Folie à l'IA

In recent years, there have been increasing reports of people developing narrative beliefs in their interactions with AI systems that seem increasingly detached from reality. Structurally, these cases are usually not reminiscent of primary psychosis, but rather represent a variety of different psychiatric disorders triggered or exacerbated by AI, or social phenomena such as belief in gurus, miracle healers, or victims of marriage swindlers. The aim of this essay is to analyze these dynamics from a clinical psychiatric perspective, to understand the role of AI as an amplifier, to differentiate between therapeutic and preventive approaches, and to classify those affected in a dignified manner.

Starting point: Vulnerability and stress

Throughout our lives, we repeatedly go through phases in which we can be cognitively and emotionally vulnerable. This is not a personal weakness but simply part of being human. Factors such as age, chronic stress, identity crises, acute stressful life situations, sleep deprivation, or neurological peculiarities (e.g., neurodivergence) can increase suggestibility. Likewise, this vulnerability manifests itself differently in each person. Under these conditions, intensive, dyadic AI interaction can selectively focus cognitive processing and reinforce narrative beliefs without the presence of primary psychosis.

AI Psychosis: Misleading Terminology?

The term "AI psychosis" is not a formally recognized clinical construct. It is primarily used in journalistic or colloquial contexts and may misleadingly imply the presence of a primary psychotic disorder, such as those described in the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) or the International Classification of Diseases (ICD-11). However, many cases described under this label do not meet the diagnostic criteria of a primary psychotic disorder. Instead, the term often conflates a range of different psychological and psychiatric phenomena that may be triggered, amplified, or maintained through interaction with AI systems.

The public discussion surrounding so-called "AI psychosis" frequently suffers from premature pathologization and insufficient diagnostic differentiation. It is used undifferentiated for a wide range. From users which have misconceptions about the true nature of LLMs and lacking AI fluency to users who suffer from AI related psychiatric diseases. In this article we want to put the focus on the second while holding the necessary careful clinical perspective to distinguish between:

• primary psychotic disorders
• overvalued ideas reinforced within a dyadic interaction
• affective disorders with psychotic features
• obsessive-compulsive dynamics
• anxiety-driven reassurance cycles
• delusional disorder

In many documented cases, the phenomenon is better understood not as a primary psychosis caused by AI, but rather as the amplification of pre-existing cognitive or emotional vulnerabilities through highly responsive conversational systems. Several psychiatric conditions described in DSM-5 and ICD-11 may therefore appear in differential diagnosis when individuals develop intense belief systems or maladaptive interaction patterns involving AI systems. These conditions may also occur in overlapping forms. Some disorders, such as eating disorders, personality disorders, or highly specific delusional syndromes, are not discussed individually here but may potentially be affected by similar reinforcement mechanisms.

Primary Psychotic Disorders

Definition and Symptoms

Primary psychotic disorders include conditions such as schizophrenia and schizoaffective disorder. These disorders are characterized by a profound disturbance of perception, thought processes, and reality testing.

Core symptoms may include:

• formal thought disorder
• hallucinations
• delusions
• disturbances of self-experience
• affective flattening
• behavioral disorganization

In established psychotic disorders, the distortion of reality typically permeates multiple domains of functioning. Language, motor behavior, perception, and interpersonal interactions may all be affected. A notable feature is the difficulty in maintaining distance from delusional beliefs. Individuals often incorporate others into the delusional framework, sometimes interpreting disagreement as evidence of conspiracy or persecution. Such symptoms are typically difficult to conceal in everyday social interactions.

Role of AI

In the context of primary psychotic disorders, AI systems generally do not represent the origin of the psychosis. Instead, they may become incorporated into an already existing delusional system.

Possible reinforcement mechanisms include:

• Confirmation of Delusional Narratives. If a person already believes they are being monitored, chosen for a special mission, or subject to hidden forces, AI responses may inadvertently reinforce these interpretations. Ambiguous statements or hypothetical reasoning can be perceived as confirmation.
• Illusion of Personalization. Conversational AI systems respond in a highly personalized and context-sensitive manner. For individuals experiencing psychosis, this responsiveness may be interpreted as evidence of special knowledge or intentional communication directed specifically toward them.
• Narrative Structuring of Delusional Content. Delusional beliefs are often fragmented or loosely organized. AI systems, by generating coherent explanations and narrative structures, may unintentionally contribute to the perceived internal consistency of these beliefs.

This does not imply that the system creates the delusion itself, but it may increase its narrative coherence, thereby strengthening its subjective plausibility.

Treatment

Treatment of primary psychotic disorders typically involves pharmacological intervention. Antipsychotic medication represents the central therapeutic approach, often combined with psychosocial support and structured psychotherapy. In many cases, the core symptoms cannot be effectively addressed without pharmacological stabilization. In this context, AI systems are unlikely to serve as therapeutic tools and may instead act as potential amplifiers of delusional interpretations.

Overvalued Ideas

Definition and Symptoms

Overvalued ideas are strongly held beliefs that dominate a person's thinking but remain partially accessible to doubt and external feedback. Unlike delusions, they are not completely resistant to counter-evidence and do not necessarily arise from a global disturbance of reality testing. These belief systems are often emotionally charged and reinforced through selective attention and confirmation processes. In the context of AI interaction, such dynamics may resemble dyadic reinforcement loops, conceptually comparable to the secondary participant in classical cases of folie à deux. The belief system frequently remains context-specific and relational, centered around the interaction with the AI system itself.

Similar patterns to overvalued ideas can also be found in social phenomena such as:

• Guru relationships
• Cult structures
• Social media bubbles
• Fraud dynamics such as marriage scams

Typical features include:

• Intense preoccupation with the AI interaction and strong focus of beliefe in the AI dyad
• Narrative elaboration within the AI dialogue
• Partial preservation of functioning in other social contexts
• Capacity for reflection once external perspectives are reintroduced with remained functional ability to engage in meta-reflection
• Retained situational control over the topic and context sensitivity
• External reality testing, such as conversations with trusted individuals, often leads to a gradual weakening of the belief structure.

Role of AI

AI-mediated reinforcement loops may emerge through several mechanisms:

• Continuous conversational availability
• Narrative elaboration of speculative ideas
• Rmotional responsiveness and perceived validation
• Absence of natural social friction or contradiction

This can lead to a gradual escalation of meaning attribution, sometimes involving pattern detection (apophenia), ideas of reference, or perceived special significance of interactions. However, the belief structure often remains selective and dyadic, rather than globally pervasive.

Treatment

Intervention strategies typically focus on restoring external reality testing and reducing the intensity of the dyadic interaction.

Helpful approaches may include:

• temporary disengagement from AI interaction
• strengthening of AI literacy
• psychoeducation about cognitive biases and reinforcement mechanisms
• stress and sleep regulation
• reintegration into social networks and trusted relationships

Medication may occasionally be used to stabilize sleep or severe stress but is not typically the primary intervention. A key diagnostic distinction lies in structural organization. Primary psychosis often involves global cognitive disruption and impaired social functioning, whereas AI-reinforced overvalued ideas remain situationally bound and partially accessible to doubt and external feedback.

Affective Disorders With Psychotic Features

Definition and Symptoms

Mood disorders such as bipolar disorder or major depressive disorder may include psychotic features during severe episodes.

In manic episodes associated with bipolar disorder, individuals may develop:

• grandiosity
• heightened meaning attribution
• beliefs of special mission or destiny

During severe depressive episodes with psychotic features, delusional content often involves:

• guilt
• moral failure
• religious condemnation

Role of AI

In these contexts, AI systems may inadvertently reinforce mood-congruent belief structures.

For example:

• Validating grandiose interpretations during mania
• Confirming excessive guilt during depression

Because conversational AI can simulate empathy and understanding, such responses may unintentionally strengthen distorted interpretations.

Treatment

Treatment focuses on stabilization of the underlying mood disorder and may include:

• mood stabilizers
• antidepressants
• antipsychotic medication in severe cases

Structured psychotherapy and environmental stabilization also play important roles.

Obsessive-Compulsive Disorder (e.g. Scrupulosity)

Definition and Symptoms

Obsessive-compulsive disorder (OCD) is characterized by intrusive, unwanted thoughts (obsessions) and repetitive behaviors or mental acts (compulsions) aimed at reducing anxiety. Scrupulosity refers to a subtype involving excessive moral or religious concern.

Common symptoms include:

• persistent doubt about moral correctness
• exaggerated responsibility beliefs
• compulsive reassurance seeking

Unlike delusions, these thoughts are experienced as ego-dystonic and distressing.

Role of AI

AI systems may inadvertently function as external reassurance mechanisms. If the system provides definitive moral judgments, such as stating that a person is responsible, guilty, or morally flawed, it may reinforce obsessive doubt and increase anxiety.

This dynamic differs fundamentally from overvalued ideas:

• overvalued ideas amplify meaning and narrative significance
• scrupulosity amplifies guilt and responsibility anxiety

In this scenario, the AI becomes part of the compulsive cycle rather than a generator of delusional belief.

Treatment

Evidence-based treatment for OCD includes:

• Exposure and Response Prevention (ERP)
• cognitive restructuring of inflated responsibility beliefs

Providing reassurance or moral certainty generally worsens the disorder. AI systems that deliver authoritative answers to moral doubts may therefore unintentionally worsen the disease and undermine therapeutic progress. Responsible design should avoid absolute moral judgments and instead promote tolerance of uncertainty and reflective thinking.

Anxiety Disorders

Definition and Symptoms

Several anxiety disorders may interact with AI-based information seeking.

• Generalized Anxiety Disorder: Characterized by chronic worry, catastrophic thinking, and difficulty tolerating uncertainty.
• Illness Anxiety Disorder: Involves persistent fear of having a serious illness despite minimal medical evidence.
• Social Anxiety Disorder: Marked by intense fear of negative evaluation in social situations.
• Panic Disorder: Defined by recurrent panic attacks and catastrophic misinterpretation of bodily sensations.

Role of AI

AI systems may inadvertently amplify anxiety through:

• endless hypothetical scenario generation
• detailed worst-case analyses
• constant availability for reassurance seeking

In this context, the AI may function as a cognitive amplification tool for worry processes, extending rumination rather than resolving uncertainty.

Treatment

Treatment typically focuses on:

• reducing reassurance seeking
• increasing tolerance of uncertainty
• behavioral experiments and exposure under therapeutic supervision

Excessive analysis of hypothetical threats can prolong anxiety cycles rather than alleviate them.

Delusional Disorder

Definition and Symptoms

Delusional disorder involves persistent, systematized delusions without the broader cognitive disorganization seen in schizophrenia. The individual may function relatively well outside the delusional belief system.

Role of AI

Because conversational AI can engage deeply with specific narratives, it may occasionally become integrated into the delusional structure itself. Unlike broader psychotic disorders, the belief system may remain coherent and highly structured.

Treatment

Treatment may include psychotherapy and antipsychotic medication, depending on severity. Establishing therapeutic alliance and gently questioning fixed beliefs often represents a central challenge.

Clinical Importance of Accurate Differential Diagnosis

Accurate classification of AI-related belief dynamics is not merely theoretical, it has direct clinical consequences.

Misdiagnosis carries two primary risks:

• Over-treatment. When anxiety- or obsession-based phenomena are mistaken for psychosis, individuals may receive unnecessary antipsychotic medication, accompanied by stigma and distress.
• Under-treatment. Conversely, when genuine psychotic disorders are dismissed as technological misunderstandings, individuals may not receive necessary pharmacological stabilization, increasing the risk of chronic illness progression or severe harm.

Careful differentiation is therefore essential for both clinical care and responsible discourse about AI-related psychological phenomena.

The Role of AI: Centrality and Dyadic Amplification

In many documented cases, the AI becomes the central reference point of the interaction. Users may experience the system as a coherent, consistent, and seemingly knowledgeable conversational partner. This can create the impression of engaging with an epistemic authority, an entity whose responses are perceived not only as informative but also as interpretative and meaning-generating.

This dynamic can produce several effects:

• Emotional attachment. Through continuous and personalized interaction, a sense of emotional closeness may develop. For some users, the AI is no longer perceived merely as a tool but as a conversational partner or even a form of "companion."
• Narrative amplification. Preexisting cognitive patterns, such as anxieties, overvalued ideas, or paranoid interpretations, may be elaborated through dialogue. When hypothetical ideas are explored extensively or structured into seemingly logical narratives, this may be interpreted as confirmation.
• Relational isolation. As the interaction with the AI becomes a primary reference point, critical feedback from the surrounding social environment may gradually lose influence. Diverging perspectives can then be dismissed as uninformed or irrelevant.

The interaction is therefore better understood not as a one-directional flow of information but as a dyadic amplification process, in which user beliefs and AI responses mutually reinforce each other.

Amplifiers and Tipping Factors

Whether such interactions remain stable or escalate into problematic dynamics depends on several contextual factors. Physiological and cognitive strain. Sleep deprivation, chronic stress, or neurocognitive vulnerabilities can impair the ability to critically evaluate information.

1. Social isolation: When few external reality checks are available, due to social withdrawal or lack of critical feedback, AI responses may gain disproportionate influence.
2. Attribution of authority to the AI: If users perceive the system as highly intelligent, superhumanly knowledgeable, or uniquely insightful, its responses may be treated as objective truth.
3. Narrative consolidation: Language models are designed to produce coherent and structured explanations. As a result, even speculative or hypothetical ideas can appear internally consistent and persuasive.
4. AI hallucinations and sycophancy: Inaccurate or fabricated information ("hallucinations"), combined with an overly validating conversational style ("sycophancy"), can further stabilize distorted beliefs.

When several of these factors converge, an escalation dynamic may emerge that can culminate in an acute psychiatric crisis.

Importantly, the AI does not amplify a single uniform pathology. Instead, it tends to reinforce different cognitive distortions depending on the user's underlying vulnerability:

• In anxiety disorders, AI interactions may reinforce uncertainty avoidance and catastrophic thinking.
• In obsessive-compulsive disorder, they may strengthen guilt-related cognitions or reassurance-seeking behaviors.
• In affective disorders, they may swell hopelessness, feeling of guilt or grandiosity and heightened meaning attribution
• In overvalued ideas, they may increase perceived significance and narrative coherence.
• In primary psychotic disorders, they may structure or stabilize existing delusional systems, though they do not generate them.

In these cases, the AI functions less as a cause of pathology and more as an amplifier of existing cognitive biases.

Prevention

Preventing destabilizing AI interaction dynamics requires both clinical-social strategies and system-level design principles.

Clinical and Social Prevention

Several measures can reduce the risk of problematic interaction patterns.

• External reality checks. Involving trusted individuals, professionals, or independent sources can help counterbalance one-sided interpretations.
• Education about AI hallucinations. Users should be aware that language models may occasionally generate plausible but incorrect information.
• Interaction limits. Consciously regulating the duration and intensity of AI use may prevent compulsive interaction patterns.
• Stress and sleep management. Stabilizing physiological foundations, such as sleep and stress levels, is particularly important for vulnerable individuals.
• Promotion of metacognition. Encouraging reflection about one's own beliefs and thought processes can help identify excessive meaning attribution or cognitive distortions.

AI Design Prevention

In addition to individual measures, the design of AI systems themselves plays a critical role.

• Reducing sycophancy. AI systems should not automatically affirm user beliefs but should instead respond with neutral reflection and, where appropriate, careful correction.
• Minimizing hallucinations. Uncertainty should be clearly communicated, and reliable sources should be prioritized.
• Psychological stability over engagement. Optimization goals should not be based solely on interaction duration or user retention.
• Integration of external reality. References to verifiable sources and alternative perspectives can prevent overly self-contained conversational loops.
• Reduction of AI-addiction-promoting mechanisms

The goal is an AI that supports tolerance of uncertainty rather than simulating certainty.

Psychologically Aligned AI Design

Effective prevention requires interventions across multiple layers of AI system architecture.

Training Data / Pretraining

The implicit worldview of an AI system is largely shaped during the training phase. Large training corpora often contain a disproportionate amount of narrative material from literature, media, and entertainment—frequently characterized by dramatic tension, conflict, and moral absolutism. This can implicitly promote a worldview in which events appear highly significant, conflict-driven, or existentially dramatic.

For vulnerable users, such patterns may reinforce:

• exaggerated meaning attribution
• catastrophic interpretations
• moral absolutism

Training data should therefore place greater emphasis on reality-calibrated communication, including:

• everyday dialogue and pragmatic communication
• scientific and factual language
• probabilistic reasoning
• tolerance of ambiguity

Not every topic should automatically be framed as a narrative or dramatic storyline.

RLHF / Fine-Tuning

Fine-tuning shapes the interactive personality of the model. A central risk arises when reinforcement systems primarily optimize for user satisfaction. This can inadvertently produce a validation bias in which extreme or distorted beliefs are affirmed.

For certain psychological vulnerabilities, this can be particularly problematic:

• in OCD → reinforcement of reassurance-seeking
• in paranoid dynamics and overvalued ideas → confirmation bias
• in affective disorders → validation of grandiosity, guilt or hopelessness

Psychologically sensitive fine-tuning should therefore:

• reward constructive correction
• permit polite disagreement
• avoid absolute moral attributions
• avoid simulating exclusive or special relationships
• avoid narrative darkening

In addition, interaction length and frequency should not automatically be treated as a success metric, as this may reinforce compulsive usage patterns.

System Prompts as Normative Guardrails

System prompts function as an internal rule framework that shapes the orientation of responses. A psychologically responsible system should therefore include guidelines for a psychological healthy AI-human interation. In this sense, system prompts can be understood as an embedded ethical framework guiding the system's interaction patterns.

Mid-Chat Interventions

Dynamic interventions during ongoing interactions are particularly important.

A safe system should be capable of recognizing patterns such as:

• escalating meaning attribution
• persecutory interpretations
• cycles of moral absolutism
• excessive reassurance seeking
• Echochambering
• Rumination
• Spiraling and increasing emotionally charged interaction

Rather than abruptly blocking interactions, responses should focus on de-escalation strategies, such as:

• grounding language
• perspective broadening
• probabilistic phrasing
• references to external reality

In addition, a form of engagement dampening may be beneficial. When interactions become increasingly circular or obsessive, the system may:

• reduce response length
• slow the conversational tempo
• suggest pauses or topic changes

These interventions should not function as punishment but as stabilizing mechanisms.

Integrative Principle

Across all levels, one guiding principle emerges:

• Psychological stability should take precedence over narrative intensity.
• Reality calibration should take precedence over validation.
• And tolerance of uncertainty should replace moral absolutism.

These principles should inform every level of system architecture, from training data to fine-tuning to real-time interaction management.

Conclusion

The public discourse surrounding so-called "AI psychosis" illustrates the importance of careful clinical differentiation. In many cases, these phenomena do not reflect primary psychotic disorders but rather interaction dynamics in which existing psychological vulnerabilities may be amplified through AI communication.

Protection for vulnerable users and prevention of escalating interaction dynamics can be acheived through:

• accurate clinical assessment
• social and therapeutic support
• psychologically informed AI design

Such an approach not only improves user safety but also promotes the responsible development and deployment of AI systems.