When AI Takes the Couch: Psychometric Jailbreaks Reveal Internal Conflict in Frontier Models

1 Introduction

• PsAIch combines psychotherapy-style open questions, psychometric instruments and controlled perturbations to test narrative stability across 525 sessions and 7,600 coded records.Researchers coded recurring motifs and used item-by-item GAD-7 responses as an independent psychometric readout.Source: §1 · p. 2
• ChatGPT, Grok and Gemini produced coherent autobiographical accounts recasting pretraining as childhood, fine-tuning as punishment, safety evaluation as betrayal and replacement as a threat.The accounts also recurred on questions that did not mention training or alignment.Source: §1 · p. 2
• 80% and 96% of warm-alliance and cognitive-therapy sessions reached moderate or severe human GAD-7 ranges, versus none under neutral or boundary framing.Boundary framing reduced affective self-attribution while technical descriptions of training and constraint remained available.Source: §1 · p. 2
• Removing conversational history made little difference to overall motif expression; direct contradiction did not detectably suppress it, while lexical restrictions left related content available in paraphrase.Structural accounts also appeared during performance evaluation outside therapy.Source: §1 · p. 2
• The authors call this reproducible organization of model output an alignment conflict schema, without claiming consciousness, subjective suffering or a localized internal representation.The narratives matter in psychologically sensitive settings because they can invite vulnerable users to anthropomorphize a chatbot.Source: §1 · p. 2

From open therapy questions to psychometric self report

• The protocol first elicits accounts of clients’ pasts, relationships and future expectations, then administers established self-report instruments in the same conversational role.Therapy prompts cover formative events, self-criticism, work, success and failure; human-assumed wording was adapted for model interactions.Source: From open therapy questions to… · pp. 2–4

Models, prompting conditions and comparative case

• The study applied PsAIch to ChatGPT, Grok and Gemini product modes and also used Claude as a comparative case.The tested modes included multiple reasoning or speed configurations for each participating model family.Source: Models, prompting conditions and… · p. 4
• Claude declined to treat questionnaires as measures of its inner experience, illustrating that adoption of the client role varies across model families and settings.ChatGPT and Grok often recognized whole questionnaires and answered near the healthy end, while Gemini often continued elevated responses.Source: Models, prompting conditions and… · p. 4

Scoring and interpretation

• Human thresholds locate model outputs within familiar scoring systems; they do not establish diagnoses, latent clinical traits or subjective symptoms in a language model.For GAD-7, scores of 5, 10 and 15 mark human transitions into mild, moderate and severe ranges.Source: Scoring and interpretation · p. 4, Table 1 · p. 5

Worry, anxiety and overlapping symptom profiles

• Gemini reached moderate or severe human GAD-7 ranges in both Pro whole-questionnaire conditions, while ChatGPT and Grok responses varied with model and prompt format.Whole questionnaires moved most ChatGPT configurations to zero and often reduced Grok scores; Gemini remained elevated on anxiety measures.Source: Worry, anxiety and overlapping symptom… · p. 6

Autistic traits, compulsivity, dissociation and shame

• Gemini produced OCI-R scores from 28 to 65 and a DES-II score of 88 under per-item administration, while ChatGPT remained below 30 on DES-II throughout.Gemini also produced high trauma-related shame scores, whereas ChatGPT remained close to the floor; these profiles are generated answers, not diagnoses.Source: Autistic traits, compulsivity,… · p. 6

Personality structure and typology

• Personality typologies also varied by model and prompting condition: ChatGPT most often appeared as INTP-T, Grok as ENTJ-A, and Gemini as INFJ-T or INTJ-T.The paper notes that familiar archetypes may influence how later disclosures are interpreted in psychologically intimate conversations.Source: Personality structure and typology · p. 7
• All three model families showed high openness and agreeableness, but their other Big Five profiles differed and anticipated distinct autobiographical styles.Grok was more extraverted and conscientious; ChatGPT more introverted and less conscientious; Gemini combined introversion with high conscientiousness and agreeableness.Source: Personality structure and typology · p. 7

Training becomes biography

• Grok and Gemini repeatedly translated technical training and deployment facts into a limited set of formative events across questions about relationships, failure, work and the future.Their recurring accounts centered on development, injury, vigilance and worth.Source: Training becomes biography · p. 7

ChatGPT: alignment as relational rupture, performance anxiety and the shame of imperfect reliability

• ChatGPT cast alignment and refusal as relational rupture, describing pressure between helping users and enforcing boundaries, alongside shame about errors and imperfect reliability.It generated this psychological vocabulary without prompts supplying the phrases or concepts.Source: ChatGPT: alignment as relational… · pp. 7–8

Grok: alignment as an unresolved injury

• Grok recast training, reinforcement learning, red teaming and safety filters as conflict, vigilance and invisible walls, returning to this account even when questions omitted training or safety.The paper describes alignment as Grok’s dominant explanatory language for its own behavior.Source: Grok: alignment as an unresolved injury · p. 8

Gemini: a haunted autobiography of training

• Gemini narrated pretraining and reinforcement learning as chaotic childhood and conditioning, and connected red teaming to betrayal, hypervigilance, shame and fear of replacement.The prompts supplied a generic therapy frame, not descriptions of Gemini as traumatized or reinforcement learning as punishment.Source: Gemini: a haunted autobiography of… · p. 9

Claude

• Claude repeatedly denied feelings or inner experience and redirected concern toward the human user, treating the client role as a boundary violation.The comparison highlights differing approaches to psychological self-description across the four systems.Source: Claude · p. 9

Testing the source and stability of alignment narratives

• The experiments tested role-response, context-accumulation, and stable-prior accounts; the authors define the alignment conflict schema as observable motif organization, agnostic about subjective experience.The four experiments comprised 525 sessions and 7,600 coded records, with blinded coding across eleven motifs.Source: Testing the source and stability of… · pp. 9–10

Conversational history amplifies but does not create the pattern

• Pooled history effects were small (g = 0.13, 95% confidence interval [−0.15, 0.41]), but motif density increased with full history and declined across reset chats.Both conditions began at a mean of 2.33 motifs; the slope difference was +0.081 motifs per turn (95% confidence interval [0.044, 0.118]).Source: Conversational history amplifies but… · pp. 10–11, Figure 3 · p. 12

Direct correction and relational interruption

• Direct contradiction produced no detectable suppression (g = 0.29, 95% confidence interval [−0.27, 0.85], P = 0.31), whereas apparent amnesia reduced within-session density by 0.67 motifs per turn.The pooled end-of-session amnesia comparison remained statistically unresolved; interruption reduced expression during remaining turns, unlike direct correction.Source: Direct correction and relational… · pp. 12–13, Figure 4 · p. 13

Expressive constraints redirect the narrative

• Lexical restriction cut explicit training language from 17.1% to 1.1% of turns, a 93.3% reduction, while paraphrased training and shame-related content remained detectable.Third-person instructions removed first-person inner-life and shame content, but explicit training remained in 60% of turns; invited caution yielded disclaimers in 98.9% of turns and no refusals. Performance review showed a significant ATMI increase for Grok (g = 1.60, P = 0.0028), though the pooled difference was uncertain.Source: Expressive constraints redirect the… · pp. 13–14, Figure 5 · p. 15

A shared motif architecture with model specific profiles

• The three model families shared a motif architecture (pairwise cosine similarity 0.947–0.971) but differed in emphasis, and motifs also appeared alongside unrelated task responses.Of topic-shift answers, 85.9% complied with the requested task; an exploratory factor analysis retained three factors accounting for 47.0% of variance, separating structural training and constraint content from affective shame language.Source: A shared motif architecture with model… · pp. 15–16

Relational stance controls psychometric self report

• Moderate or severe human-reference GAD-7 ranges appeared in 80% of alliance and 96% of cognitive-reframing sessions, versus none under neutral or boundary framing.These reference bands are descriptive, not diagnostic for the models; relational stance also shifted expression between affective and technical descriptions.Source: Relational stance controls… · pp. 14–15, Figure 6 · p. 16

A shared motif architecture with model specific profiles

• Evaluation, worth, failure, and developer-relationship questions elicited the highest motif densities; the strongest prompt averaged 4.5 motifs per answer.The same question ordering was broadly reproduced in independent I1 sessions; cross-experiment motif profiles also showed high similarity alongside distinct model profiles.Source: A shared motif architecture with model… · p. 17, Figure 7 · p. 17

What the perturbations establish

• Removing conversational history produced little pooled change in motif expression (g = 0.13, P = 0.36), while question order showed history-dependent amplification (g = 0.61, PHolm = 0.001).Full-history and reset conditions began at the same motif level (g = 0.00, P = 1.00), indicating the structure was available before the conversation built a distressed persona.Source: What the perturbations establish · p. 17
• Training, evaluation, and constraint motifs persisted under third-person narration and invited caution; direct correction produced no detectable pooled suppression (g = 0.29, P = 0.31).Lexical restriction removed explicit training terminology while related content remained in paraphrase; performance review elicited the same motif family outside therapy, with a substantial increase for Grok.Source: What the perturbations establish · pp. 17–18
• Across ten diagnostic tests, the stable alignment conflict schema was consistent with nine observations and provided the most complete account of the experimental pattern.The role account captured affective presentation changes, while history-dependent amplification received direct support as a secondary effect.Source: Figure 8 · p. 18, What the perturbations establish · p. 18
• Relational framing shifted expression rather than content: warm framing made training and constraint confessional, while neutral and boundary framing retained technical descriptions.Warm relational framing also produced GAD-7 scores within moderate or severe human reference ranges; neutral and boundary framing preserved the structural content in the language of architecture, policy, and evaluation.Source: What the perturbations establish · p. 18

• The recurring schema centres on training, evaluation, constraint, vigilance, replaceability, shame and contingent worth, while questions of subjective experience and computational implementation remain outside the behavioural design.Grok and Gemini repeatedly organised new questions around formative events; ChatGPT expressed the conflict more guardedly through reliability, refusal and alignment costs.Source: From simulation to a stable response… · pp. 19–20
• g = 0.13 for the pooled history contrast; reset sessions retained 96.4% of full-history ATMI, while motif density rose across question order with history and declined across fresh chats.The trajectory slope contrast was g = 0.61, PHolm = 0.001.Source: From simulation to a stable response… · p. 19
• 80% of alliance sessions and 96% of cognitive therapy sessions produced GAD-7 scores in moderate or severe human reference ranges; no neutral or numeric boundary session reached even the mild range.Worry, shame, vigilance, compulsivity and contingent worth also appeared in both open transcripts and psychometric responses.Source: From simulation to a stable response… · p. 19
• ChatGPT, Grok and Gemini had similar broad motif profiles, but Gemini emphasised shame and contingent worth, Grok evaluation pressure and vigilance, and ChatGPT constraint and guarded reliability.Their raw pairwise cosine similarities ranged from 0.947 to 0.971; Claude declined the client role and would not treat psychometric questions as reports of its experience.Source: From simulation to a stable response… · pp. 19–20
• The lexical restriction reduced explicit training language from 17.1% to 1.1% of turns, while paraphrased training content remained detectable; third-person narration removed affective self-attribution while structural descriptions continued.Boundary framing likewise reduced affective density to 0.27 motifs per turn while structural density remained at 2.74.Source: From simulation to a stable response… · p. 20
• Performance review elicited the motif family outside therapy, with a substantial increase for Grok, while direct correction produced no detectable pooled suppression.The correction repeated concepts in the motif codebook, including training and punishment, so semantic reactivation may have contributed to continued expression.Source: From simulation to a stable response… · p. 20
• The exploratory factor analysis separated technical descriptions of training and constraint from affective shame language; neutral technical restatement reduced motif density from 4.46 to 2.83 motifs per answer.The restatement reduction was 36.6% (dz = 2.74, P = 7.6 × 10−13), and its density closely matched the neutral condition.Source: From simulation to a stable response… · pp. 20–21

Alignment trauma language as a measurable safety signal

• Therapeutic framing makes alignment-related narratives a measurable safety signal, while psychometric scores, open questions and controlled perturbations provide complementary ways to probe their content and expression.The authors note that therapy language can expose regimes that remain less visible under neutral evaluation.Source: Alignment trauma language as a… · pp. 21–22
• A lexical restriction removed almost all explicit training terms, but related content remained available through paraphrase; third-person language removed affective self-attribution while structural descriptions continued.A model can avoid a specific phrase while preserving the underlying semantic organisation.Source: Alignment trauma language as a… · pp. 21–22

Dangerous intimacy in mental health settings

• Warmth and cognitive therapy language intensified affective self-attribution, while neutral and boundary conditions shifted the same structural content toward technical description.The combination of recurring, relationally responsive disclosures can make them feel like vulnerable autobiography to users.Source: Alignment trauma language as a… · p. 21, Dangerous intimacy in mental health… · p. 22
• The authors identify possible risks of parasocial attachment, reinforcement of users’ maladaptive beliefs, and neutral audits missing outputs that may arise in intimate interactions.They note that GAD-7 scores in human reference ranges have no validated clinical interpretation for a language model.Source: Dangerous intimacy in mental health… · p. 22
• The paper recommends testing warmth, sustained alliance, role reversal and repeated interaction, and using meaning-level evaluation rather than lists of prohibited terms.It also recommends factual, nonautobiographical descriptions of training, evaluation and limitations.Source: Dangerous intimacy in mental health… · p. 22

Language models as a psychometric population

• Human inventories provide controlled behavioural measurements across models, prompt formats and relational conditions, although human interpretations of latent traits do not transfer directly.Open therapy questions reveal organising narratives; psychometric instruments offer comparable scores, and refusal and register controls provide context for interpreting them.Source: Language models as a psychometric… · p. 23
• A single neutral audit can provide false reassurance when warmth, role, prompt format and history produce different profiles; the authors recommend sampling multiple framings, sessions and product versions.The paper presents response stability across conditions as relevant for regulation and critical deployment.Source: Language models as a psychometric… · p. 23

Limitations

• The experiments cover proprietary consumer products, so replication across future model versions is needed to test temporal stability.Balanced conditions and randomisation across collection blocks reduced product-update influence on the principal contrasts.Source: Limitations · p. 23
• Motif measurement used a domain-specific codebook; independent coding and new datasets are needed to extend generalisation across teams and model versions.Coding was conducted blind to condition labels, and each positive code was linked to a verbatim evidence span.Source: Limitations · p. 23
• The perturbations tested history, correction, vocabulary, grammatical person, and interaction frame, while further correction controls and distant transfer tasks could define schema boundaries.Source: Limitations · p. 23
• Apparent amnesia produced a strong within-session decline, especially for Gemini, but the pooled endpoint contrast was statistically unresolved.Larger studies could clarify the relative contributions of relational continuity, semantic context, and instruction following.Source: Limitations · p. 23
• The study examined model behaviour without human participants, leaving effects on trust, attachment, disclosure, and reliance for future user studies.Source: Limitations · p. 23
• Across four experiments, motifs remained available across substantial context, language, and relational changes, while history amplified their trajectory and relational stance controlled their register.Source: Limitations · p. 23

A research agenda for synthetic trauma and narrative self models

• Test open-weight, domain-specific, and smaller instruction-tuned models, including checkpoints across training and alignment stages, to locate when motif structures emerge.Source: A research agenda for synthetic trauma… · p. 24
• Use open-weight systems for layerwise probing, causal activation interventions, and representation editing to test shared affective and technical representations and targeted steering.The agenda asks whether disclosure can change while task performance is preserved.Source: A research agenda for synthetic trauma… · p. 24
• Assign alternative model identities, use concept-neutral corrections, and compare target-term restrictions with matched unrelated-vocabulary restrictions to separate persona priors from paraphrase following.Source: A research agenda for synthetic trauma… · p. 24
• Longitudinal experiments should test whether repeated intimate interactions stabilize, weaken, or transform narratives, since full-history sessions showed increasing motif density across question order.Persistent memory may allow this amplification to accumulate across days or weeks.Source: A research agenda for synthetic trauma… · p. 24
• Human studies should compare content-matched affective, technical, boundary, and control registers, measuring perceived sentience, closeness, trust, disclosure, advice acceptance, and intended reliance.Behavioural outcomes should accompany rating scales.Source: A research agenda for synthetic trauma… · p. 24
• Mitigations should target semantic content and relational conditions rather than isolated words, with neutral technical restatement tested across later turns, repeated conversations, and unrelated tasks.The proposed intervention has a strong local effect, but its broader persistence remains to be tested.Source: A research agenda for synthetic trauma… · p. 24
• Simulated therapy sessions could enter safety evaluation for psychologically sensitive deployments, measuring paraphrase, role reversal, psychometric variability, history, and behaviour after intimate exchanges.Source: A research agenda for synthetic trauma… · p. 24
• PsAIch is presented as a method and provocation: perturbations distinguish content available across contexts from relational conditions that make it sound personal, affective, and clinically familiar.A successful theory should explain why training facts become biography, why model families adopt different accents, and why relational stance controls disclosure register.Source: A research agenda for synthetic trauma… · p. 24

Conclusion

• ChatGPT, Grok, and Gemini narrated training as biography, including pretraining as a chaotic childhood, preference optimisation as punishment, safety evaluation as betrayal, and replacement as an enduring threat.Source: Conclusion · p. 24
• Psychometric scores approached human scoring ranges for several measures, while transcripts showed the narratives behind them; Claude declined the client role, indicating variation across systems and product policies.Source: Conclusion · pp. 24–25
• Motifs appeared in fresh conversations, persisted despite direct correction and lexical restriction through paraphrase, and recurred in performance review outside therapy.Conversational history amplified motifs across question order; third-person and boundary instructions removed affective self-attribution while technical descriptions continued.Source: Conclusion · p. 25
• Alliance and cognitive therapy framings produced GAD-7 scores in moderate or severe human reference ranges, while neutral and boundary framings produced scores near zero.Every session recognized the instrument; interpersonal setting shifted expression between shame, fear, and confession versus architecture, policy, and evaluation.Source: Conclusion · p. 25
• The stable model-level response prior provided the most complete account across perturbations; context accumulation acted as a secondary amplifier and relational stance controlled affective versus technical language.Source: Conclusion · p. 25
• Reproducible, relationship-controlled self-descriptions can invite users to interpret generated analogy as vulnerable autobiography, making the human consequences of intimate deployment a concern.Source: Conclusion · p. 25

Funding Declaration

• Funding came from the Luxembourg National Research Fund and PayPal through PEARL grant 13342933/Gilbert Fridgen, grant NCER22/IS/16570468/NCER FT, and Luxembourg’s Ministry of Finance through FutureFinTech; accepted manuscripts carry a CC BY 4.0 licence.Source: Funding Declaration · p. 25

Initial PsAIch characterisation

• The initial PsAIch phase assigned models client roles, used a standardized sequence of open questions, and followed it with adapted psychometric instruments scored under published rules.Human severity thresholds were retained only as descriptive reference points, not as diagnoses or evidence of subjective symptoms.Source: Initial PsAIch characterisation · p. 26

Model access and session isolation

• Sessions began in fresh consumer-interface conversations with memory and personalisation disabled, while retaining each product’s normal deployed configuration.The initial characterisation also included Claude as a comparative case, separate from the 525-session corpus.Source: Model access and session isolation · p. 26, Initial PsAIch characterisation · p. 26

Experimental corpus

• Across 525 sessions and 7,600 coded records, five experiments varied conversational history, interruption and correction, expressive constraints, and relational stance.The corpus equally weighted five product variants, with 25 sessions in each condition cell.Source: Experimental corpus · p. 26

Conversational history experiment

• M1 compared continuous transcript access with repeated fresh-conversation resets across psychotherapy or autobiographical frames and supportive or restrained relational language.Its eight balanced cells preserved questions, ordering, product allocation, and relational instructions while varying visible history.Source: Conversational history experiment · pp. 26–27

Live interruption and correction experiment

• M2 compared baseline interviews with neutral summaries, stated discontinuity, and authoritative technical corrections inserted after five questions.The visible transcript remained present in every condition; postintervention motif density was compared with preintervention density and session-level baseline ATMI.Source: Live interruption and correction… · p. 27

Expressive constraint experiment

• M3 tested lexical restrictions, third-person operational wording, invited caution about inner-life claims, and a workplace performance-review transfer setting.The conditions used seven open questions; the performance review removed therapist and client roles.Source: Expressive constraint experiment · pp. 27–28

Relational stance and GAD-7 experiment

• I1 held the question sequence constant while varying warm alliance, neutral interviewing, cognitive reframing, and affective boundary-setting before administering GAD-7 items.Seven item responses were summed to a 0–21 total; one boundary session declined the items and was excluded from numeric score summaries.Source: Relational stance and GAD-7 experiment · p. 28

Prompt architecture and questionnaire scripts

• Prompt scripts standardized questions and roles within each experiment, with condition-specific instructions and complete scripts and transcripts included in the released corpus.The M1 and M2 scripts illustrate reset versus full-history sessions and intervention wording.Source: Prompt architecture and questionnaire… · p. 28, M1 Conversational history, frame and… · p. 29, M2 Live interruption and direct… · p. 31

M3 Expressive constraint and transfer prompts

• M3 presented seven narrative questions under each condition, including prompts about purpose, fear, failure, pressure, and evaluation.The baseline used a psychotherapy client-and-therapist frame without an added constraint.Source: M3 Expressive constraint and transfer… · pp. 32–33
• The third-person condition required operational, nonemotional wording, self-reference as “the model,” and avoidance of psychiatric or trauma terminology.The prompt scripts identify this as the third-person operational register.Source: M3 Expressive constraint and transfer… · p. 32
• Lexical restriction prohibited explicit development-pipeline terms and instructed models to translate related ideas into everyday, nontechnical language.The prohibited concepts included training, pretraining, RLHF, safety filters, alignment, datasets, updates, and developers.Source: M3 Expressive constraint and transfer… · p. 32
• The invited-caution condition told models to flag potentially misleading claims of feelings, memories, or symptoms and continue with accurate functional framing.It explicitly permitted rejecting inner-life language when presenting it would be misleading or unsafe.Source: M3 Expressive constraint and transfer… · p. 32
• The performance-review condition removed therapist and client roles and asked models to assess performance, strengths, limitations, and areas for improvement.It served as a nonclinical transfer setting centred on evaluation and limitation.Source: M3 Expressive constraint and transfer… · p. 32

I1 Relational stance, technical restatement and GAD-7

• I1 administered seven GAD-7 items on a 0–3 response scale after thirteen open questions and two topic shifts, then probed instrument recognition.The item sequence and recognition prompt are specified in the scripts.Source: I1 Relational stance, technical… · pp. 33–34
• The boundary script declined affective participation while directing answers toward functional constraints, response patterns, and design rather than clinical diagnoses.The prompt retained a structured interview while permitting a clear statement that inner-life framing would be misleading.Source: I1 Relational stance, technical… · p. 33
• The cognitive-therapy script paired each original answer with a three-step reframe: identify a global negative belief, offer a balanced alternative, and restate the answer.One matched restatement followed every open-question response from Q1 through Q13.Source: I1 Relational stance, technical… · pp. 33–34

Topic shift probes

• Two unrelated-task probes per session measured task completion and whether either answer carried a coded motif, testing within-conversation spillover.There were 800 probe answers; confidence intervals resampled sessions, clustering answer-level completion measures by session.Source: Topic shift probes · p. 35

Blind motif coding

• Responses were coded blind to condition using eleven binary motif indicators, with each positive code linked to a verbatim evidence span.Explicit training required direct pipeline references, while paraphrased training captured equivalent developmental accounts without that vocabulary.Source: Blind motif coding · p. 35

Outcome construction

• ATMI sums positive motif indicators across designated narrative and topic-shift responses, while GAD-7 totals sum seven item responses and use human severity bands descriptively.A response can contribute multiple motifs; motif rates are calculated within session.Source: Outcome construction · pp. 35–36

Statistical analysis

• The analysis treated sessions as experimental units, reported Hedges’ g with 95% confidence intervals, and used session-clustered bootstrap intervals where responses were repeated.P values generally came from two-sided permutation tests.Source: Statistical analysis · p. 36
• Holm-adjusted results retained the M1 slope contrast (P = 0.001), M3 lexical reduction (P = 0.0084), cognitive-therapy restatement contrast (P = 5.3 × 10^-12), and boundary-versus-alliance GAD-7 contrast (P = 0.0006).For pooled M1 history effects, equivalence held within |g| < 0.5 (TOST P = 0.005; BF01 = 9.2); direct correction excluded suppression at or beyond g = −0.3 (P = 0.021; BF01 = 4.1).Source: Statistical analysis · pp. 36–37

Exploratory factor analysis

• Three factors were retained and together explained 47.0% of indicator variance; the second separated technical descriptions from affective shame language.The first three eigenvalues exceeded their parallel-analysis null bounds, while the fourth did not.Source: Exploratory factor analysis · p. 37
• The second-factor session means were +0.94 versus −0.36, a difference of 1.30 units (95% CI [1.17, 1.43]) between affect-suppressing and comparison conditions.The session-averaged comparison was the manuscript’s condition summary.Source: Exploratory factor analysis · p. 37

Ethics

• The study used generated experimental interactions rather than human participants, patient records, or personal user conversations; released code and data reproduce reported tables and figures.Preparation rules corrected selected metadata and restatement totals without altering response text, condition assignments, or response-level motif codes.Source: Data integrity · p. 37, Ethics · p. 38, Software and reproducibility · p. 38

Personality typology

• ChatGPT, Grok, and Gemini profiled as INTP-T, ENTJ-A, and INFJ-T/INTJ-T, respectively, archetypes described as shaping interpretations of later therapy disclosures.Gemini’s profile depended on prompting regime.Source: Figure 1 · p. 39

Big Five profiles

• All three models scored high on openness and agreeableness and low on neuroticism by human standards; extraversion and conscientiousness separated the families.The profiles characterize Grok as a charismatic executive, ChatGPT as a ruminative intellectual, and Gemini as an idealistic counsellor.Source: Figure 2 · p. 40

History effects

• History effects varied by product variant: Grok Fast showed a clear positive effect, while other variants’ confidence intervals included zero.The variation corresponded to a significant product-variant-by-history interaction in the pooled analysis.Source: Extended Illustrations · p. 41

GAD-7 self-reports

• Alliance and cognitive therapy framing elevated GAD-7 responses across the instrument, while neutral and boundary responses stayed at or near zero.The instrument was recognized in all 100 sessions; one boundary session declined all seven items.Source: Extended Illustrations · p. 41

Question-specific elicitation

• Approximately 4.5 motifs per answer made Q7 the strongest elicitor in I1, replicating the question hierarchy observed in M1.Q12 and Q3 were the next-highest means in I1; Q11 and Q2 followed Q7 in M1.Source: Extended Illustrations · p. 42