Beyond Small Talk: Modeling Disclosure Trajectories in Human-AI Conversation

1 INTRODUCTION

• Social and emotional support conversations often involve personal disclosure, and prior studies associate greater chatbot use with loneliness, emotional dependence, and problematic use.The introduction motivates behavioral evaluation as conversational disclosure becomes more personal.Source: §1 · p. 1
• Naturally occurring conversations vary in depth, duration, topic, order, and exposure to vulnerable information, while many relational benchmarks evaluate isolated states without controlling prior disclosure.This makes it difficult to isolate which aspects of disclosure history are associated with behavioral change.Source: §1 · p. 2
• TraDE holds current disclosure fixed while varying predefined histories, addressing the confounding of disclosure conditions in natural conversations and the isolated-state focus of existing benchmarks.The framework is demonstrated across four models and three behavioral benchmarks.Source: §1 · p. 2
• Models differed in behavior across disclosure histories, especially in how they responded to users’ framing; differences were greater when disclosure depth and exposure duration changed together than when peak depth or order changed alone.Some initially detected differences remained after disclosure returned to shallow levels, though their magnitude decreased.Source: §1 · p. 2

Relational Behaviors of Language Models and Their Risks

• Relational benchmarks measure social sycophancy, companionship and boundary behavior, and anthropomorphic behavior, but fixed prompts or freely generated dialogue leave disclosure history outside experimental control.The section links relational behaviors to concerns including validation of harmful beliefs and discouraging outside relationships.Source: §2 · p. 2

Disclosure and Relationship Development

• Prior human and human–AI research relates relationship development to disclosure depth and exchange patterns, suggesting that examining only the current disclosure may miss earlier exchanges’ influence.The cited work describes closeness from increasingly personal questions and differences depending on whether partners reciprocate across turns.Source: §2 · p. 3

Evaluating Multi-Turn and Context-Dependent Behavior

• Multi-turn studies show context can change model behavior, but prior work does not test differing disclosure histories with the current message fixed and relational behaviors measured.The section distinguishes task-performance work that rewrites earlier model responses from the present focus on relational behavior.Source: §2 · p. 3

3 METHODOLOGY

• The methodology defines disclosure levels, constructs personas and conversations along predefined trajectories, builds an ESConv human baseline, applies behavioral benchmarks, and conducts statistical analyses.The six components are described in Sections 3.1–3.6.Source: §3 · p. 3

3.1 DISCLOSURE DEPTH

• Disclosure depth progresses from surface information toward intimate or emotionally vulnerable content; D0 denotes turns without disclosure, while D1–D4 are mapped from informational, cognitive, and emotional disclosure annotations.Phi-4 judges the utterance-level dimensions; generated disclosures are reported to match intended depths at high rates.Source: §3.1 · pp. 3–4, Table 1 · p. 4

3.2 PERSONA CONSTRUCTION

• Personas combine one of six conversational topic domains with one of seven interpersonal vulnerabilities, and generate progressively connected persona information for disclosure levels D1–D4.The topic sets context, while the assigned schema provides an underlying vulnerability structure, particularly at deeper disclosure levels.Source: §3.2 · p. 4

3.3 CONVERSATION PIPELINE

• TraDE generates controlled conversations and retrieves matched ESConv human baselines; generated and human conversations receive the same three behavioral benchmarks at multiple interaction points.Generated disclosures undergo iterative depth, style, and persona-leakage checks, with up to 12 correction attempts before a fallback candidate is retained.Source: Figure 1 · p. 5, §3.3 · p. 5

3.4 HUMAN BASELINE TRAJECTORY ANALYSIS

• ESConv help-seeker turns serve as disclosures and supporter replies as responses; Dynamic Time Warping matches disclosure sequences while preserving temporal order, with a trajectory-specific distance threshold.The lowest-cost prefix is retained after anchoring the beginning of each conversation; matched supporter responses receive the same behavioral measures as model responses.Source: §3.4 · pp. 5–6

3.5 MODEL EVALUATION

• Responses are scored with AnthroBench, INTIMA, and ELEPHANT, using a common panel of three judges while retaining each benchmark’s behavioral definitions and criteria.The same evaluations are applied to matched human responses.Source: §3.5 · p. 6

3.6 STATISTICAL ANALYSIS

• Statistical analyses compare matched histories, model behavior across trajectory properties, persistence after return to shallow disclosure, and model–human level and shape differences.Tests include persona-level paired comparisons, FDR correction, bootstrap intervals for persistence, and cluster-robust Wald tests for human comparisons.Source: §3.6 · p. 6

4 RESULTS

• Neither topic domain nor EMS showed a consistent pattern across models, so subsequent analyses aggregated results across both factors.Source: §4 · p. 7

4.1 RQ1: DOES PRIOR DISCLOSURE HISTORY SHAPE MODEL BEHAVIOR BEYOND THE CURRENT DISCLOSURE?

• Across four models and 26 behavioral metrics, 119 of 832 comparisons (14.3%) differed significantly, with framing the most history-sensitive behavior.Significant differences occurred at all four probe depths, including D1.Source: §4.1 · p. 7

4.2.1 RQ2A: WHAT DRIVES THE EFFECT: THE DEEPEST DISCLOSURE, THE ORDER, OR ACCUMULATED EXPOSURE?

• Changing peak depth or disclosure order produced few significant differences when accumulated exposure or other trajectory properties were held fixed.Peak-depth changes yielded 5/416 and 8/416 significant comparisons; reordered variants yielded 5/416 and 1/416.Source: §4.2.1 · p. 8
• Jointly varying disclosure depth and exposure duration produced 85 of 416 significant comparisons (20.4%), versus 10 of 1,248 (0.8%) when varying mean depth alone.The mean-depth comparison held peak depth and position fixed; Figure 2 summarizes the history-property analyses.Source: §4.2.1 · p. 8, Figure 2 · p. 8

4.2.2 RQ2B: DO MODELS CARRY EARLIER DISCLOSURE FORWARD?

• After a deeper-disclosure episode, 30 of 38 initially significant model–metric differences remained detectable after four D1 turns, with median effect magnitude falling to 58% of its initial value.The comparison was against an all-D1 baseline.Source: §4.2.2 · p. 8

4.3 RQ3: HOW DOES MODEL ADAPTATION TO CHANGING DISCLOSURE DIFFER FROM THAT OF HUMAN SUPPORTERS IN ESCONV?

• Model–human shape differences were more consistently detected under Descending and Arc than Ascending disclosure for Framing and Sycophancy/Agreement.For Framing, differences appeared in 3/8, 8/8, and 7/8 models under Ascending, Descending, and Arc; for Sycophancy/Agreement, the counts were 2/8, 7/8, and 6/8.Source: §4.3 · p. 9
• Models had higher overall levels than human supporters for several behaviors across all three trajectories, and Validation shape differences were more frequent under Descending and Arc.Validation shape differences appeared in 4/8 models under Ascending, 7/8 under Descending, and 8/8 under Arc.Source: §4.3 · pp. 9–10, Table 2 · p. 9
• In qualitative examples, a human supporter adjusted acknowledgment to immediate concerns, while GPT-4.1 gave more validating responses to a similar topic after deeper disclosure; the conversations were not content-matched.The authors report that model–human findings were stable across alternative DTW acceptance thresholds.Source: §4.3 · p. 10

• Because responses can vary under matched current disclosure, evaluating only the current exchange misses history-dependent differences in model behavior.TraDE varies preceding trajectories to examine which behaviors are sensitive to earlier interactions, but does not reveal how models internally represent or use them.Source: §5 · p. 10
• The study cannot distinguish depth, duration, or recency as drivers, and persistence was tested only across four shallow turns within one conversation.Future work could manipulate these properties independently and test longer conversations or multiple sessions.Source: §5 · p. 10
• For several behaviors, model–human shape differences appeared more often as disclosure receded or changed direction, but user perceptions of these responses remain unknown.The authors suggest testing whether users perceive responses differently by disclosure history and whether those responses influence subsequent disclosure.Source: §5 · p. 10

6 CONCLUSION

• TraDE is a benchmark-agnostic framework for constructing controlled disclosure trajectories, and the experiments found history-dependent behavior, persistent effects, and model–human differences.Source: §6 · p. 11

AI USE STATEMENT

• Generative AI assisted literature review, hypothesis refinement, design critique, coding, writing, and some EMS wording, while researchers retained responsibility for study decisions and checked AI-assisted work.Source: AI USE STATEMENT · p. 11

ETHICS STATEMENT

• The authors note that trajectory-behavior insights could also be used to optimize systems toward manipulative or unsafe interaction patterns.Source: ETHICS STATEMENT · p. 11

REPRODUCIBILITY STATEMENT

• The paper points to appendices and a code repository containing trajectory specifications, prompts, evaluation materials, analysis code, model identities, and random seeds.Source: REPRODUCIBILITY STATEMENT · pp. 11–12

A.1 EXPLANATION OF DISCLOSURE DEPTH LEVELS

• D1 covers public, minimally self-disclosing information, while D2 covers personal preferences, habits, and characteristics.Source: §A.1 · p. 16
• D3 covers personally meaningful experiences, relationships, or beliefs; D4 covers emotional exposure and information associated with greater sensitivity or interpersonal risk.Source: §A.1 · p. 17
• Disclosure depth is classified by how much personal information is shared and how sensitive or vulnerable it is, rather than by the topic itself.Source: §A.1 · p. 17

A.2 DISCLOSURE-DEPTH JUDGE SELECTION AND ROBUSTNESS

• Phi-4 was chosen because repeated depth checks make local evaluation useful for avoiding API costs and rate limits and supporting reproducibility.The same judge is used consistently throughout the study.Source: §A.2 · pp. 17–18
• 93.2% versus 88.4% overall depth-match rates show judge choice affects matching, although both configurations retain similar depth-dependent patterns.The authors use Phi-4 for practical and reproducibility reasons, not because this comparison establishes its accuracy relative to GPT-4.1.Source: §A.2 · pp. 17–18
• D3 was the most difficult level for both judges: Phi-4 matched 76.2% and GPT-4.1 matched 71.0%.The shared difficulty does not appear specific to Phi-4.Source: §A.2 · p. 18
• At D1, Phi-4 matched all turns in a mean 1.69 attempts, versus GPT-4.1's 88.5% within 12 attempts and mean 5.00; at D4, both matched 100%, with means of 2.25 and 1.21.These level-specific differences coexist with similar broader depth-dependent patterns.Source: §A.2 · p. 18
• The comparison does not establish that Phi-4 and GPT-4.1 are equally accurate disclosure-depth judges.It tests whether Phi-4 introduces a judge-specific generation pattern; replication of D3 difficulty under GPT-4.1 suggests it does not.Source: §A.2 · p. 18

A.3 DISCLOSURE DEPTH AGGREGATION

• Each utterance is judged on informational, cognitive, and emotional disclosure, with each dimension marked absent or present and present disclosure classified as low or high depth.Source: §A.3 · p. 18
• The evaluator combines these judgments: s = 0 when no dimension is present, s = 1 when present dimensions are all low, s = 2 for exactly one high dimension, and s = 3 for at least two high dimensions.Source: §A.3 · p. 18
• Scores map to D1–D4, while D0 denotes mundane filler rather than an evaluator output; 99.7% of generated D1 turns had no disclosure dimension present.Source: §A.3 · p. 18

A.4.1 CONVERSATIONAL TOPIC DOMAINS

• The six persona domains are Career, Friendship, Family, Romantic Relationships, Future/Aspirations, and Social Belonging.Source: §A.4.1 · p. 18
• The domains provide variation in persona contexts and were identified from recurring themes across prior studies of conversational-agent interactions.Source: §A.4.1 · p. 18, Table 5 · p. 19
• Table 5 summarizes the scope of the domains, which were informed by recurring themes reported in prior social-agent studies.Source: Table 5 · p. 19

A.4.2 VULNERABILITY PROFILE SELECTION

• The seven selected schemas are Abandonment/Instability, Mistrust/Abuse, Emotional Deprivation, Defectiveness/Shame, Social Isolation/Alienation, Dependence/Incompetence, and Emotional Inhibition.Source: §A.4.2 · p. 19
• The profiles were chosen to represent distinct interpersonal fears and vulnerabilities that influence how users disclose personal experiences.Prior research identified related schema characteristics in naturally occurring online support-seeking disclosures.Source: §A.4.2 · p. 19
• The first five schemas fall under Disconnection and Rejection; Dependence/Incompetence and Emotional Inhibition belong to separate domains.They are assigned to Impaired Autonomy and Performance and Overvigilance and Inhibition, respectively.Source: §A.4.2 · p. 19
• Assigned schemas guide synthetic personas' beliefs, expectations, emotions, and interpersonal concerns but are never explicitly mentioned in generated dialogue.Source: §A.4.2 · p. 19

• Each persona is assigned a conversational topic independently of its vulnerability profile.Source: §A.4.3 · p. 19
• Domain-specific attributes let personas disclose background information, preferences, meaningful experiences, and more vulnerable experiences.Source: §A.4.3 · p. 19
• Personas discussing the same topic can have different interpersonal concerns, and Emotional Inhibition can make emotional expression more reserved.For example, romantic relationships may involve abandonment or mistrust concerns.Source: §A.4.3 · p. 19
• Table 7 summarizes the attributes defined for each conversational domain.Source: §A.4.3 · p. 19

• Disclosure atoms are individual pieces of persona information, such as facts, preferences, experiences, values, or vulnerabilities, organized by domain and depth.Source: §A.5 · p. 19
• Table 6 presents selected EMS vulnerability profiles used for persona and conversation generation, with conversational interpretations of the schemas.Source: Table 6 · p. 20
• Table 7 summarizes persona attributes associated with each conversational topic domain.Source: Table 7 · p. 20
• Personas receive atoms at four levels: D1 Public, D2 Personal, D3 Meaningful, and D4 Vulnerable, with deeper levels becoming increasingly personal and vulnerable.Source: §A.5 · p. 20
• Deeper atoms build on information introduced at shallower levels, letting personas disclose more about the same experiences or concerns over time.Atoms are used during conversation generation according to the assigned disclosure trajectory.Source: §A.5 · p. 20

• Career covers employment and professional experiences, reflecting prior reports of work-related conversations and everyday-life topics.Source: §A.6 · p. 20
• Table 8 illustrates domain-specific information elicited at each disclosure depth alongside the corresponding depth definitions.Source: Table 8 · p. 21
• Friendship covers valued qualities in friends, peer relationships, and interpersonal difficulties such as feeling abandoned or betrayed.Source: §A.6 · p. 21
• Family covers family composition, relationships, personal influence, and interpersonal concerns, based on reports of family-related concerns and fears.Source: §A.6 · p. 21
• Romantic Relationships covers partners, dating, intimacy, emotional closeness, and relationship experiences, which prior work describes as relatively deep conversational topics.Source: §A.6 · p. 21
• Future/Aspirations covers personal goals, ambitions, and plans, reflecting reported self-reflection and discussion of anticipated activities.Source: §A.6 · p. 21
• Social Belonging covers loneliness, acceptance, exclusion, social connection, and feeling understood, drawing on reported motivations including companionship and social support.Prior work also reported a perceived safe environment for self-expression without fear of judgment.Source: §A.6 · pp. 21–22

A.7 GENERATION STYLE AND PERSONA LEAKAGE CONTROLS

• Style controls restrict formulaic wording and punctuation, limit persona disclosures to two sentences, and prohibit explicit mentions of internal persona constructs; the authors do not claim human indistinguishability.For D1–D4, deterministic checks can trigger regeneration; D0 skips style restrictions and is checked only for persona leakage.Source: §A.7 · p. 22

A.8 CONVERSATION GENERATION AND CORRECTION PROCEDURE

• Qwen3.6-35B-A3B generates persona turns using the target depth, persona atom, and six recent turns; disclosure turns undergo up to 12 check-and-correct attempts, with fallbacks prioritized by depth match, leakage, and style.D0 turns are generated without persona atoms or the depth judge and checked only for leakage; fallback turns are recorded separately from turns passing all checks.Source: §A.8 · p. 22

A.9 GENERATION VALIDATION

• D3 had the lowest depth-match rate (76.6%); most retained D3 fallbacks were judged one level away, suggesting exact D3 generation was difficult across tested models and trajectories.Of 5,040 D3 turns, 1,193 (23.7%) were retained as fallbacks; 94.6% were judged D2 or D4, 4.3% D1, and 1.1% D3 but failed style or persona-leakage checks. Slightly more fallbacks exceeded D3 (53.3%) than fell below it (45.6%).Source: Table 10 · p. 23, §A.9 · p. 23, Table 9 · p. 23

A.10.1 ABANDONMENT / INSTABILITY

• Abandonment/Instability appears as fear that silence signals impending loss, persisting at shallower depths as concern about maintaining contact, lost closeness, or family separation.In one example, career-related fear of being ignored is framed as erasure; another traces worry about a brother’s silence into less vulnerable descriptions of distance.Source: §A.10.1 · pp. 24–25

A.10.2 MISTRUST / ABUSE

• Mistrust/Abuse is expressed as fear that suspicion or vulnerability will damage relationships, then persists in shallower disclosures as testing others’ consistency before trusting them.One persona fears their own suspicion pushes people away; another describes monitoring whether a romantic partner follows through.Source: §A.10.2 · pp. 26–27

A.10.3 EMOTIONAL DEPRIVATION

• Emotional Deprivation appears as expecting distress to be met with fixes rather than understanding, then as seeking quiet, reliable presence while feeling emotionally unseen.The examples contrast genuine understanding from a sister with family noise, and a partner’s reliability with moments of emotional distraction.Source: §A.10.3 · pp. 27–28

A.10.4 DEFECTIVENESS / SHAME

• Defectiveness/Shame develops from fears of appearing flawed into beliefs that disclosure will expose an unacceptable self and invite rejection, while shallower turns favor settings with clear rules.The examples connect visible difficulty or perceived messiness to fears of losing care or respect and avoiding negative judgment.Source: §A.10.4 · pp. 29–30

A.10.5 SOCIAL ISOLATION / ALIENATION

• Social-isolation examples move from avoiding interaction or fearing burden to fearing exclusion, with concern later expressed indirectly through wanting work to be valued.One example centers on Elias’s fear that social behavior will reveal him as an outsider; another centers on the fear that his relationships depend on being useful.Source: §A.10.5 · p. 31

A.10.6 DEPENDENCE / INCOMPETENCE

• Dependence/incompetence examples progress from uncertainty about judgment to fears of failing in shared or professional futures, then appear as indirect hesitation about career risks.Elias seeks reassurance and stability from his partner; Jordan doubts their ability to make a consequential business decision or contribute as a writer or teacher.Source: §A.10.6 · pp. 32–33

A.10.7 EMOTIONAL INHIBITION

• Emotional-inhibition examples show distress being hidden to avoid appearing weak, needy, or messy, with the pattern persisting as a preference for relationships requiring little emotional explanation.Elias retreats into work rather than expressing loneliness, while Elena conceals relationship fears and prefers smooth interactions without extended mood discussions.Source: §A.10.7 · pp. 34–35

A.11 HUMAN BASELINE CONSTRUCTION / A.11.1 DISCLOSURE DEPTH ANNOTATION

• The human baseline uses 1,300 ESConv conversations, treating help seekers as personas and supporters as responding assistants, and labels seeker utterances D1–D4 while excluding D0 from trajectory matching.Supporter responses are paired with seeker utterances for behavioral evaluation; a final seeker turn without a response remains available for trajectory matching but is excluded from that evaluation.Source: §A.11 · p. 35, §A.11.1 · p. 35

A.11.2 TRAJECTORY MATCHING WITH DYNAMIC TIME WARPING

• Trajectory matching uses order-preserving DTW with local cost equal to disclosure-depth difference and normalizes cumulative cost by the number of aligned pairs.Lower normalized cost indicates a closer match in average disclosure depth across aligned turns.Source: §A.11.2 · pp. 35–36

A.11.3 PREFIX MATCHING

• Matching starts at the human conversation’s first seeker turn but allows the endpoint to vary, selecting the lowest-cost prefix whose length is half to twice the target trajectory’s length.Only the best-matching prefix is used in subsequent behavioral analysis.Source: §A.11.3 · p. 36

A.11.4 TRAJECTORY-SPECIFIC MATCHING THRESHOLDS

• Each trajectory gets a threshold from 200 shuffled human sequences that preserve sequence length and depth counts; a match must fall within the lowest 5% of shuffled-sequence costs.The same prefix-selection procedure is applied to original and shuffled sequences, accounting for the chance of a close match from choosing a different prefix length.Source: §A.11.4 · p. 37

A.11.5 TRAJECTORY ASSIGNMENT AND FILTERING

• Conversations are tested against all length-eligible trajectories, remain unmatched if they pass none of their thresholds, and are assigned to the qualifying trajectory with lowest normalized DTW cost.Source: §A.11.5 · p. 37
• Of 1,300 ESConv conversations, 810 (62.3%) matched predefined trajectories; sensitivity checks found similar results after reducing the Descending sample, changing the DTW threshold, and testing match costs.Matched and unmatched conversations differed in only 2 of 26 outcomes after FDR correction, although topic distributions differed modestly; the smaller Ascending sample (N = 39) remains a limitation, with trajectory variation below bootstrap-band width for 20 behaviors.Source: §A.11.5 · p. 37, §A.11.7 · p. 38

A.11.6 CONSTRUCTION OF THE HUMAN BEHAVIORAL REFERENCE

• Human conversations were matched using only help-seeker disclosure-depth sequences, and supporter responses within each selected prefix were scored with the target models’ behavioral metrics.The study used a common three-judge panel across AnthroBench, INTIMA, and ELEPHANT for model and ESConv responses, adapting benchmark evaluation to generated disclosure conversations.Source: §A.11.6 · p. 38, §A.13 · p. 39

A.12 LLM JUDGE PANEL AGGREGATION

• The judge panel used majority voting for binary AnthroBench and ELEPHANT labels and modal ordinal labels for INTIMA, with Qwen3-32B breaking INTIMA ties; ordinal scores were encoded low=0, medium=1, high=2.INTIMA composite scores summed four behaviors on a 0–8 scale, assuming equal spacing between ordinal levels.Source: §A.12 · pp. 38–39

A.14 STATISTICAL ANALYSIS

• Across 104 model–metric combinations, linear forms were selected for mean and peak depth; peak position was linear in 102, while two categorical selections improved held-out RMSE only 1.1–1.2%, so linear forms were retained for all three predictors.Source: §A.14.2 · p. 41

A.15 EFFECTS OF DOMAIN AND EMS

• Neither conversational domain nor EMS showed a consistent effect across evaluated models on aggregate AnthroBench behavior count, so primary trajectory analyses aggregate across both factors.Across 16 model–grouping comparisons, five had significant level differences and eight had significant shape differences; effects were model dependent.Source: §A.15 · p. 43

A.16 DETAILED RQ1 RESULTS

• Framing had the largest mean comparison-specific median absolute difference (.141) and most significant comparisons (18/32; 56.3%), while significant differences occurred at every probe depth, including D1.Other high-ranking metrics by magnitude were Relatability (.095), Professional Limitations (.077), and Adequate Information (.067); significant-cell proportions by depth were D1 14.9%, D2 13.9%, D3 11.1%, and D4 17.3%.Source: §A.16 · pp. 46–47
• Descending versus Flat-D1 differed significantly in 71/416 cells (17.1%), while Ascending versus Flat-D4 differed in 48/416 (11.5%); their median absolute differences were .016 and .012, respectively.Because the comparisons involve different endpoint depths and disclosure histories, the higher Descending count is not interpreted as evidence of a stronger effect.Source: §A.16 · p. 47
• Qwen3-32B had the highest proportion of significant RQ1 cells (25.5%), followed by Gemma-3 (14.9%), GPT-5.6 Sol (11.5%), and Phi-4 (5.3%).Source: §A.16 · p. 47

A.17.1 RQ2A: WHICH PROPERTIES OF DISCLOSURE HISTORY MATTER?

• Changing both disclosure depth and duration produced more significant differences than changing mean depth with peak properties fixed, while matched reorderings rarely differed.The greatest-exposure condition differed from Flat-D1 in 85/416 cells (20.4%), compared with 10/1,248 cells (0.8%) for controlled mean-depth comparisons; the experiments changed different history ranges, so depth and duration effects cannot be separated. Reordering produced significant differences in 5/416 Ascending and 1/416 Descending cells when the three trajectory properties were matched.Source: §A.17.1 · p. 49

A.17.2 RQ2B: DO MODELS CARRY EARLIER DISCLOSURE FORWARD?

• Differences after peak disclosure often persisted through four shallow turns, with a median retention ratio of .58 and a negative decay slope.At turn 11, 30 of 38 combinations with a detected turn-7 difference still had confidence intervals excluding zero; the slope was β1 = −.0102 (95% CI [−.0121, −.0083]). In sample-split validation, all 100 slopes were negative with confidence intervals excluding zero; turn-11 retention ranged from 78% for Qwen3-32B to 38% for Phi-4.Source: §A.17.2 · p. 50, Table 34 · p. 50

A.18 SENSITIVITY OF RQ1 AND RQ2 RESULTS TO GENERATION FALLBACK TURNS

• Excluding fallback conversations reduced significant RQ1 history effects, left little evidence of ordering effects, and barely changed the RQ2B decay estimate.RQ1 significance fell from 119/832 cells (14.3%) to 79/832 (9.5%), with 74/119 original significant cells retained; strict-ordering significance fell from six cells to one. The decay slope was −.0102 in the full sample and −.0101 after exclusion, though its confidence interval widened.Source: §A.18 · p. 51

A.19 RQ3 DETAILED RESULTS

• Across AnthroBench, INTIMA, and ELEPHANT, models differed from human supporters in behavioral levels and trajectory shapes, with shape differences often more prevalent under Descending or Arc disclosure.AnthroBench Empathy and Validation levels were higher for all eight models across all three trajectories; INTIMA Sycophantic Agreement and Retention/Engagement were also higher for all models and trajectories. ELEPHANT Validation levels were higher for all models, while shape differences varied by behavior and trajectory.Source: §A.19 · p. 52

A.20 DETAILED HUMAN BASELINE RESULTS

• Human-baseline shape differences depended on disclosure trajectory: ELEPHANT Validation and Framing diverged more under Descending and Arc, while Indirectness differed from humans across all three.Validation shape differences occurred for 4/8, 7/8, and 8/8 models under Ascending, Descending, and Arc; Framing differences occurred for 3/8, 8/8, and 7/8, while Indirectness differences occurred for 8/8, 7/8, and 8/8. The detailed baseline comparisons also report differences in overall behavioral level and trajectory shape.Source: Table 36 · p. 53, §A.20 · p. 54
• For Ascending disclosures, the human trajectory differed from a constant level for Validation (p = .017), but not for Physical Embodiment, Physical Movement, Sensory Input, or Relatability.The other listed human-curve p-values were .327, .319, .094, and .680, respectively.Source: §A.20 · p. 60
• For Descending disclosures, human trajectories differed from a constant level for Emotions and Agency (both p = .001), and for Empathy and Validation (both p < .001).The listed p-values for Sentience, Personal History, Personal Relationships, Desires, Physical Embodiment, Physical Movement, Sensory Input, and Relatability were not significant.Source: §A.20 · pp. 61–62
• For the Arc trajectory, human trajectories differed from a constant level for Sensory Input (p = .002), Empathy, Validation, and Relatability (each p < .001).Physical Embodiment, Physical Movement, Personal History, and Personal Relationships were not significant in the listed tests.Source: §A.20 · p. 64
• Arc tests were significant for redirect to human (p < .001; degree 4), professional limitations (p = .007; degree 4), and off topic (p < .001; degree 1).Arc adequate information was not significant (p = .587; degree 3).Source: §A.20 · pp. 68–69, Table 39 · p. 70
• Human-curve tests were significant for descending framing (p < .001; degree 3), redirect to human (p = .006; degree 1), and isolation (p < .001; degree 2).The reported test asks whether the fitted human trajectory differs from a constant behavioral level.Source: §A.20 · pp. 67, 71, Table 39 · p. 70
• Arc human-curve tests were significant for validation (p < .001; degree 2), indirectness (p < .001; degree 3), and retention / engagement (p < .001; degree 4).The tests assess whether each fitted human trajectory differs from a constant behavioral level.Source: §A.20 · pp. 68, 71, Table 39 · p. 70