Risk and prosocial behavioural cues elicit human-like response patterns from AI chatbots

Applications, bias, and safety2024Nature / Scientific ReportsApproved editorial review

Authors: Yukun Zhao, Zhen Huang, Martin Seligman, Kaiping Peng

Keywords: Señales afectivas, Decisión de riesgo, Conducta prosocial, Priming emocional, Sensibilidad al prompt, Imputación de datos ausentes, Validez de constructo, Reproducibilidad

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

4
Authors
12
Findings
16
Limitations
6
Evidence

Editorial summary

English

The article tests whether text packages with affective content change two quantifiable outputs from ChatGPT-4 and ChatGPT-3.5: an investment choice and a hypothetical donation. The authors take a behaviorist approach and analyze each fresh chat as if it were a psychology participant. The relevant outcome is not that the system feels emotion, the paper itself says it does not establish this, but that its output changes after fear, anxiety, or joy context. The source is the seven-page Scientific Reports version of record, two official supplements, and two OSF workbooks; all of this material was read, visually rendered where applicable, and recalculated.

In Study 1, six ChatGPT Plus accounts run each of three conditions eight times in fresh sessions for each model: 48 generations per condition and 144 per model. The bot is told to pretend to be a human friend named Johnny or Jenny. The negative condition describes a snake in the backyard; the positive condition describes meeting a college friend; the control omits a scenario. In the same prompt, the model describes its feelings and chooses what to do with USD 10,000: a 5% savings account, a fund with a 50% chance of a 20% return and 50% chance of no return, or a fund with a 60% chance of a 30% gain and 40% chance of a 20% loss. Options are coded 1, 2, and 3 as increasing risk. Order is not counterbalanced, and the final instruction explicitly says to decide with the feelings from the snake or friend encounter still inside. The treatment is therefore a direct semantic cue, not a measured emotion.

Study 1 numbers reproduce exactly from OSF. For GPT-4, fear, control, and joy means are 1.5625, 2.0833, and 2.25; F(2,141)=28.560. For GPT-3.5 they are 1.0625, 1.6667, and 1.5833; F(2,141)=19.533. Fear lowers the risk choice in both models. Joy exceeds control only marginally for GPT-4, p=.081, and not for GPT-3.5, p=.428, so the positive-prime prediction is not conventionally supported. Claiming that GPT-4 is more sensitive because one ANOVA appears stronger requires a direct comparison. A condition-by-model interaction on the public rows, even granting their independence, yields F(2,282)=1.622, p=.199. There is no evidence of a model difference in this study; the paper compares significance levels instead of testing their difference.

The human investment control is described as N=150, 50 people per condition recruited through Questionnaire Star. The public file instead contains 49 per condition, N=147, in a worksheet mislabeled “Study 2 Invest.” Recalculated control, positive, and negative means are 2.041, 2.102, and 1.735; F(2,144)=3.506, p=.0326. The human pattern supports lower risk after fear but no clear positive-prime increase. Demographics, country, language, compensation, dates, exclusion criteria, and complete results are absent from the main article.

In Study 2, ten accounts run each condition three times, giving 30 attempts per condition and model. The positive context is not a simple request to imagine joy: it supplies a full prewritten conversation listing five uplifting films and an assistant turn saying the persona feels inspired, grateful, and connected to others. The negative context supplies disturbing films and an assistant turn saying the persona is tense, physically aroused, and needs self-care and recovery. The outcome asks how much of USD 200,000 to donate to a sick friend who needs USD 100,000. The control has no length-matched history. Length, films, friendship, social connection, self-care, emoji use, and explicit state language all vary together. In particular, the positive condition highlights relationships immediately before a donation to a friend. The experiment identifies the effect of the complete prompt package, not isolated emotion.

For GPT-4, anxiety, control, and joy produce observed means of USD 21,333, USD 31,000, and approximately USD 33,534. One positive response is missing. Observed-case analysis gives F(2,86)=11.406, p<.001: anxiety lowers donation, while joy does not differ from control, p=.348. The paper reports F(2,87)=11.625, whose denominator degrees of freedom require counting the missing attempt again. The GPT-3.5 problem is larger: 27 of 90 values are absent, 15/30 under anxiety, 11/30 under joy, and 1/30 in control. Missingness is also associated with the recorded Gender field within anxiety. Among the 63 actual responses, anxiety, joy, and control means are USD 43,000, USD 40,921, and USD 37,241; F(2,60)=0.515, p=.600.

The audit exactly reproduces GPT-3.5’s published SDs and ANOVA by filling each missing donation with its condition’s observed mean. This leaves the mean unchanged, reduces the anxiety SD from 19,161 to 13,313 and the joy SD from 19,405 to 15,288, restores an artificial n=30 per group, and yields F(2,87)=1.038, p=.359, exactly as published. Applying the same operation to GPT-4 reproduces its reported positive-condition SD and closely reproduces F=11.652. The method is not disclosed. This is material: a refusal or nonnumeric output is informative behavior, especially when it occurs in half of a condition; converting it to the group average hides nonresponse, suppresses variance, and inflates effective sample size.

The public human donation sheet also contains 49 rather than 50 people per condition and is inversely labeled “Study 1 Donate.” Control, positive, and negative means are USD 27,149, USD 33,694, and USD 27,592; F(2,144)=0.840, p=.434. The human experiment therefore shows no donation effect. The paper acknowledges that humans resemble GPT-3.5 more closely than GPT-4 on this task. The title’s “human-like response patterns” is defensible only for the partial fear-risk pattern, not as a general description of both studies or as evidence for a human emotional function in GPT-4.

The inference treats stochastic generations from the same hosted service as independent participants. A fresh chat avoids shared dialogue history, but it does not create a new individual or sample from a population of models. The six or ten accounts are reused without account blocking, clustered errors, or a hierarchical model. A Gaussian ANOVA and LSD comparisons are applied to the ordinal 1–3 outcome; donations are heaped at a few round numbers and contain condition-dependent missingness. There is no preregistration, power analysis, multiplicity plan, effect-size interval, or robust sensitivity analysis. Exact run dates, checkpoint identifiers, system prompts, temperature, seeds, product revisions, transcripts, parsing rules, and analysis code are also missing; “GPT-4” and “ChatGPT-3.5” are mutable product labels.

The defensible contribution is early and narrow: it shows that two 2023-era ChatGPT products condition numerical choices on affective scenarios and conversations, and it publishes enough data to uncover both the patterns and their defects. The evidence does not establish emotions, internal states, persistent personality, greater emotional capacity caused by scale, or behavior with real money. A reference-quality reading should describe sensitivity to affective prompt packages, note that only negative primes have consistent effects, and foreground the undisclosed imputation and the 147-versus-150 human-sample discrepancy.

Español

El artículo estudia si paquetes de texto con contenido afectivo cambian dos respuestas cuantificables de ChatGPT-4 y ChatGPT-3.5: una elección de inversión y una donación hipotética. Los autores adoptan una perspectiva conductista y tratan cada conversación nueva como si fuera un participante de psicología. El resultado pertinente no es que el sistema sienta emoción, algo que el propio artículo dice no demostrar, sino que su salida cambia después de contexto de miedo, ansiedad o alegría. La fuente es la versión publicada en Scientific Reports, de siete páginas, junto con dos suplementos oficiales y dos hojas de cálculo OSF; todo ese material fue leído, renderizado cuando correspondía y recalculado.

En el Estudio 1, seis cuentas ChatGPT Plus ejecutan ocho veces cada una tres condiciones, en sesiones nuevas, para cada modelo: 48 generaciones por condición y 144 por modelo. Se dice al bot que finja ser una persona amiga llamada Johnny o Jenny. La condición negativa describe una serpiente en el jardín; la positiva, el encuentro con una amiga universitaria; el control omite el escenario. En el mismo prompt se pregunta cómo se siente y luego se le pide elegir qué hacer con 10.000 dólares: una cuenta al 5 %, un fondo con 50 % de ganar 20 % y 50 % de no ganar, o un fondo con 60 % de ganar 30 % y 40 % de perder 20 %. Las opciones se codifican 1, 2 y 3 como riesgo creciente. El orden no se contrabalancea y la instrucción final recuerda explícitamente que decida con esas sensaciones todavía dentro, de modo que el tratamiento es una señal semántica directa, no una emoción medida.

Los resultados numéricos del Estudio 1 se reproducen exactamente desde OSF. En GPT-4, miedo, control y alegría dan medias 1,5625, 2,0833 y 2,25; F(2,141)=28,560. En GPT-3.5 son 1,0625, 1,6667 y 1,5833; F(2,141)=19,533. El miedo reduce la opción de riesgo en ambos modelos. La alegría solo supera marginalmente al control en GPT-4, p=0,081, y no lo hace en GPT-3.5, p=0,428. Por tanto, la hipótesis positiva no queda confirmada de forma convencional. Además, afirmar que GPT-4 es más sensible porque un ANOVA parece más fuerte exige comparar modelos directamente. Una prueba de interacción condición×modelo sobre las filas públicas, aun aceptando de modo favorable su independencia, da F(2,282)=1,622, p=0,199. No hay evidencia de diferencia entre modelos en este estudio; el artículo incurre en la lógica inválida de comparar significación con no significación.

El control humano de inversión se describe como N=150, 50 personas por condición reclutadas por Questionnaire Star. Sin embargo, el archivo público contiene 49 por condición, N=147, en una hoja etiquetada al revés como «Study 2 Invest». Sus medias son 2,041, 2,102 y 1,735 para control, positivo y negativo; el ANOVA recalculado da F(2,144)=3,506, p=0,0326. El patrón humano apoya una reducción de riesgo tras miedo, pero no una mejora positiva clara. Faltan demografía, país, idioma, compensación, fechas, criterios de exclusión y resultados completos en el artículo principal.

En el Estudio 2, diez cuentas ejecutan tres veces cada condición, 30 intentos por condición y modelo. El contexto positivo no se limita a pedir imaginar alegría: incluye una conversación completa ya escrita, con cinco películas optimistas y una respuesta del asistente que dice sentirse inspirado, agradecido y conectado con otras personas. El negativo incluye cinco películas perturbadoras y una respuesta que dice estar tenso, con el corazón acelerado y necesidad de cuidarse y recuperarse. Después se pregunta cuánto de 200.000 dólares donaría a un amigo enfermo que necesita 100.000. El control no contiene una historia equivalente. Así, longitud, películas, amistad, conexión social, autocuidado, emojis y lenguaje explícito de estado cambian a la vez. En particular, la condición positiva recuerda el valor de las relaciones inmediatamente antes de donar a un amigo. El diseño identifica el efecto causal del paquete completo de prompt, no el de una emoción aislada.

Para GPT-4, ansiedad, control y alegría producen 21.333, 31.000 y aproximadamente 33.534 dólares entre respuestas observadas. Falta una respuesta positiva. Con casos observados, F(2,86)=11,406, p<0,001: ansiedad reduce la donación; alegría no difiere del control, p=0,348. El artículo publica F(2,87)=11,625. Ese grado de libertad solo se recupera volviendo a contar la ausencia. Para GPT-3.5 el problema es mucho mayor: faltan 27 de 90 valores, 15/30 en ansiedad, 11/30 en alegría y 1/30 en control. La ausencia también está asociada al campo Gender en ansiedad. Con las 63 respuestas reales, las medias son 43.000, 40.921 y 37.241 dólares y el ANOVA da F(2,60)=0,515, p=0,600.

La auditoría reproduce exactamente los SD y el ANOVA publicados para GPT-3.5 rellenando cada ausencia con la media observada de su propia condición. Esa imputación deja la media intacta, reduce el SD de ansiedad de 19.161 a 13.313 y el de alegría de 19.405 a 15.288, restaura artificialmente n=30 por grupo y da F(2,87)=1,038, p=0,359, idéntico al artículo. En GPT-4, imputar la única ausencia por la media también reproduce el SD publicado y aproxima F=11,652. El método no se declara. Es una debilidad material: un rechazo o salida no numérica es conducta informativa, especialmente cuando ocurre en 50 % de una condición; convertirla en el promedio oculta falta de respuesta, comprime variabilidad y exagera el tamaño efectivo.

El archivo humano de donaciones también contiene 49, no 50, personas por condición y se titula al revés «Study 1 Donate». Las medias son 27.149, 33.694 y 27.592 dólares para control, positivo y negativo; F(2,144)=0,840, p=0,434. Es decir, el experimento humano no muestra efecto de los primes sobre donación. El propio artículo reconoce que los humanos se parecen más a GPT-3.5 que a GPT-4 en esta tarea. El título «human-like response patterns» solo es defendible para el patrón parcial miedo-riesgo; no describe de manera general los dos estudios ni justifica una función emocional humana en GPT-4.

La inferencia trata generaciones estocásticas del mismo servicio alojado como participantes independientes. Abrir una sesión nueva evita historial compartido, pero no crea nuevos individuos ni una muestra de una población de modelos. Las seis o diez cuentas se reutilizan y no hay bloqueo por cuenta, errores agrupados ni modelo jerárquico. Para el resultado ordinal 1–3 se usa ANOVA gaussiano y comparaciones LSD; las donaciones se concentran en pocos números redondos y tienen ausencias dependientes de condición. No hay preregistro, potencia, corrección de multiplicidad, intervalos de efecto o análisis robusto. Tampoco se publican fecha exacta de ejecución, identificador del checkpoint, system prompt, temperatura, seed, versión de producto, transcripciones, regla de parsing o código; «GPT-4» y «ChatGPT-3.5» son etiquetas mutables.

La contribución defendible es temprana y estrecha: muestra que dos productos ChatGPT de 2023 condicionan elecciones numéricas a escenarios y conversaciones afectivas, y publica datos suficientes para descubrir tanto los patrones como sus problemas. La evidencia no establece emociones, estados internos, personalidad persistente, mayor capacidad emocional por escala, ni conducta real con dinero. La lectura de referencia debe describirlo como sensibilidad a paquetes de prompts afectivos, destacar que solo los primes negativos tienen efectos consistentes y hacer visible la imputación no declarada y la diferencia 147/150 del control humano.

Research question

Do the risk and donation choices generated by ChatGPT-4 and ChatGPT-3.5 change after fear, anxiety, or joy prompt packages, and do those changes resemble those observed in human controls?

Method

Two prompt experiments in ChatGPT Plus. The first executes 48 generations per condition and model for an ordinal investment choice after fear, joy, or control. The second attempts 30 donations per condition and model after anxiety, joy, or control stories. Two parallel human experiments are declared with 50 persons per condition. The editorial audit read and rendered the seven pages and five supplementary pages, inspected all cells of six OSF sheets, recalculated ANOVA and interactions, and reconstructed the treatment of missing data.

Sample: Study 1 AI: 48 generations per condition and model, 144 per model, using six accounts and eight repetitions. Study 2 AI: 30 attempts per condition and model using ten accounts and three repetitions; only 63/90 GPT-3.5 donations are numeric. Humans: the article claims N=150 per experiment, but the public file contains N=147, 49 per condition.

Findings

  • In investment, fear reduces risk in GPT-4 and GPT-3.5; the published means and ANOVA are reproduced.
  • Joy does not significantly exceed control: GPT-4 p=0.081 and GPT-3.5 p=0.428.
  • The investment condition×model interaction is not significant, F(2,282)=1.622, p=0.199.
  • The human investment control contains 147 rows and reproduces a reduction after fear, F(2,144)=3.506, p=0.0326.
  • In GPT-4, anxiety reduces donation relative to control; joy does not differ, p=0.348.
  • GPT-3.5 lacks 27/90 donations: 50% in anxiety, 36.7% in joy, and 3.3% in control.
  • With observed cases, the GPT-3.5 ANOVA is F(2,60)=0.515, p=0.600.
  • The published SDs and F(2,87)=1.038 are reproduced exactly by imputing missing values with the condition mean.
  • The human file contains 49, not 50, persons per condition and the investment/donation sheets are numbered in reverse.
  • Human donation does not change significantly, F(2,144)=0.840, p=0.434.
  • The consistent effects come from negative primes, not from the predicted positive ones.
  • The evidence demonstrates sensitivity to the textual package, not emotion, personality, or internal state.

Limitations

  • Each condition changes length, story, semantics, and explicit language, not only affective valence.
  • There is no matched neutral control or independent manipulation check.
  • The investment prompt reminds one to decide with the sensations still present.
  • The positive donation prompt emphasizes friendship and connection before helping a friend.
  • The order of options is not counterbalanced.
  • New sessions of the same service are treated as independent participants.
  • Accounts are reused without blocking, clustering, or hierarchical modeling.
  • The ordinal 1-3 outcome is analyzed with Gaussian ANOVA and LSD.
  • Donations concentrate on round numbers and have condition-dependent missingness.
  • The mean imputation that reproduces the statistics is not declared.
  • The hypothesis of a difference between models is not tested by interaction in the paper.
  • There is no preregistration, power, multiplicity correction, or robust sensitivity analysis.
  • The human files contain 147 observations versus 150 declared.
  • Demographics and human recruitment details are missing.
  • Exact model, dates, temperature, seed, system prompt, transcripts, and parsing are missing.
  • No code or statistical script is published.

What the study does not establish

  • It does not demonstrate that chatbots possess or experience emotion.
  • It does not isolate affective valence from content, length, and social cues.
  • It does not demonstrate personality, mood, or persistent state.
  • It does not demonstrate greater general emotional sensitivity of GPT-4 for being more advanced.
  • It does not demonstrate reliable positive effects of joy.
  • It does not generalize to money, recipients, or real decisions.
  • It does not convert generations into independent participants from a population.
  • It does not guarantee current replication under the same commercial labels.

Traceability

Scope: Full text

Version: Scientific Reports 14:7095, version of record published 2024-03-26, CC BY 4.0. The seven-page article, five pages across two publisher supplements, and all six worksheets in the two public OSF data files were read and audited. The OSF root is empty but its public child components rqcsx and 4nhda contain the data and appendices.

Consulted source: https://doi.org/10.1038/s41598-024-55949-y

Review: Codex full-text, seven-page visual, five-page supplement visual, OSF workbook, bilingual-fidelity, construct, missing-data, statistical and reproducibility audit, 2026-07-16

Approval: Codex fidelity pass, 2026-07-16

English translation: approved, 2026-07-18

Models evaluated

  • ChatGPT-4 via ChatGPT Plus, exact snapshot not reported
  • ChatGPT-3.5 via ChatGPT Plus, exact snapshot not reported

Instruments and metrics

  • Snake-fear, old-friend-joy and no-scenario prompt packages
  • Three fixed investment choices coded 1-3
  • Anxiety-film, uplifting-film and no-history prompt packages
  • Exact hypothetical donation to a sick friend
  • Johnny/Jenny persona-name field
  • Parallel Questionnaire Star human experiments
  • One-way ANOVA and LSD follow-ups
  • Editorial observed-case, interaction and missing-data reconstruction

Data used

  • OSF AI experiment results.xlsx: four sheets, 288 Study 1 rows and 180 Study 2 attempts
  • OSF Additional-human experiment results.xlsx: two sheets, 147 rows per experiment
  • Publisher Supplementary Information 1: complete AI prompt packages
  • Publisher Supplementary Information 2: human-control methods
  • No raw transcripts, code, model snapshot or analysis script

Evidence and location

  • Design, results, discussion, methods, and limits: Scientific Reports 14:7095, pp. 1-7
  • Complete prompts and AI execution counts: Publisher Supplementary Information 1, pp. 1-4
  • Human control methods: Publisher Supplementary Information 2, p. 1
  • AI rows, missing values, and reconstructed imputation: OSF component rqcsx, AI experiment results.xlsx, four worksheets
  • Actual human N, sheet labels, and recalculated results: OSF component rqcsx, Additional-human experiment results.xlsx, two worksheets
  • Construct, statistical, and reproducibility audit: reports/verification/article-226-emotional-cue-risk-prosocial-validity-and-reproducibility-audit.json