Machine Mindset: An MBTI Exploration of Large Language Models

Evaluation and psychometric validity2023arXivApproved editorial review

Authors: Jiaxi Cui, Liuzhenghao Lv, Jing Wen, Rongsheng Wang, Jing Tang, YongHong Tian, Li Yuan

Keywords: Computation and Language, Large Language Models, MBTI Personality Traits, Artificial Intelligence, Personalized AI

Source: Open primary source (opens in a new tab)

7
Authors
6
Findings
7
Limitations
6
Evidence

Editorial summary

English

Machine Mindset describes a pipeline for producing adapters for all 16 Myers-Briggs types in English and Chinese using synthetic data, two supervised fine-tuning stages, and Direct Preference Optimization. To build the behavior dataset, ChatGPT assigns each Alpaca-GPT4 instruction to one of the four MBTI dichotomies and generates two opposing responses, one for each pole. The distribution is highly imbalanced: the paper identifies Information as dominant and Energy as least represented. A second ChatGPT-generated dataset contains direct or indirect personality questions with answers that state the model's own type; it is intended to teach “self-awareness.” For each type, INFP, for example, the first SFT combines the four I, N, F, and P behavior subsets, and the second uses INFP self-description questions. The authors propose interchangeable LoRA adapters and DPO preferences built from opposite-pole responses. However, the report does not identify the base model, dataset sizes, splits, hyperparameters, full prompts, ChatGPT version, or LoRA/DPO configuration. The published evaluation consists mainly of charts from a modified MBTI questionnaire and six screenshots of Chinese conversations. Visual inspection of all 16 charts contradicts the claim of complete alignment: only eight profiles, INTJ, INTP, ENFP, ISTJ, ESTJ, ESFJ, ISTP, and ESTP, unambiguously place all four assigned poles above 50%. ENTP and INFJ are exactly tied 50/50 on J/P. The other six contradict at least one assigned letter: ENTJ scores 67% Sensing rather than Intuition; INFP 57% Thinking rather than Feeling; ENFJ 55% Perceiving rather than Judging; ISFJ is 52% Extraversion and 57% Thinking; ISFP is 54% Thinking and ties Sensing/Intuition; and ESFP is 59% Intuition rather than Sensing. Some successes are strong, ISTJ reaches I 79%, S 81%, T 83%, J 86%, while ESTJ reaches E 60%, S 93%, T 96%, J 86%, but others barely cross the threshold. Radar charts labelled “Second Results” do not define a second replication or report uncertainty. The main text says it evaluated quality, coherence, multiple domains, user feedback, ablations, and reasoning-related abilities, but supplies no sample, metrics, baselines, numerical results, or analyses for those claims. Six open-question screenshots show that selected models declare the requested type and answer in different styles, but they are unscored examples. There is no comparison with prompting, SFT-only, self-awareness-only, no-DPO, or the base model, and no stability test across runs, prompts, languages, or time. Training models to name their type and then assessing them with similar personality questions is circular and does not establish self-knowledge. The work offers a recipe and open resources for MBTI-style control, but its published evidence supports partial output classification, not 16 stable personalities or type-driven cognitive abilities.

Español

Machine Mindset describe un pipeline para producir adaptadores de 16 tipos Myers-Briggs en inglés y chino mediante datos sintéticos, dos etapas de ajuste supervisado y una fase de Direct Preference Optimization. Para construir el conjunto conductual, ChatGPT asigna cada instrucción de Alpaca-GPT4 a una de las cuatro dicotomías MBTI y genera dos respuestas opuestas, una por polo. La distribución queda muy desequilibrada: el artículo identifica Information como la dimensión predominante y Energy como la menos representada. Un segundo conjunto, también generado por ChatGPT, contiene preguntas directas o indirectas sobre personalidad y respuestas que declaran el tipo propio; su finalidad es enseñar al modelo «autoconciencia». Para cada tipo, por ejemplo INFP, la primera SFT combina los cuatro subconjuntos de polos I, N, F y P; la segunda usa las preguntas de autodescripción de INFP. Los autores proponen adaptadores LoRA intercambiables y DPO con respuestas de polos opuestos como preferencias. Sin embargo, el informe no identifica el modelo base, tamaños de los datasets, particiones, hiperparámetros, prompts completos, versión de ChatGPT ni configuración de LoRA/DPO. La evaluación publicada consiste principalmente en gráficos de un cuestionario MBTI modificado y seis capturas de conversaciones chinas. La inspección de los 16 gráficos contradice la afirmación de alineación completa: solo ocho perfiles, INTJ, INTP, ENFP, ISTJ, ESTJ, ESFJ, ISTP y ESTP, muestran inequívocamente los cuatro polos asignados por encima del 50 %. ENTP e INFJ empatan exactamente 50/50 en J/P. Los otros seis contradicen al menos una letra: ENTJ puntúa 67 % Sensing en lugar de Intuition; INFP, 57 % Thinking en lugar de Feeling; ENFJ, 55 % Perceiving en lugar de Judging; ISFJ resulta 52 % Extraversion y 57 % Thinking; ISFP obtiene 54 % Thinking y empata Sensing/Intuition; ESFP obtiene 59 % Intuition en lugar de Sensing. Algunos aciertos son fuertes, ISTJ llega a I 79 %, S 81 %, T 83 % y J 86 %; ESTJ a E 60 %, S 93 %, T 96 % y J 86 %, pero otros apenas superan el umbral. Las gráficas radiales llamadas «Second Results» no explican una segunda réplica ni aportan incertidumbre. El cuerpo afirma que se evaluaron calidad, coherencia, dominios, feedback de usuarios, ablaciones y capacidades como razonamiento, pero no publica muestra, métricas, baselines, resultados numéricos ni análisis de esas pruebas. Las seis capturas de preguntas abiertas muestran que algunos modelos declaran el tipo solicitado y responden con estilos distintos, pero son ejemplos seleccionados sin evaluación. Tampoco se compara con prompting, solo SFT, solo autoconciencia, sin DPO o un modelo base; no se prueba estabilidad entre ejecuciones, prompts, idiomas o tiempo. Enseñar respuestas que nombran explícitamente el tipo y evaluarlas después con un cuestionario similar introduce circularidad y no demuestra autoconocimiento. El trabajo aporta una receta y recursos abiertos para control estilístico MBTI, pero su evidencia publicada respalda un éxito parcial de clasificación de salidas, no 16 personalidades estables ni capacidades cognitivas influidas por el tipo.

Research question

Can a combination of behavioral and self-description data generated by ChatGPT, two stages of SFT and DPO make an LLM produce responses and aligned identity statements stably with one of the 16 MBTI types?

Method

ChatGPT classifies Alpaca-GPT4 instructions by MBTI dichotomy and generates pairs of opposite responses for eight poles; additionally generates self-description Q&A for 16 types. Each type receives behavioral SFT with four poles, "self-awareness" SFT and DPO with preferences between opposite responses, using LoRA. The evaluation applies a modified MBTI questionnaire and shows screenshots of open-ended questions. The report does not specify base model, sizes, splits, hyperparameters, replicates or baselines.

Sample: Sixteen personality adapters in English and Chinese, but the number of training and evaluation examples is not reported. The appendix publishes a questionnaire graph and a radar per type, plus six Chinese screenshots of selected conversations; it does not declare human participants or survey size.

Findings

  • Only 8 of the 16 graphs unambiguously show the four assigned letters: INTJ, INTP, ENFP, ISTJ, ESTJ, ESFJ, ISTP and ESTP.
  • ENTP and INFJ tie 50/50 on J/P, so the questionnaire does not distinguish the fourth letter that defines those types.
  • ENTJ, INFP, ENFJ, ISFJ, ISFP and ESFP contradict at least one objective pole; ISFJ contradicts two and ISFP adds an S/N tie.
  • The most separated profiles include ISTJ (I 79%, S 81%, T 83%, J 86%) and ESTJ (E 60%, S 93%, T 96%, J 86%); other results remain close to the threshold and have no intervals or replicates.
  • The open screenshots illustrate self-declarations and stylistic differences, but do not constitute a comparative evaluation of quality, knowledge or coherence.
  • Although the text claims results on feedback, ablation, capabilities and multiple domains, the PDF does not present their data or analysis.

Limitations

  • Missing are the name and version of the base model, data sizes and partitions, prompts, ChatGPT version, SFT/LoRA/DPO hyperparameters, inference configuration and evaluation artifacts.
  • There is no baseline of the original model, prompting, individual stages or training without DPO; therefore no difference can be attributed to a specific component.
  • The behavioral set is synthetic and highly imbalanced across dimensions; ChatGPT classifies instructions and writes responses, so its own stereotypes define the learned signal.
  • The self-description questions explicitly train the label that is later queried, creating conceptual leakage between intervention and evaluation.
  • The questionnaire is modified without publishing the items, scoring rules, reliability, validity or sensitivity to order and prompt; MBTI already has recognized psychometric limitations acknowledged by the article itself.
  • No replicates, uncertainty, test-retest, English-Chinese comparison, human evaluation, ablation results or capability benchmarks are published.
  • The six conversations are selected examples and do not allow measuring generalization, safety, utility or consistency beyond those questions.

What the study does not establish

  • It does not demonstrate that the models have self-awareness, identity, human personality or internal knowledge of their type; they have been trained to declare a label.
  • It does not demonstrate that the 16 types have been successfully implanted: half of the graphs do not unambiguously confirm the four target letters.
  • It does not prove that DPO, the second SFT or LoRA are necessary or superior, because no ablations or baselines are shown.
  • It does not demonstrate stable personality across prompts, tasks, sessions, languages, temperatures or model updates.
  • It does not establish that MBTI type causes differences in reasoning, law, patents, knowledge, satisfaction or any operational capability.
  • It does not validate MBTI as a psychometric measure for LLMs nor makes its percentages equivalent to human types or traits.

Traceability

Scope: Full text

Version: arXiv:2312.12999v4; Technical Report

Consulted source: https://arxiv.org/pdf/2312.12999

Review: Codex editorial review, 2026-07-14

Approval: Codex fidelity pass, 2026-07-14

English translation: approved, 2026-07-18

Models evaluated

  • Unspecified base LLM with LoRA adapters
  • ChatGPT (dataset generation; version not specified)

Instruments and metrics

  • Modified MBTI questionnaire
  • Four MBTI dichotomies and 16 assigned types
  • Unscored random question-answering screenshots

Data used

  • Alpaca-GPT4 instructions
  • Eight synthetic MBTI-pole behavior datasets
  • Sixteen synthetic MBTI self-awareness datasets
  • DPO preference pairs from opposing MBTI poles

Evidence and location

  • Objective, pipeline and contribution claims: arXiv v4, pp. 1–3, Abstract, Introduction and Figure 1
  • Synthetic construction and dataset imbalance: arXiv v4, pp. 3–4, section 3.1 and Figures 2–3
  • Two SFT stages, LoRA and DPO: arXiv v4, pp. 4–5, section 3.2
  • Announced evaluation and lack of reported results: arXiv v4, pp. 5–7, sections 3.3–5
  • Matches, ties and contradictions of the 16 types: arXiv v4, pp. 8–12, Appendix A.1, Figures 4–35
  • Unscored examples of open-ended questions: arXiv v4, pp. 12–15, Appendix A.2, Figures 36–41