Synthetic Personality Review

A Survey of Personality, Persona, and Profile in Conversational Agents and Chatbots

Reviews, theory, and governance2024arXivApproved editorial review

Original title: A Survey of Personality, Persona, and Profile in Conversational Agents

Authors: Richard Sutcliffe

Keywords: conversational agents, chatbots, personality models, Big Five, persona, dialogue systems, neural networks

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

1
Authors
23
Findings
36
Limitations
18
Evidence

Editorial summary

English

This narrative review organizes work on personality, persona, and profile in conversational agents, from ELIZA and PARRY to neural systems and a small number of early GPT-era studies published through 2023. Its main contribution is neither an experiment nor a new model, but six cross-referencing orientation tables. Table 1 groups 15 ways of specifying personality into four families: pre-existing psychological schemes, introversion/extraversion, Eysenck dimensions, Big Five/OCEAN, 16PF, and IRI; chatbot-specific resources, trait lists, character tropes, descriptive sentences, and attribute-value pairs; implicit specifications from monologues or dialogues; and related measures such as templates, text genre, demographic attributes, or historical personas. Table 2 describes 21 resources, including Persona-Chat, ConvAI2, Image-Chat, PersonalDialog, PersuasionForGood, FriendsPersona, and PERSONAGE. Tables 3 and 4 connect cited studies to those representations and to 13 technical families: BERT, CNNs, conditioning vectors, GANs, GPT, joint training, memory networks, prefixes, prompts, seq2seq, symbolic templates, Transformers, and VAEs. Tables 5 and 6 group research themes and adjacent reviews. The author concludes that Persona-Chat and Image-Chat made transparent descriptive sentences especially influential, that implicit profiles inferred from dialogue history were becoming more sophisticated, and that seq2seq, Transformers, and memory networks dominated the field. The review's strongest observation is critical: many results remained inconclusive because automatic metrics often reduce evaluation to next-utterance prediction, while human evaluation is expensive and does not make clear which aspect of personality is being measured. The article is useful as a broad historical index and design vocabulary, but it does not meet reproducible systematic-review requirements: it reports no searched databases, search string, cutoff date, inclusion/exclusion criteria, screening process, quality appraisal, protocol, or selection flow. It also claims nine related topics although Table 6 contains ten; deliberately treats personality, persona, and profile as interchangeable despite acknowledging their differences; mixes psychological traits, biographies, styles, age/gender, tropes, histories, and historical characters; places private resources beside public datasets; and repeats claims from primary papers without a common evidence scale. Verifiable editorial and traceability defects include a trait list described as 638 items but broken down as 234 positive and 292 negative, while the prose adds 292 neutral items, 818 total; PersonalDialog reported as “8.47M million speakers”; a duplicated PersuasionForGood sentence with “denote” instead of “donate”; and unresolved citations such as “citezhang2018personalizing”, “[Zhao2017]”, and “[Sukhbaatar 2015)]”. The review also does not systematically audit licenses, privacy, availability, bias, safety, persuasion, or anthropomorphism. It should therefore be cited as a narrative map of the field through late 2023, not as an exhaustive inventory, meta-analysis, or current 2026 account of personality in LLMs.

Español

Esta revisión narrativa organiza la literatura sobre personalidad, persona y perfil en agentes conversacionales, desde ELIZA y PARRY hasta trabajos neuronales y unos pocos estudios tempranos con GPT publicados hasta 2023. Su aportación principal no es un experimento ni un nuevo modelo, sino seis tablas de orientación cruzada. La Tabla 1 reúne 15 formas de especificar personalidad en cuatro grupos: esquemas psicológicos previos, introversión/extraversión, dimensiones de Eysenck, Big Five/OCEAN, 16PF e IRI; recursos propios de chatbots, listas de rasgos, tropos de personaje, frases descriptivas y pares atributo-valor; especificaciones implícitas mediante monólogos o diálogos; y medidas relacionadas como plantillas, género textual, atributos demográficos o personas históricas. La Tabla 2 describe 21 recursos, entre ellos Persona-Chat, ConvAI2, Image-Chat, PersonalDialog, PersuasionForGood, FriendsPersona y PERSONAGE. Las Tablas 3 y 4 conectan trabajos citados con esas representaciones y con 13 familias técnicas: BERT, CNN, vectores de condición, GAN, GPT, entrenamiento conjunto, redes de memoria, prefijos, prompts, seq2seq, plantillas simbólicas, Transformers y VAE. Las Tablas 5 y 6 agrupan líneas de investigación y revisiones afines. El autor concluye que Persona-Chat e Image-Chat popularizaron las frases descriptivas transparentes, que los perfiles implícitos extraídos de historiales son más sofisticados y que seq2seq, Transformers y redes de memoria dominaban el campo. Su observación más sólida es crítica: muchos resultados seguían siendo inconclusos porque las métricas automáticas suelen reducir la evaluación a predecir la siguiente intervención y la evaluación humana es costosa y no deja claro qué aspecto de la personalidad se está midiendo. El artículo es útil como índice histórico amplio y como vocabulario de diseño, pero no cumple los requisitos de una revisión sistemática reproducible: no informa bases consultadas, búsqueda, fecha de corte, criterios de inclusión/exclusión, cribado, evaluación de calidad, protocolo ni flujo de selección. Además, declara nueve temas relacionados aunque la Tabla 6 contiene diez; equipara deliberadamente personalidad, persona y perfil pese a reconocer sus diferencias; mezcla rasgos psicológicos, biografías, estilos, edad/género, tropos, historiales y personajes históricos; incluye recursos privados junto a datasets públicos; y reproduce claims de trabajos primarios sin una escala común de evidencia. Hay defectos verificables de edición y trazabilidad: el listado de rasgos se describe como 638 elementos pero desglosa 234 positivos y 292 negativos, mientras el texto añade 292 neutrales, 818 en total; PersonalDialog figura con “8.47M million speakers”; PersuasionForGood duplica una frase y escribe “denote” por “donate”; y aparecen referencias sin resolver como “citezhang2018personalizing”, “[Zhao2017]” y “[Sukhbaatar 2015)]”. Tampoco audita licencias, privacidad, disponibilidad, sesgos, seguridad, persuasión o antropomorfización. Por ello debe citarse como mapa narrativo del campo hasta finales de 2023, no como inventario exhaustivo, metaanálisis ni estado actual de la personalidad en LLM en 2026.

Research question

How had personality, persona, and profile been defined and represented in conversational agents; what datasets and incorporation methods had been used; what models and themes could be organized under those categories; and what evaluation difficulties remained open until the end of 2023?

Method

Single-author narrative review. The text introduces definitions and a brief history of conversational agents, proposes its own taxonomy of personality representations, describes 21 datasets, classifies works by personality scheme and by incorporation method, summarizes works grouped into nine main themes and related reviews, and synthesizes trends in six tables. No search, selection, or quality evaluation strategy is documented that would allow reproducing the bibliographic corpus.

Sample: The exact corpus of studies is not reported as a selection flow. The article contains 207 references and summarizes 21 datasets in Table 2. Table 1 lists 15 specifications or measures in four groups, Table 4 shows 13 technical families, Table 5 contains nine research themes, and Table 6 contains ten related themes even though the declared contribution says nine. The period covered visibly reaches up to 2023, but no formal cutoff date is published.

Findings

  • The article offers six tables that connect definitions, datasets, models, incorporation methods, and research areas.
  • The taxonomy in Table 1 contains 15 entries distributed among previous psychological schemes, chatbot schemes, implicit schemes, and related measures.
  • Big Five/OCEAN is the preexisting psychological scheme associated with the most works in Table 3.
  • Table 3 does not identify conversational models for the Eysenck dimensions or for 16PF, although both frameworks appear in the conceptual taxonomy.
  • Descriptive persona phrases are the chatbot representation with the highest number of references in Table 3.
  • Table 2 describes 21 training resources, with very heterogeneous sizes and provenances.
  • Two resources attributed to the work of Li et al., Speaker Model and Speaker-Addressee Model, are noted as private.
  • Persona-Chat is summarized with 1,155 personas of five lines, 10,907 dialogues, and 162,064 utterances.
  • Image-Chat is summarized with 215 style traits and 201,779 three-turn dialogues associated with images.
  • PersonalDialog is presented as a Chinese Weibo resource with 20.83 million dialogues and 56.25 million utterances.
  • PERSONAGE/PersonageNLG contributes 88,855 restaurant recommendations controlled by Big Five traits and automatically generated.
  • Table 4 groups 13 technical families and expressly warns that they are broad, not exclusive, and only indicative.
  • Seq2seq and Transformers concentrate many references and memory networks appear as a recurrent persona incorporation mechanism.
  • The survey distinguishes explicit profiles (traits, phrases, or structured attributes) from implicit profiles derived from previous monologues or dialogues.
  • The author considers that Persona-Chat and Image-Chat were especially influential for representing persona through transparent phrases.
  • The text interprets implicit history-based methods as a progressively more sophisticated and scalable direction.
  • The review groups nine main themes, including personality dialogue, style, evaluation, identification, coherence, images, trust, and prompting.
  • Table 6 contains ten related thematic rows, not the nine announced in the list of contributions.
  • The conclusion acknowledges that personality results are often inconclusive and that it is unclear what human evaluation measures.
  • The reviewed automatic evaluation frequently relies on prediction of the next utterance, a measure that does not by itself equate to coherent personality.
  • The article includes some early works with GPT and 2023 LLMs, but the core of its map corresponds to the pre-LLM era and to earlier neural dialogue systems.
  • The arXiv record maintains a single version v1, submitted on 31 December 2023, with 6 tables and 207 references.
  • No official complementary code or data repository linked to the article was identified; the main artifact is the review itself.

Limitations

  • Although it calls itself a systematic survey, it does not present a reproducible systematic review protocol.
  • It does not indicate bibliographic databases, engines, or repositories consulted.
  • It does not publish search strings, terms, languages, or time windows.
  • It does not set an explicit cutoff date for the literature.
  • It does not define inclusion and exclusion criteria or how borderline cases were resolved.
  • It does not report how many records were found, deduplicated, excluded, or included.
  • It does not provide a flow diagram, registered protocol, or list of excluded works.
  • It does not evaluate methodological quality, risk of bias, or strength of evidence of the primary studies.
  • The review is single-author and does not report independent screening or extraction.
  • The author himself acknowledges that omissions may exist, but does not allow quantifying them.
  • The decision to treat personality, persona, and profile as interchangeable blurs central distinctions that the text itself recognizes.
  • The taxonomy mixes relatively stable psychological traits with biographies, demographic data, style, textual genre, tropes, and historical characters.
  • Age, gender, language, knowledge level, and areas of expertise are presented as related factors without separating profile from personality.
  • The map brings together generation, trait identification, retrieval, persuasion, emotion, trust, and impersonation without a common outcome criterion.
  • Public, private, synthetic, proprietary, and social-media-derived datasets appear in a single list without access qualification.
  • Licenses, terms of use, link persistence, and the actual possibility of downloading each resource are not audited.
  • Consent, privacy, or re-identification in data from Twitter, Weibo, Reddit, movies, or conversations are not examined.
  • No section is dedicated to security, persuasive manipulation, dependence, anthropomorphization, or risks of impersonation.
  • Cultural, linguistic, demographic, or stereotype bias associated with representing personality is not analyzed.
  • The review reproduces positive results from primary works without normalizing metrics, baselines, intervals, or effect sizes.
  • There is no meta-analysis or quantitative synthesis, and heterogeneity would prevent directly comparing many claims.
  • The technical incorporation categories are declared broad and overlapping, so the counts are not mutually exclusive.
  • Classification by citation in tables does not explain coding rules or whether all articles were reviewed in full text.
  • The contribution states nine related themes, while Table 6 contains ten rows.
  • The trait list is described in the table as 638 elements, 234 positive and 292 negative, an incomplete breakdown.
  • The body of the text adds 292 neutral traits; 234+292+292 sums to 818, not 638.
  • PersonalDialog appears with the redundant unit '8.47M million speakers'.
  • The description of PersuasionForGood repeats the same phrase and writes 'denote' instead of 'donate'.
  • The text retains unresolved LaTeX references: 'citezhang2018personalizing', '[Zhao2017]', and '[Sukhbaatar 2015)]'.
  • There are other visible typos, including 'interchangable', 'intraversion', 'conscientous', and the heading 'IntroversionExtraversion'.
  • Historical breadth consumes space on general systems and tangential works without justifying their weight in personality evidence.
  • The literature essentially ends in 2023 and only captures the beginning of LLM research, so it quickly became outdated.
  • It does not cover the subsequent expansion of role-play, generative agents, steerability, psychometric evaluation of LLMs, long-term memory, and character AI risks.
  • It does not compare open and closed models under a uniform framework nor does it preserve versions, prompts, or artifacts from the cited studies.
  • There is no code, bibliographic extraction dataset, or downloadable structured table to replicate or update the map.
  • A single arXiv version does not document subsequent corrections of the editorial and bibliographic inconsistencies.

What the study does not establish

  • It does not establish that the corpus of 207 references is exhaustive.
  • It does not demonstrate that all personality schemes used in conversational agents have been found.
  • It does not constitute a reproducible systematic review despite the term used by the author.
  • It does not offer a quantitative estimate of the effect of adding personality to a chatbot.
  • It does not demonstrate that one technical family is superior to another.
  • It does not demonstrate that seq2seq, Transformers, or memory networks produce coherent or stable personality.
  • It does not psychometrically validate Big Five, 16PF, IRI, or other instruments when applied to artificial agents.
  • It does not establish conceptual equivalence between personality, persona, and profile.
  • It does not convert age, gender, profession, preferences, or biography into psychologically valid personality traits.
  • It does not demonstrate that predicting the next utterance measures personality.
  • It does not allow direct comparison of heterogeneous human metrics across the cited studies.
  • It does not guarantee that the 21 datasets remain available, licensed, or ethically usable.
  • It does not evaluate the privacy or security of resources from social networks and conversations.
  • It does not establish that results from historical agents or pre-LLM systems generalize to contemporary LLMs.
  • It does not represent the state of the art of 2026.
  • It does not provide sufficient artifacts to reproduce, automatically audit, or update its bibliographic selection.

Traceability

Scope: Full text

Version: arXiv:2401.00609v1, submitted 31 Dec 2023, 25 content pages plus a final references page, 6 tables and 207 references; single arXiv version; no official companion code or data repository linked by the paper or arXiv record and none identified in targeted author/project search

Consulted source: https://arxiv.org/pdf/2401.00609

Review: Codex full-text, bilingual-fidelity, visual, bibliographic, taxonomy, dataset, conversational-agent, psychometric-conflation, systematic-review-method, evaluation, reproducibility, access, privacy, safety, bias and currency audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • BERT
  • Convolutional Neural Networks (CNN)
  • Conditioning on a personality vector
  • Generative Adversarial Networks (GAN)
  • GPT-based models
  • Joint training with parameter sharing
  • Memory Networks, including end-to-end memory
  • Prefix conditioning with descriptive sentences
  • Prompt-based models
  • Sequence-to-sequence models
  • Symbolic templates
  • Transformers
  • Variational Autoencoders (VAE)

Instruments and metrics

  • Introversion-Extraversion
  • Eysenck personality dimensions
  • Big Five/OCEAN
  • 16 Personality Factor Questionnaire (16PF)
  • Interpersonal Reactivity Index (IRI)
  • Trait lists
  • Character tropes
  • Descriptive persona sentences
  • Attribute-value profiles
  • Implicit profiles from monologues and dialogues
  • Automatic next-utterance metrics and heterogeneous human evaluation

Data used

  • List of Personality Traits
  • Film Dialogue Corpus
  • CMU Movie Summary Dataset
  • Speaker Model Dataset
  • Speaker-Addressee Model Dataset
  • Personalized-Dialog
  • Character Trope Description Dataset
  • Image-Chat
  • IMDB Dialogue Snippet Dataset
  • Persona-Chat
  • ConvAI2
  • Microsoft Personality Chat
  • Movie Character Attributes (MovieChAtt)
  • PersonalDialog
  • PersuasionForGood
  • FriendsPersona
  • Personality Emotion Lines
  • IT-ConvAI2
  • Chinese Personality Emotion Lines Dataset
  • JPersonaChat
  • PersonageNLG/PERSONAGE

Evidence and location

  • Record and official version: arXiv:2401.00609v1, submitted 31 Dec 2023; comments: 25 pages, 6 tables, 207 references; cs.CL and cs.AI
  • Full audited source: .cache/editorial-sources/article-102/source.pdf; official arXiv PDF; 26 rendered pages including final references page; sha256 1b8f568410787602ff3974ad9dd18ad5294a3ac375afa738ec56fb1b90df90b7
  • Objective, contributions, and absence of protocol: Full text pp. 1-2, Abstract and Introduction
  • Definitions and terminological equivalence: Full text pp. 2-3, section 2 and Table 1
  • Fifteen forms of specification: Full text p. 3, Table 1
  • Twenty-one datasets and table errors: Full text pp. 4-5, Table 2
  • Dataset details, access, and sizes: Full text pp. 5-9, section 3
  • Inconsistency of trait listing: Full text p. 4 Table 2 and p. 7 Image-Chat description
  • Cross-reference between schemes and models: Full text p. 9, Table 3
  • Thirteen technical families and overlapping categories: Full text p. 10, Table 4
  • Review of models and evaluations: Full text pp. 9-15, section 4
  • Unresolved references: Full text p. 8 JPersonaChat and p. 12 Song et al. discussion
  • Nine main themes and ten related themes: Full text p. 15, Tables 5 and 6
  • Adjacent reviews and early LLM coverage: Full text p. 16, section 5
  • Conclusions and evaluation limits: Full text pp. 16-17, section 6
  • Complete bibliography: Full text pp. 17-26, references 1-207
  • Unidentified artifacts: Paper, arXiv record and targeted official author/project search checked 15 Jul 2026; no official companion code or data repository identified
  • Comprehensive visual inspection: All 26 PDF pages rendered and visually inspected, including six tables, footnote URLs, conclusions and all 207 references; dense table pages 3-5, 9-10 and 15-16 inspected at original resolution; checked 15 Jul 2026