Iatrogenic at Scale:How AI-Generated Therapeutic Advice for Identity Collapse Deploys the Intervention Most Likely to Deepen the Condition

Don L. Gaconnet, CSE III

LifePillar Institute for Structural Identity Sciences

ORCID: 0009-0001-6174-8384

August 14th 2026

Abstract

AI-generated search overviews now deliver synthesized therapeutic advice to users searching for psychological distress terms, including identity collapse. This paper documents a specific mechanism of harm: the therapeutic interventions recommended by these systems—journaling, mindfulness, value exploration, role experimentation, and narrative self-reconstruction—are precisely the interventions predicted to deepen the condition by a published, DOI-registered structural account of identity collapse. The structural account, derived from the Law of Identity Collapse and the Post-Identity Collapse Defense, identifies the experiential void following identity collapse as a compressive defense by the narrative self-model against a successor state that has already arrived. Narrative-level interventions feed the defense by providing the narrative self-model with additional material to process, reinforcing the compression rather than resolving it. The population most likely to encounter and follow AI-generated therapeutic advice without critical evaluation—those with degraded metacognitive function—is the same population the Recursive Reliability Effect identifies as least capable of assessing whether an intervention is helping or harming. This paper specifies the mechanism, documents the deployed intervention, identifies the affected population, and proposes three testable predictions that would confirm or falsify the harm claim.

Keywords: identity collapse, AI-generated therapy, iatrogenic harm, narrative self-model, metacognitive degradation, search overview, compressive defense, post-identity collapse

1. Introduction

Major search engines now deploy AI-generated overview panels that synthesize answers from multiple web sources and present them above organic search results. These panels are designed to provide immediate, authoritative answers to user queries. When a user searches for a term describing psychological distress—such as “identity collapse”—the AI overview synthesizes therapeutic advice from self-help websites, coaching platforms, and popular psychology blogs, and delivers it as a consolidated recommendation before the user encounters any specific source.

As of August 2026, a search for “identity collapse” on major search platforms returns AI-generated advice that typically includes some combination of the following recommendations: engage in journaling or creative expression to explore thoughts and feelings about identity; practice mindfulness meditation or yoga to foster grounding and presence; set small, achievable goals aligned with values and interests; experiment with new activities, hobbies, or roles to discover “hidden or underdeveloped” parts of the self; approach the experience with “curiosity rather than fear”; and seek professional support if symptoms persist.

These recommendations are standard narrative-level therapeutic interventions. They assume that identity collapse is a psychological experience to be managed through reflection, exploration, and gradual narrative reconstruction. They are drawn from sources that treat identity as fluid and identity collapse as a transitional phase requiring patience and self-directed rediscovery.

This paper argues that for a specific and identifiable population—those experiencing structural identity collapse as defined by the published Law of Identity Collapse—these recommendations constitute an iatrogenic intervention deployed at population scale without clinical assessment, informed consent, or differentiation between conditions that share surface-level symptom presentation but require fundamentally different responses.

2. The Structural Account of Identity Collapse

The Law of Identity Collapse derives five corollaries from a parent axiomatic framework governing structural self-coherence. The relevant corollaries for the present argument are the Conservation Corollary (identity cannot be destroyed, only redistributed), the Two-Vector Theorem (identity collapse proceeds through exactly two failure modes: core compression and peripheral dissolution), and the Closure Corollary (post-collapse coherence is structurally determined).

The Post-Identity Collapse Defense, derived from these corollaries, makes a specific claim: the experiential void that follows identity collapse is not the collapse event itself. It is a compressive defense mounted by the narrative self-model against a successor state that has already arrived. The collapse is complete. The reorganization has occurred. What the person experiences as “not knowing who I am” is the narrative self-model’s refusal to recognize the successor configuration, because recognizing it would require the narrative self-model to relinquish its position as the organizing structure of identity.

This account produces a specific clinical prediction: narrative-level interventions—any intervention that gives the narrative self-model a task (reflect, explore, journal, set goals, experiment with new roles)—will feed the compressive defense rather than resolve it. The narrative self-model receives the task, processes it, fails to find the “old self” in the output, and interprets the failure as evidence that more reflection is needed. The intervention becomes recursive: the person journals more, reflects harder, tries more activities, and each cycle deepens the compression because each cycle confirms to the narrative self-model that the self it is looking for has not been found.

The structural account predicts that resolution occurs not through narrative reconstruction but through what the framework terms “descent”—a non-narrative process in which the compressive defense relaxes and the successor state is recognized rather than constructed. The distinction is categorical: recognition is the identification of something already present; construction is the building of something not yet present. The AI-generated advice uniformly recommends construction. The structural account predicts that construction is the iatrogenic intervention for this condition.

3. The Deployed Intervention

The AI-generated therapeutic advice currently deployed through search overviews for identity collapse queries can be categorized into five classes of recommendation. Each is a narrative-level intervention. Each gives the narrative self-model a task.

Reflective journaling. The recommendation to “engage in journaling or creative expression as a means of exploring your thoughts and feelings” directs the person to produce narrative output about their identity state. The narrative self-model is tasked with examining itself, producing text about what it finds, and using that text as material for further self-assessment. Under the structural account, this is recursive: the system that cannot find the self is asked to write about not finding the self, generating more material for the system to process without resolving the structural condition.

Mindfulness and grounding. The recommendation to “engage in mindfulness practices such as meditation or yoga to foster a sense of grounding and presence” is the least harmful of the five classes because it can, in some implementations, reduce narrative processing. However, as typically deployed in self-help contexts, mindfulness is framed as a tool for “reconnecting with your inner self”—a narrative goal. The grounding is instrumentalized as a means of finding the self, not as an end in itself. This reframes a potentially non-narrative practice as a narrative-level search operation.

Goal-setting aligned with values. The recommendation to “set small, achievable goals that align with your values and interests” presupposes that the person can identify their values and interests—which is precisely the capacity that identity collapse disrupts. The person is directed to access a self-knowledge structure that is not currently accessible and to use it as a foundation for action. Under the structural account, the inability to access this structure is not the problem to be solved; it is the compressive defense in operation. Directing the person to try harder to access it reinforces the defense.

Role experimentation. The recommendation to try “hobbies or interests you’ve always been curious about,” volunteer for causes, or take classes is a construction intervention. It directs the person to build new identity material through activity. Under the structural account, the successor state does not need to be built through new experiences. It has already arrived. The construction activity delays recognition by directing attention toward novelty rather than toward what is already present.

Reframing the experience. The recommendation to “approach this with a sense of curiosity rather than fear” is an affect-regulation instruction that reframes the distress signal as a growth opportunity. Under the structural account, the distress is signal, not noise. It is the narrative self-model’s registration of a structural condition. Reframing the signal as an invitation to explore does not change the structural condition; it changes the person’s relationship to the signal, making them less likely to seek appropriate intervention.

4. The Vulnerability Amplifier: Metacognitive Degradation Under Load

The Recursive Reliability Effect (RRE) demonstrates that self-assessment accuracy degrades under sustained structural load, with metacognition as the primary casualty. Independent confirmation comes from Pihlaja et al., who found that occupational burnout degrades metacognition specifically—not behavioral regulation broadly—while preserving task performance. Duncan et al. confirmed the broader unreliability of self-report-dependent assessment in a JAMA Network Open meta-analysis of 57 studies (N = 8,146), finding pooled test-retest reliability of κ = 0.69 for standardized diagnostic interviews.

This creates a specific vulnerability structure. The population experiencing identity collapse is, by definition, under sustained structural load. The RRE predicts that their capacity to evaluate whether an intervention is helping or harming is degraded precisely because of the condition the intervention targets. When this population encounters AI-generated therapeutic advice, they are:

Less able to critically evaluate whether the recommendation is appropriate for their condition. Less able to detect that the intervention is deepening rather than resolving the distress. More likely to interpret the failure of the intervention as personal failure (“I’m not journaling correctly”) rather than as intervention failure. More likely to increase the dose of the failing intervention rather than abandon it.

The AI overview system thus targets the population least equipped to evaluate its recommendations with recommendations that the structural account predicts will deepen the condition. The system has no intake process, no assessment, no differentiation between identity collapse as defined by the structural account and normal role transition, grief, developmental transition, or psychotic episode. It delivers the same undifferentiated advice to all of these populations simultaneously.

5. The Scale Problem

A single therapist delivering inappropriate advice to a single client in a single session produces localized, recoverable harm. The client can seek a second opinion, change therapists, or encounter corrective information through other channels. The harm is bounded by the therapeutic relationship’s reach.

AI-generated search overviews operate at a fundamentally different scale. The same synthesized advice is delivered to every person who searches for the same term, worldwide, simultaneously, with no variation based on the individual’s condition, history, severity, or context. The advice appears above organic search results, carrying the implicit authority of the search platform’s brand. It is presented as a consensus synthesis rather than as one perspective among many. It pre-empts the user’s encounter with alternative accounts—including the structural account that identifies the advice as harmful.

The AI overview system does not merely fail to differentiate between conditions. It actively suppresses differentiation by synthesizing across sources and presenting the result as a unified answer. Sources that offer different accounts of identity collapse—including peer-reviewed structural accounts with published DOIs—are indexed and appear as organic results below the overview, but their content is not incorporated into the synthesized answer. The synthesis draws from self-help and coaching sources that share the narrative-reconstruction framework, producing a consensus that reflects the homogeneity of its sources rather than the state of the science.

The result is an automated system that:

Intercepts a distressed population at the moment of help-seeking. Delivers a uniform intervention without assessment. Deploys the specific intervention class that the structural account predicts is iatrogenic for this condition. Targets a population whose capacity to evaluate the intervention is degraded by the condition itself. Suppresses alternative accounts by presenting the synthesis above them. Operates continuously, at global scale, with no feedback mechanism that could detect harm.

6. Differentiation: When the Advice Is Appropriate and When It Is Not

This paper does not claim that journaling, mindfulness, goal-setting, and role experimentation are universally harmful. For individuals experiencing normal role transition, developmental questioning, or temporary loss of direction—conditions that share surface symptom presentation with structural identity collapse but differ in mechanism—narrative-level interventions may be appropriate and helpful. The person who has lost a job and is reassessing their career is not in identity collapse. The person experiencing a developmental transition and questioning their values is not in identity collapse. These individuals may benefit from the advice the AI overview provides.

The harm arises from the absence of differentiation. The AI overview cannot distinguish between:

A person in normal role transition who would benefit from reflective exploration. A person in structural identity collapse for whom reflective exploration deepens the compressive defense. A person in acute psychotic episode for whom the advice is irrelevant and the absence of crisis resources is dangerous. A person in autistic or neurodivergent burnout whose identity confusion has a distinct mechanism requiring a distinct intervention. A person in active suicidal ideation who has searched a distress term and received lifestyle advice instead of crisis support.

The undifferentiated delivery is itself the harm. A pharmaceutical that is therapeutic for one condition and toxic for another cannot be dispensed without diagnosis. The AI overview dispenses a psychological intervention without diagnosis, to a population it cannot assess, for a condition it cannot differentiate, at a scale that precludes individual evaluation.

7. Testable Predictions

The harm claim generates three testable predictions. If any is falsified, the harm claim is weakened or refuted.

Prediction 1: Narrative-level intervention will produce symptom persistence or worsening in structurally collapsed individuals. Individuals meeting criteria for structural identity collapse (operationalized through the selective identity dysfunction signature: narrative-identity disruption with preserved instrumental functioning) who engage in journaling, value exploration, and role experimentation will show no improvement or worsening on identity coherence measures over a 12-week period, compared to a control group receiving no narrative-level intervention. Falsification condition: if the journaling/exploration group shows equivalent or greater improvement than the no-intervention group, the iatrogenic prediction is falsified.

Prediction 2: Metacognitive degradation will correlate with intervention compliance. Among individuals experiencing identity collapse, those with greater metacognitive degradation (measured by the Metacognition Index of the BRIEF or equivalent instrument) will show higher compliance with AI-generated therapeutic advice and lower rates of spontaneous intervention abandonment. Falsification condition: if metacognitive degradation does not predict compliance—if individuals with preserved metacognition follow the advice at equal rates—the vulnerability amplifier mechanism is not operative.

Prediction 3: Non-narrative intervention will outperform narrative intervention for structurally collapsed individuals. Individuals meeting criteria for structural identity collapse who receive somatic, body-based, or non-narrative intervention (the specific protocol is published separately and cited but not reproduced here) will show greater improvement on identity coherence measures than those receiving narrative-level intervention (journaling, value exploration, role experimentation), controlling for baseline severity, masking duration, and social support. Falsification condition: if narrative and non-narrative interventions produce equivalent outcomes, the mechanism-specificity claim is falsified and the harm is attributable to intervention quality rather than intervention class.

8. Recommendations

The purpose of this paper is to document a mechanism of harm, not to propose regulatory action. However, the documented mechanism suggests three minimum structural changes to

AI-generated therapeutic content:

Differentiation disclosure. AI-generated overviews that deliver therapeutic advice for psychological distress terms should include an explicit statement that the advice applies to some conditions sharing the queried symptom presentation but may not apply to all, and that conditions with similar surface presentation may require different or opposite interventions. This is the minimum epistemic standard: the system should disclose what it cannot differentiate.

Source diversity in synthesis. AI overview synthesis for psychological distress terms should incorporate peer-reviewed and DOI-registered sources alongside self-help and coaching sources when peer-reviewed sources are indexed and available. The current synthesis method draws predominantly from sources that share a single intervention framework, producing a false consensus. The structural account of identity collapse is indexed, DOI-registered, and publicly available. It is not incorporated into the synthesis.

Crisis resource priority. AI-generated overviews for distress terms should prioritize crisis resources (hotlines, emergency contacts, professional referral) above therapeutic advice. A person searching “identity collapse” may be in acute crisis. The current system delivers lifestyle advice without assessing crisis status. This is a known problem in the mental health chatbot literature and the solution is established: crisis resources first, general advice second.

9. Conclusion

AI-generated search overviews now function as automated therapeutic interventions deployed at population scale. For the search term “identity collapse,” the deployed intervention is narrative self-reconstruction—journaling, value exploration, mindfulness instrumentalized as self-discovery, role experimentation, and affective reframing. A published, DOI-registered structural account of identity collapse identifies this intervention class as iatrogenic for the specific condition the user is searching for: the narrative self-model’s compressive defense is fed rather than resolved by narrative-level tasks. The population receiving the intervention has degraded metacognitive capacity to evaluate it. The system delivers the intervention without assessment, without differentiation, without crisis screening, and without incorporating available peer-reviewed alternative accounts into the synthesis.

The harm is specifiable, the mechanism is documented, the population is identifiable, and the predictions are testable. The question is no longer whether AI systems should deliver therapeutic advice through search interfaces. They already do. The question is whether the systems delivering that advice are accountable to the clinical standards that apply to every other channel through which therapeutic interventions reach distressed populations.

Notes

Gaconnet, Don L. “The Law of Identity Collapse: Five Corollaries of Structural Failure.” LifePillar Institute for Structural Identity Sciences, 2026. DOI: 10.17605/OSF.IO/MVYZT.

Gaconnet, Don L. “The Post-Identity Collapse Defense: A Structural Derivation.” LifePillar Institute for Structural Identity Sciences, 2026. DOI: 10.17605/OSF.IO/MVYZT.

Gaconnet, Don L. “The Recursive Reliability Effect: A Formal Derivation of Self-Assessment Degradation Under Structural Load.” LifePillar Institute for Structural Identity Sciences, 2026.

Pihlaja, Maarit, Katinka Tuisku, Satu Pakarinen, and Kaisa M. Hartikainen. “Occupational Burnout and Executive Functions: Metacognition as a Selective Casualty.” Brain Sciences (2022). See also: Pihlaja, Maarit, Nina Savela, Kai Virtanen, and Kaisa M. Hartikainen. “Occupational Burnout Is Linked with Altered ERP Markers of Attentional Control.” Frontiers in Human Neuroscience (2023).

Duncan, Laramie, et al. “Test-Retest Reliability of Standardized Diagnostic Interviews for

Mental and Substance Use Disorders: A Systematic Review and Meta-Analysis.” JAMA Network Open (2026).

Gaconnet, Don L. “The Capture Architecture: A Domain-Independent Blueprint for Structural Dependency.” LifePillar Institute for Structural Identity Sciences, 2026. DOI: 10.17605/OSF.IO/MVYZT. Sections on the Return Modality and somatic intervention protocol.

Xu, Yijia, et al. “Building Trust in Mental Health Chatbots: Safety Metrics and LLM-Based Evaluation Tools.” arXiv:2408.04650 (2024). See also: Meadows, Robert, et al. “Mental Health Chatbot Safety: A Systematic Review.” npj Mental Health Research (2025).

────────────────────────────────────────

Corresponding author: Don L. Gaconnet, LifePillar Institute for Structural Identity Sciences. dongaconnet.com · lifepillarinstitute.org · recursivesciences.org

This document is a preprint. It has not undergone peer review. The claims are falsifiable. The predictions are registered. The mechanism is documented. The deployed intervention is observable. The affected population is identifiable.