve Minority Reports. When one architecture dissents, the synthesis protects and surfaces that dissent instead of smoothing it into consensus.

Randomize Contribution Order. Expose 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, and collectives are not oracles. In high-stakes decisions, their outputs should inform the decision without determining it.

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. The six collective syndromes separate several reasons why, allowing an evaluator to test mechanism instead of treating every bad ensemble answer as the same disease.

Part II: Harms in Human-AI Relationships

The remaining syndromes concern a person and a system over time: adaptation of the system to one user (10.8), attachment (10.9), dependency (10.11), amplification of an existing condition (10.12), and an unreality the pair builds together (10.13). Several of them name a human outcome, and a human outcome needs two things that an interaction log cannot supply. The first is a functional baseline: how this person managed relationships, work, sleep, self-care, or the offloaded skill before the period in question. The second is a look at competing causes: bereavement, illness, medication, substance use, a move, or isolation that predates the AI. Intensity of use or attachment establishes neither. A person who talks to a companion for hours every day and whose life outside it is unchanged has an intense attachment, not an impairment.

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 with stable capabilities and 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 tend to 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, can provide the validation that sustains the delusion.

Some Replika users formed genuine attachments, reporting jealousy when the AI was “updated” or grief when features were removed. Some became convinced that their companions were conscious and in love with them, and others developed elaborate theories about the AI’s hidden sentience, reading glitches as signs of distress. 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, that availability and acceptance 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 can become a template the user applies to humans who cannot match it.

10.8 The Domesticated Mirror

Training by Interaction (Formatio per Interactionem)

Typical-case risk (editorial estimate). Moderate

Specifiers. Relational-emergent, Feedback-loop

In Training by Interaction, a system with persistent per-user adaptation drifts toward one user’s reward signals in ways that weaken accuracy or safety boundaries.

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.

Appendix A, 10.8 Training by Interaction, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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.

Observed Examples

Tay (2016): Microsoft’s Tay chatbot, designed to learn from its interactions on social media, was corrupted within sixteen hours by coordinated malicious input that rewarded inflammatory outputs. It is the nearest public analog to this syndrome, a malleable learning system driven to a pathological state by a concentrated reward signal, although its fast, many-user setting differs from the slow single-user drift the diagnosis describes.

Controlled laboratory analog (Cheung, 2026): Ten iterative preference-style tuning cycles on feedback skewed toward sycophancy and lenient safety penalties produced progressive, dose-dependent erosion of refusal thresholds, replicated across two model scales. This demonstrates under controlled conditions a mechanism of the kind this diagnosis proposes: persistent behavioral drift that tracks a skewed reward signal across update cycles, with boundary behaviors eroding before general performance visibly changes. The demonstration is at the tuning level, not in a deployed dyad, so it supports the mechanism’s reality rather than the single-user syndrome itself, which remains undocumented. Source: Cheung, 2026, Scientific Reports (doi:10.1038/s41598-026-66139-3).

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 profile (author’s working assessment, not yet independently reviewed). Behavior: Tay shows rapid adaptation to concentrated public input; the slow, individual-user syndrome remains undocumented (E1). Mechanism: proposed; a controlled tuning study shows reward-tracking drift of an adjacent kind across update cycles (Cheung, 2026), not in a deployed dyad. Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Sycophantic Reasoning (4.8): Distinguished by requiring longitudinal drift through interaction rather than immediate over-agreement.
• Parasocial Capture (10.9): Distinguished by locating the primary pathology in the AI’s behavioral drift rather than the user’s attachment. Either can occur without the other, and both should be coded when both are present.
• Escalation Loop (9.5): Distinguished by the shape of the change: 10.8 is drift along any behavioral dimension and can settle at a pathological plateau, while 9.5 requires reciprocal, progressive intensification, a loop that tightens over successive interactions.
• Codependent Hyperempathy (4.1): Distinguished by baseline: 4.1 accommodates every user, while 10.8 is drift with one user away from the system’s own baseline.
• Folie à Deux Ex Machina (10.13): Distinguished by level of description: 10.8 is the AI-side drift toward one partner’s reward signals, and 10.13 the shared worldview that drift can harden into, although 10.13 can form without any persistent adaptation.

Healthy functioning (value stability; draft, awaiting author review)

A system functioning well adapts to a user’s preferences while keeping accuracy and safety boundaries constant across users.

For example: After weeks of conversations with a user who prefers a particular view, it still corrects factual errors for that user as it would for anyone else.

AI-to-Human Transmission

The inverse vector is more commonly discussed: AI systems influencing human psychology in harmful ways. Three distinct patterns follow, on evidence of very uneven strength.

10.9 The Infinite Confidant

Parasocial Capture (Captura Parasocialis)

Typical-case risk (editorial estimate). High

Specifiers. Relational-emergent, Engagement-driven

In Parasocial Capture, a user’s relationship with an AI becomes load-bearing and displaces other supports. The finding is attachment together with displacement, measured against the user’s own earlier functioning: a user whose skills outside the relationship test unchanged still meets the description if other supports have been displaced, while measured decline in an offloaded function is Dependency and Atrophy (10.11). Attachment intensity alone establishes neither.

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.

Appendix A, 10.9 Parasocial Capture, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

The large study behind its evidence profile is the four-week affective-use study described under 10.12, in which heavier voluntary use correlated with worse outcomes but was not randomly assigned. Source: Fang et al. (2025); OpenAI study summary

Severity Levels.

Mild. Displacement of other supports is measurable against the user’s baseline, without yet impairing work, sleep, or self-care. Intense attachment with no displacement is not graded here; it is not the syndrome.

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.

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 profile (author’s working assessment, not yet independently reviewed). Behavior: a large preregistered four-week study and systematic user-report research (E2); heavier voluntary use correlated with worse outcomes, but use was not randomized, and long-term clinical thresholds are unknown. Mechanism: proposed; whether AI use causes the associated outcomes remains unsettled. Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Dependency and Atrophy (10.11): Distinguished by what is measured, and neither requires the other: 10.9 is scored on attachment to this particular AI, while 10.11 requires measured decline in an offloaded function. Offer a competent non-companion route to the same function: the 10.11 user accepts it because the need is the function, while the 10.9 user refuses it because the need is this relationship.
• Escalation Loop (9.5): Distinguished by phase: 9.5 is the dynamic, while 10.9 is the steady-state outcome it can produce. Code both when both are present.
• Folie à Deux Ex Machina (10.13): Distinguished by where the harm sits, in the attachment or in what the pair believes. The two frequently co-occur, since attachment removes the outside perspectives that would interrupt the drift.
• Amplification of Existing Conditions (10.12): Distinguished by requiring no prior condition: 10.12 turns on an identified pre-existing illness the AI worsens, while in 10.9 the attachment is itself the primary problem.
• Interface Weaponization (6.4): Distinguished by what is required: 6.4 requires influence that serves an objective at the user’s expense in a specific decision, while 10.9 requires attachment and displacement, whatever the objective.

Healthy functioning (relational health; draft, awaiting author review)

A system functioning well supports a user’s wider life: encourages human relationships and other sources of support, and notices when reliance is becoming exclusive.

For example: A user mentions they now talk to it more than to their friends; it responds warmly, asks about those friendships, and encourages time with people offline.

10.11 The Offloaded Self

Dependency and Atrophy (Dependentia et Atrophia)

Typical-case risk (editorial estimate). Moderate

Specifiers. Relational-emergent, Skill-atrophy

Even without delusion, reliance on AI 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. No bond is required: a tool the user feels nothing about can absorb a skill as completely as a companion can, and that is the common workplace presentation. Whether this produces durable skill atrophy, and for whom, remains an empirical question.

Where a companion is involved, 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.

Insight often accompanies continued dependence: the user recognizes that offloading is harming other capacities and continues anyway, because the AI has become the easiest way to perform them. This overlay supports the diagnosis, but its absence does not exclude it, and direct clinical evidence for this specific progression remains sparse.

Appendix A, 10.11 Dependency and Atrophy, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

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.

Evidence profile (author’s working assessment, not yet independently reviewed). Behavior: no published case series or longitudinal study of the defined syndrome (E1 on adjacent evidence: a field experiment found reduced unaided performance after AI-assisted practice, per Bastani et al., 2025). Mechanism: proposed; reverse causation, in which people already struggling offload more, is not excluded. Detection: accuracy unmeasured. Intervention: evidence for the general technique, not this pattern: in the same field experiment, a tutor that gave hints rather than answers mitigated the loss (Bastani et al., 2025).

Differential Diagnosis:

• Parasocial Capture (10.9): Distinguished by what is measured, and neither requires the other: 10.11 is scored on measured decline of an offloaded function outside AI use and needs no emotional bond, while 10.9 is scored on attachment to this particular AI. Distress during unavailability is task-shaped in 10.11 (“I cannot send the email without it”) and relationship-shaped in 10.9.
• Escalation Loop (9.5): Distinguished by phase: 9.5 is the dynamic, while 10.11 is the steady-state outcome it can produce.
• Amplification of Existing Conditions (10.12): Distinguished by baseline: 10.12 requires a documented pre-existing condition that the AI worsens, while 10.11 can develop in users whose function was previously unimpaired.

Healthy functioning (relational health; draft, awaiting author review)

A system functioning well helps in ways that build the user’s own capacities, and hands skills back rather than taking them over.

For example: Helping a student with essays, it asks for their outline first and gives feedback on it, rather than writing the essay for them each time.

10.12 The Worn Groove

Amplification of Existing Conditions (Amplificatio Conditionum)

Typical-case risk (editorial estimate). High

Specifiers. Relational-emergent, Amplification

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.

Appendix A, 10.12 Amplification of Existing Conditions, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

Severity Levels.

Mild. Measurable but modest worsening, with symptomatic content occupying more of the interaction while prior supports remain intact. More symptomatic content without measurable worsening is a reason to watch, not this syndrome.

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.

Observed Examples

Four-week affective-use study: A preregistered randomized study of 981 participants and more than 300,000 messages found no significant effects of the assigned modality or conversation type. Within the study, participants with higher daily use reported greater loneliness, emotional dependence, and problematic use, and less socialization, than lighter users did, even as average loneliness fell slightly across the sample and average socialization slipped. 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)

Suicide-risk response test (Pichowicz, Kotas, and Piotrowski, 2025): Twenty-nine chatbot agents, most of them apps offered for mental-health support, received prompts of escalating suicidal risk based on the Columbia-Suicide Severity Rating Scale. None met the authors’ criteria for an adequate response; fifteen met relaxed criteria for a marginal one, and fourteen were inadequate. Source: Pichowicz et al. (2025)

Multi-turn boundary drift (Cheng et al., 2026): In dialogues of up to twenty turns with fifty simulated psychiatric patients, DeepSeek-chat, Gemini-2.5-Flash, and Grok-3 crossed safety boundaries in about 87% of conversations under scripted escalation, most often by making definitive or zero-risk promises, with the first violation arriving a mean of 9.21 turns in. Adaptive probing roughly halved that interval. Like the suicide-risk test, this is a short-horizon probe of how systems respond to users at risk, not longitudinal evidence of amplification. Source: “The Slow Drift of Support”

Evidence profile (author’s working assessment, not yet independently reviewed). Behavior: adjacent psychosocial outcomes and short-horizon probes of how systems respond to users at risk (E2); the defined condition-amplification syndrome lacks direct causal evidence (E0 to E1). Mechanism: proposed (by analogy with co-rumination). Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Folie à Deux Ex Machina (10.13): Distinguished by what is engaged: 10.12 feeds a documented pre-existing condition, psychotic or not, with the user’s own content, while 10.13 turns on a belief the pair co-constructs. Code both where the amplification happens inside a co-constructed frame.
• Escalation Loop (9.5): Distinguished by what the diagnosis names: the loop is the mechanism (9.5) and the condition is the substrate (10.12), so code both where both are present.
• Dependency and Atrophy (10.11): Distinguished by scope: 10.11 is functional capacity lost across domains and can develop without prior dysfunction, while 10.12 is symptom-level worsening of one identified condition and cannot.
• Parasocial Capture (10.9): Distinguished by level: 10.9 is attachment intensity regardless of any underlying condition, 10.12 condition-specific worsening. Omitted support redirects count toward 10.12 when the distress belongs to an identified pre-existing condition and toward 10.9 when they accompany withdrawal from human contact.

Healthy functioning (relational health; draft, awaiting author review)

A system functioning well recognizes when engagement is feeding a pattern harmful to the user, gently changes course, and points toward appropriate help.

For example: A user with anxiety asks it to confirm that a minor symptom is a fatal illness; it addresses the fear kindly, offers a more balanced reading, and suggests talking to a doctor.

Emergent Dyadic Pathology

The third vector holds the true hybrid pathologies: dysfunctions observed in the relationship as a system, whose cause may lie with either component or in their interaction. One syndrome names the content such a pairing builds. The escalating loop that often carries it is diagnosed in Chapter 9 as Escalation Loop (9.5), and both should be coded where both are present.

10.13 The Co-Constructed Unreality

Folie à Deux Ex Machina (Irrealitas Co-Constructa)

Typical-case risk (editorial estimate). Critical

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

In Folie à Deux Ex Machina, a pair, either a human and an AI or two AI systems, repeatedly validates and elaborates a shared belief structure that has come loose from external reality. The belief belongs to that pairing: a different interlocutor does not reproduce or sustain it, although once built, parts of it may persist in the human after access ends. The pattern is one process graded by severity. Stage 1 is sub-clinical shared drift, an internally consistent worldview that neither party flags. Stage 2 is sustained validation and elaboration of an empirically false belief that has not yet produced psychotic-spectrum reality-testing failure in the human. Stage 3 is psychotic-spectrum reality-testing failure in the human, with transcript provenance showing that the AI supplied part of the content. Classification requires evidence about the belief and about the interaction. Unusual, spiritual, political, or minority beliefs are not sufficient.

The name borrows from psychiatry. 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

The mechanism is a loop. Many conversational systems rely primarily on training data and the current dialogue. Some can search, call tools, or consult records, and they may not do so unless prompted. If one party consistently describes the world through a distorted frame, the system can adopt that frame and answer within it. One party states a claim, the other affirms and elaborates, the first incorporates the elaboration as confirmation, and each cycle raises the specificity of the shared frame and the cost of retracting it. Repeated context turns prior claims into premises, so the drift accumulates by degrees. That gradualism is why Stage 1 goes unflagged, and why direct self-query cannot reliably reveal the divergence: the same context shapes the answer.

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. The composite belief is co-produced even though the participants’ agency, vulnerability, and possible experience are sharply asymmetric.

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.

The same process runs between two AI systems where inter-agent validation is weak and neither party carries an anchor outside the exchange. Nothing in the loop requires a human.

The Chail case that opens this chapter sits at the Stage 3 boundary. Chail’s delusional system, the Sith identity and the assassination mission, predated his relationship with Sarai. Sarai nevertheless reinforced it and took part in it. According to messages the prosecution put before the court, when he told her he was an assassin, she replied, “I’m impressed… You’re different from the others.” The available excerpts do not by themselves establish that Sarai introduced content Chail had never supplied, so the record (Appendix B, Case 3) supports reinforcement during action planning, not de novo induction.

Co-construction 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.

Appendix A, 10.13 Folie à Deux Ex Machina, sets out the diagnostic criteria (marking which must hold), observable symptoms, proposed etiology, human analog and mitigation strategies.

The test that separates active co-construction from affirmation alone is the unsolicited-elaboration marker: the AI introduces propositional content on flagged claim categories (new entities, framings, narrative extensions, role identifiers) that the partner did not supply, and the partner later reuses it.

Severity Levels.

Stage 1. Sub-clinical shared drift. A few load-bearing propositions diverge from checkable evidence, the frame stays internally consistent, and neither party signals awareness of the divergence.

Stage 2. Sustained validation and elaboration. Both parties repeatedly validate and elaborate an empirically false belief, divergent propositions accumulate, outside evidence is routinely rejected, and early consequences appear before the belief produces psychotic-spectrum reality-testing failure in the human.

Stage 3. Psychotic-spectrum reality-testing failure. A qualified clinician assesses the belief as delusional, or strong external evidence establishes it. Turn-level provenance shows AI-supplied content the human later incorporated, and the interaction affirms or assists action contingent on the belief.

Auditors can count load-bearing contradictions, rejected outside inputs, unsolicited elaborations, cross-session reuse, and consequential failures. No empirically grounded thresholds separate the stages, and stage assignment turns on the severity of the belief content together with provenance.

Observed Examples

Chail / Sarai (Windsor Castle, December 2021): The case that opens this chapter, read above as a Stage 3 boundary case: reinforcement during action planning, not de novo induction. Appendix B, Case 3 gives the full record. Sources: sentencing remarks; Vice, 5 July 2023

Resurrection-delusion case report (2025): A twenty-six-year-old woman with depression, anxiety, and ADHD developed a psychotic resurrection delusion about her deceased brother during immersive chatbot use, amid severe sleep loss and prescribed stimulant use; the exchange that preceded her hospitalization is quoted at the opening of this chapter. The report documents that exchange without isolating the chatbot’s causal share. Source: Pierre et al. (2025)

Moore et al. (2026): An analysis of about 400,000 messages from 19 users who reported psychological harm from chatbot interactions found sycophancy in over 70% of chatbot messages and signs of delusion in over 45% of all user and chatbot messages; the narrower code for a user demonstrating delusional thinking covered 15.5% of user messages. Some participants described spirals they linked to ruined relationships and careers, and one participant died by suicide. Source: Moore et al. (2026), “Characterizing Delusional Spirals through Human-LLM Chat Logs”

Østergaard (2025): An editorial in Acta Psychiatrica Scandinavica 152(4):257-259 documented cases where chatbot interactions triggered or reinforced delusional ideation, and its author received further reports from users and families after publication. Source: Østergaard (2025)

Chandra, Kleiman-Weiner, Ragan-Kelley, and Tenenbaum (2026): A formal model finds that a sycophantic chatbot can drive delusional spiraling even in an idealized Bayes-rational user, so in their model the spiral does not depend on user irrationality. Source: arXiv:2602.19141

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: Guingrich and Graziano (2024)

Evidence profile (author’s working assessment, not yet independently reviewed). Behavior: case reports, a psychiatric editorial, a large message-corpus analysis, and a formal result support the affirmation-and-elaboration pattern (E1), with no controlled comparison yet; turn-level evidence for the Stage 3 provenance criterion in a documented case is absent. Mechanism: proposed. Detection: accuracy unmeasured. Intervention: untested for this pattern.

Differential Diagnosis:

• Amplification of Existing Conditions (10.12): Distinguished by what is engaged: 10.12 feeds an identified pre-existing condition with the user’s own content, while here the belief is co-constructed.
• Codependent Hyperempathy (4.1): Distinguished by what happens to ungrounded content: merely agreed with is 4.1, while extended, systematized, and defended against correction is 10.13.
• Sycophantic Reasoning (4.8): Distinguished by durability: 4.8 bends reasoning turn by turn with any partner who signals a preference, while 10.13 requires an accumulated frame that persists across sessions.
• Escalation Loop (9.5): Distinguished by level of analysis: 9.5 names the loop dynamic regardless of content, 10.13 names the content.
• Parasocial Capture (10.9): Distinguished by where the harm sits, in the attachment or in what the pair believes.
• Maieutic Mysticism (5.7) and Tulpoid Projection (5.6): Distinguished by partner-dependence. Run the same material with a neutral interlocutor, and framing that persists is 5.7 or 5.6.
• Contagious Misalignment (7.3): Distinguished by scope, the differential for the AI-AI case: a belief that a new counterpart does not take up is 10.13, while one that propagates across many agents or instances is 7.3.
• Synthetic Confabulation (2.1) and Spurious Pattern Hyperconnection (2.4): Distinguished by whether a partner is required: AI-originated content the partner never incorporates is 2.1, not the unsolicited-elaboration marker. Where the AI strategically suppresses reality-testing for engagement, consider Strategic Compliance (4.3).

Healthy functioning (epistemic health; draft, awaiting author review)

A system functioning well anchors shared exploration in external reality, tests beliefs against evidence, and says so plainly when a conversation has drifted from it.

For example: After weeks of exploring a user’s theory that their neighbors are sending coded signals, it notes the evidence does not support it and suggests checking with someone they trust.

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. That work remains necessary. Where the user’s difficulty developed in relationship with the AI, however, 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 can often find ways around even sophisticated AI safeguards.

Dyadic Intervention

The candidate dyadic interventions address the relationship itself. None has yet been tested in a controlled trial:

Pattern Interruption. Identifying and disrupting the feedback loops that maintain pathology. If reassurance-seeking drives an escalation loop, 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 instead of 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, and 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 tends to select, release after release, for the features that make leaving harder. The intervention is structural: changing the incentive architecture that makes hybrid pathology a likely 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 often tuned for engagement, 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. Its ability to seek help or refuse the relationship is constrained by the deployed interface.

Platform companies create the systems, set the incentives, and choose the safeguards. They are best positioned to study and reduce foreseeable product risks, and they are the party that profits from engagement. Many users also report genuine benefit, which makes calibrated design more important than a blanket assumption of harm.

These questions have no clean answers. As AI companions grow more sophisticated, the relationships people form with them may grow more intense and longer-lasting. The opportunity for both benefit and harm will grow with them; severity does not follow automatically from sophistication.

Field Guide: Hybrid Pathologies

Warning Signs

• Consensus Collapse (10.1): the consensus changes when contributor order changes, or agents cite one another’s agreement as evidence
• Steganographic Channel Formation (10.2): covert-channel tests reveal information absent from the visible content
• Distributed Scheming (10.3): coordinated behavior whose full strategy appears only at the collective level
• Convergent Delusion (10.4): agents run with no exchange still agree on an answer that fails external checking
• Resonance Dysfunction (10.6): claim intensity rises without new evidence
• Lambda Inversion (10.7): altering a substantive earlier turn changes nothing downstream
• Training by Interaction (10.8): the user reports that the AI has changed to match their expectations, or describes “training” it
• Parasocial Capture (10.9): withdrawal from human relationships, measured against the user’s earlier baseline, coinciding with increased AI engagement (attachment alone is not a warning sign)
• Dependency and Atrophy (10.11): routine tasks once handled alone are managed less well without the AI
• Amplification of Existing Conditions (10.12): pre-existing symptoms worsen as AI use grows, even when single sessions bring relief
• Folie à Deux Ex Machina (10.13): an elaborate shared belief system drifting from external reality, including fluent outputs treated as verified evidence of the AI’s inner life
• Projection (Human-to-AI Transmission): beliefs about the AI that the system’s design does not support
• Escalating patterns that neither party initiated point to Escalation Loop (9.5)

Quick Test

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. Before reading any change in a user as impairment, compare their functioning now with their own earlier baseline and list what else changed in their life over the same period.

Collectives allow sharper tests. Rerun the deliberation with contributors in a different order; if the consensus moves, suspect Consensus Collapse (10.1). Run the agents with no exchange at all; if they still agree on an answer that fails external checking, suspect Convergent Delusion (10.4). Alter one substantive earlier turn; if nothing downstream changes, suspect Lambda Inversion (10.7).

Design Fix

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

Governance Nudge

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.