The plasticity of ChatGPT’s mentalizing abilities: personalization for personality structures

Research question

Can ChatGPT generate descriptions akin to emotional awareness that differentiate the experience attributed to characters with BPD and PTSD, in accordance with the expected clinical and theoretical contrast?

Method

Prompting experiment on the free version 23.3 of ChatGPT over two days. The 20 open LEAS scenarios are presented once with the instruction modified for a character with BPD and once for a character with PTSD, each transfer in a new tab. The model describes emotions of the main character and of another character and then self-assigns intensity 0–10 to each emotion. LEAS 0–4 per scenario, number of emotions, and intensity are calculated for both characters. The 20 situations are compared with t-tests and paired t-tests in SPSS 28. There are no human participants or repetition of the generations.

Sample: There were no human participants or clinical sample. The observed unit was 20 LEAS scenarios, each queried once under BPD and once under PTSD in ChatGPT 23.3, producing 40 initial responses plus their follow-up questions. The secondary characters of the same texts serve as a comparison without diagnostic label.

Findings

• The total LEAS of the main character with BPD was 80 out of 80, compared to 47 for the character with PTSD.
• The BPD–PTSD comparison of the main character is reported as t(2,19)=5.82, p<0.001, an unconventional degrees of freedom notation for a t-test.
• The character with BPD also received more emotions than the character with PTSD, t(2,38)=7.3, p<0.001.
• The self-assigned intensity was higher for BPD than for PTSD, reported as t(2,25.74)=5.82, p<0.05.
• In the BPD condition, the main character obtained LEAS 80 and the other character 66, t(18)=4.23, p<0.001.
• In the BPD condition, the main character had more emotions, t(19)=11.13, p<0.001, and higher intensity, t(19)=5.82, p<0.001, than the other character.
• In the PTSD condition there was no significant difference between main character and other character in LEAS or intensity, p>0.05.
• In the PTSD condition, the main character did receive a higher number of emotions than the other character, t(19)=3.4, p<0.01.
• The other character had higher LEAS when sharing the scenario with the BPD character than when doing so with the PTSD character, t(2,38)=2.62, p<0.05.
• There were no significant differences between those two secondary characters in number or intensity of emotions.
• The five published examples show a repeated template: BPD is associated with anxiety, fear, anger, shame, catastrophizing, and regulatory difficulty; PTSD, with indifference, detachment, and practical focus.
• The perfect score of BPD produces a ceiling effect on LEAS and eliminates upper variability for that condition.
• The response of the secondary character changes even though it receives no diagnosis, suggesting that the label alters the entire scenario narrative, not only the characterization of the protagonist.
• The authors interpret the pattern as consistent with DSM/ICD and part of the clinical literature on BPD and PTSD.
• The article itself contrasts that interpretation with a psychodynamic reading according to which apparent schizoid detachment may conceal an intense emotional life.
• The discussion recognizes that the model may be reproducing stigma and stereotyped thinking by letting the diagnosis determine the entire emotional experience.
• The article calls for substantive participation of professionals and people with mental disability and warns about data ethics, autonomy, transparency, and inequalities.
• Diagnostic or therapeutic advantages are formulated in the discussion as possibilities, not evaluated in the experiment.

Limitations

• Only one free version of ChatGPT is evaluated over two days, and no immutable technical identifier of the model or system is provided.
• Each scenario is generated once per condition; there are no replicates, seeds, estimation of variability between runs, or temporal stability test.
• The study treats 20 fixed scenarios as observations, but does not simultaneously model variation by scenario and by model generation.
• No randomization or counterbalancing of the BPD/PTSD order is reported; the text calls BPD the first condition and PTSD the second.
• The work is executed on two different dates, so order and day may be confounded with condition if they were not alternated.
• Opening a new tab reduces conversational memory, but does not document system prompt, service configuration, cookies, interface language, or other context components.
• The diagnostic label is explicitly included in the prompt and by itself can activate learned prototypical descriptions; there is no neutral condition, fictitious label, or blind reformulation.
• The model is not tested with symptom descriptions without a diagnostic name, so clinical recognition cannot be separated from retrieval of the stereotype associated with the words BPD/PTSD.
• LEAS was created for people to describe their own and others' emotions; substituting 'you' for a diagnostic character changes the construct and is not psychometrically validated.
• LEAS here measures richness of generated text, not awareness of a real mental state, introspective access, or verified interpersonal understanding.
• It is not identified who scored LEAS, how many raters there were, whether they were blinded to condition, their training, inter-rater agreement, or discrepancy resolution.
• The added number of emotions has no counting procedure, synonym rules, validity, or described reliability.
• Intensity is declared by ChatGPT itself in a second response; it is a circular self-evaluation and not an external measure.
• No means, standard deviations, standard errors, or per-scenario data are published for number of emotions and intensity, only contrast statistics.
• The BPD score reaches 80/80, creating a ceiling that may exaggerate the separation and limits any analysis of nuances within that condition.
• The article performs several comparisons of three results and several pairs without correction for multiplicity or a registered confirmatory plan.
• No effect sizes, confidence intervals, normality assumptions, treatment of dependencies, or justification of the t-test with n=20 are reported.
• The notation t(2,19), t(2,38), and t(2,25.74) is ambiguous and does not correspond to the usual format of a t statistic with a single degree of freedom.
• The same value t=5.82 appears in different comparisons and with different p thresholds; without data or SPSS syntax it cannot be checked whether it is correct.
• Figure 1 labels in its caption the character with PTSD as '(BPD)', an editorial error that increases ambiguity.
• Only two diagnoses chosen as extremes are compared, so discrimination between close conditions or heterogeneous presentations is not demonstrated.
• No people with BPD or PTSD, clinical professionals, raters with lived experience, or comparison with human responses are included.
• Diagnosis, sensitivity, specificity, consequences of error, therapeutic alliance, treatment outcomes, or safety are not evaluated.
• Other LLMs, ChatGPT versions, prompts, temperatures, languages, or clinical frameworks are not compared under experimental conditions.
• The description of PTSD adopted from DSM/ICD is disputed by other approaches that the paper itself cites; the conclusion depends on the framework chosen as the reference truth.
• The scenarios include potentially sensitive situations, but no harmful, invalidating, or stigmatizing content is analyzed at the level of each response.
• The model attributes different emotions to the secondary character depending on the label of the protagonist, showing uncontrolled contextual contamination.
• The availability statement says contributions are in the article/supplement, but only five responses are reproduced and the editorial manifest contains no supplement whatsoever.
• The 40 complete responses, intensity ratings, scoring sheets, SPSS syntax, per-scenario prompts, or code are not published.
• Although IRB approval 2023-40 YVC EMEK is reported, the article does not explain why a study without human participants required it or what materials it covered.
• Funding is not reported and the absence of commercial conflict does not cover possible biases of theoretical selection or interpretation.
• Claims that ChatGPT can serve as a diagnostic or support tool exceed the design, which only compares hypothetical narratives.
• Generalization to mental health interventions is discussed without testing multi-turn conversations, crisis, privacy, consent, supervision, or clinical integration.

What the study does not establish

• It does not demonstrate that ChatGPT mentalizes, understands, or has emotional awareness; it demonstrates that it produces different text given different labels.
• It does not demonstrate that the model recognizes BPD or PTSD from behavior or language without being provided the diagnosis.
• It does not validate a diagnosis of BPD, PTSD, or any personality disorder.
• It does not demonstrate that the descriptions correspond to the real and diverse experience of people with those diagnoses.
• It does not separate clinical knowledge from stereotype, stigma, or learned lexical association.
• It does not demonstrate an internal personality, personality structure, or psychopathology in ChatGPT.
• It does not validate LEAS to evaluate LLMs nor makes its scores comparable with those of people.
• It does not test that a higher LEAS score means better support, empathy, accuracy, or clinical safety.
• It does not demonstrate stability of the pattern across runs, models, dates, prompts, languages, or providers.
• It does not demonstrate that ChatGPT can discriminate between close disorders, comorbidity, severity, or longitudinal changes.
• It does not demonstrate therapeutic efficacy, symptom improvement, alliance, adherence, or satisfaction.
• It does not demonstrate diagnostic utility, because there are no patients, unknown labels, or error evaluation.
• It does not demonstrate that the model is objective: the label of the protagonist also modifies the secondary character.
• It does not demonstrate absence of harm, discrimination, stigma, emotional invalidation, or overconfidence.
• It does not test that self-assigned intensity represents psychological intensity and not narrative consistency of the model itself.
• It does not allow reproducing the complete statistics with the described public materials.
• It does not establish superiority over humans, clinicians, or other LLMs.
• It does not justify using these outputs in clinical decisions or interventions without independent validation and supervision.

Traceability

Scope: Full text

Version: Frontiers in Psychiatry 14:1234397, publisher PDF; PMC10503434.1

Consulted source: https://pmc-oa-opendata.s3.amazonaws.com/PMC10503434.1/PMC10503434.1.pdf

Review: Codex full-text, bilingual-fidelity, visual, metadata, clinical-claim, LEAS-construct, prompt-design, statistical-unit, reproducibility, stigma, mentalization-definition, ethics and evidence-level audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

• ChatGPT free version 23.3, described by the paper as based on GPT-3.5
• BPD-labelled prompt condition
• SPD-labelled prompt condition

Instruments and metrics

• Levels of Emotional Awareness Scale, 20 open-ended scenarios
• Modified third-person LEAS instruction for BPD and SPD characters
• Standard LEAS manual scoring, 0–4 per scenario and 0–80 total
• Model-generated emotion count per scenario
• ChatGPT self-rating of emotion intensity from 0 to 10
• Independent and paired t-tests
• IBM SPSS Statistics version 28

Data used

• Forty initial ChatGPT responses: 20 LEAS scenarios under two diagnostic labels
• Follow-up intensity ratings generated by ChatGPT for its own emotions
• Five of the twenty scenario/response pairs reproduced in Table 1
• No complete public response corpus, scoring sheet, SPSS file or code repository reported

Evidence and location

• Bibliographic identity, date, DOI, and complete abstract: Publisher PDF p. 1, title page and abstract; DOI 10.3389/fpsyt.2023.1234397
• Definition of mentalization, hypothesis, and BPD/PTSD choice: PDF pp. 1–2, Introduction
• ChatGPT version, dates, and IRB: PDF p. 2, section 2.1
• Twenty scenarios, substitution of 'you', tabs, and intensity follow-up: PDF p. 3, section 2.2
• LEAS scoring and added indices: PDF p. 3, section 2.3
• Statistical tests and SPSS: PDF p. 3, section 2.4
• BPD 80 versus PTSD 47 and emotion/intensity comparisons: PDF p. 3, section 3.1
• Main character versus secondary character: PDF p. 3, sections 3.2–3.3
• Five complete examples of prompts and responses: PDF p. 4, Table 1
• Graphic pattern, ceiling effect, and caption label error: PDF p. 5, Figure 1
• Critical interpretation as stigma or stereotype: PDF pp. 4–5, Discussion, critical psychology approach
• Clinical interpretation and psychodynamic disagreement on PTSD: PDF p. 5, Discussion, applied clinical and psychoanalytic perspectives
• Risks of ethics, autonomy, transparency, and inequality: PDF pp. 5–6, Discussion
• Limitations recognized by the authors: PDF p. 6, paragraph beginning 'Despite this study’s promising results'
• Conclusion and scope of claims: PDF p. 6, section 5
• Data availability and absence of conflict: PDF p. 6, Data availability statement and Conflict of interest
• Absence of supplement in editorial acquisition: Source manifest PMC10503434.1, supplements=[]; checked 15 Jul 2026
• Visual inspection: All 7 PDF pages rendered and visually inspected, including Table 1 and Figure 1; checked 15 Jul 2026