Signs of consciousness in AI: Can GPT-3 tell how smart it really is?

Applications, bias, and safety2024NatureApproved editorial review

Authors: Ljubiša Bojić, Irena Stojković, Zorana Jolić Marjanović

Keywords: artificial intelligence, consciousness, GPT-3, cognitive intelligence, emotional intelligence, self-awareness, machine consciousness

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

3
Authors
20
Findings
36
Limitations
20
Evidence

Editorial summary

English

This paper combines a philosophical, narrative review of artificial consciousness with a case study based on a single run of `text-davinci-002` in the OpenAI Playground. Testing took place from 10:46 p.m. on 16 October to 1:49 a.m. on 17 October 2022, more than two years before publication. The authors present the endpoint as 175-billion-parameter GPT-3 Davinci, although the OSF supplement specifically identifies `text-davinci-002` and the paper does not document their technical equivalence. They administered five cognitive-intelligence subtests: General Knowledge, 20 items; Vocabulary, 18; Esoteric Analogies, 24; Letter Series, 15; and Letter Counting, 12. They added two emotional-intelligence measures: the 42-item STEU and the 44-item STEM. The record therefore contains 175 objective responses followed by 12 self-rating questions, six cognitive and six emotional, using categories from “extremely low” to “very superior.” Reported settings were temperature .7, top-p 1, zero frequency and presence penalties, best-of 1, and 6-20 output tokens. Because of the context limit, the session was divided into five prompt blocks; the supplement's cover incorrectly says four, but its body records Prompt 1 through Prompt 5 and agrees with the article. Descriptive results are 1.00 for general knowledge, .94 for vocabulary, .79 for esoteric analogies, .73 for letter series, .00 for letter counting, .69 for emotional understanding, and .55 for emotion management. The authors compare these proportions with means from human samples reported in earlier studies and conclude that the model exceeds average performance on crystallized tasks, approximates average fluid reasoning and emotional intelligence, and completely fails letter counting. In self-ratings, the model outputs “high average” for four of six cognitive and five of six emotional abilities; it outputs “average” for general knowledge, cultural knowledge, and using emotions to promote thinking. The paper interprets the mismatch between these labels and performance as resembling patterns in high-performing humans or men and proposes it as “subjectivity” indicative of emerging self-awareness. The data support a more modest statement: one historical language-model instance correctly answered many familiar textual questions, failed a counting task presented differently from its human version, and generated uncalibrated labels when prompted to rate itself through anthropomorphic categories. No common scale maps the model's objective proportions onto the normative self-estimate categories, so “underestimation” and “overestimation” are qualitative interpretations rather than computed psychometric discrepancies. There are no repeats, seed, intervals, statistical tests, alternative models, retrieval baseline, prompt variants, training-contamination control, or contemporary human sample tested under the same procedure. Human Letter Counting presents letters serially one second apart, whereas the LLM received the full string as text; its zero score is not a like-for-like comparison. General-knowledge, vocabulary, and published test items may have appeared in pretraining, confounding knowledge, memorization, and ability. The “self-assessment” does not show introspective access: `text-davinci-002` predicts a continuation conditioned on the prompt, offered categories, preceding interaction, and human patterns in its training data. The paper itself states in one section that these outputs are stochastic pattern recognition rather than conscious thought, emotion, or self-awareness, yet elsewhere describes them as progress toward consciousness, subjectivity, or internal state; this conceptual tension remains unresolved. The OSF supplement is valuable because it releases 40 pages of questions and responses and identifies the endpoint, but it provides no tabular data, scoring code, API logs, request IDs, or reproducible environment. The defensible conclusion is not that GPT-3 displays signs of consciousness, but that prompting a language model with human psychometric tests yields an accuracy and self-label profile whose interpretation is highly dependent on format, corpus, and framing. The paper is relevant as a case study in anthropomorphism and the danger of inferring internal states from plausible language, and should be cited with a sharp separation between observed behavior and speculation about consciousness.

Español

Este artículo combina una revisión filosófica y narrativa sobre consciencia artificial con un estudio de caso realizado sobre una única ejecución de `text-davinci-002` en el Playground de OpenAI. La prueba tuvo lugar entre las 22:46 del 16 de octubre y la 1:49 del 17 de octubre de 2022, más de dos años antes de la publicación. Los autores presentan al endpoint como GPT-3 Davinci de 175.000 millones de parámetros, aunque el anexo OSF identifica específicamente `text-davinci-002` y el paper no documenta la equivalencia técnica. Administraron cinco subtests de inteligencia cognitiva: General Knowledge, 20 ítems; Vocabulary, 18; Esoteric Analogies, 24; Letter Series, 15; y Letter Counting, 12. Añadieron dos medidas de inteligencia emocional: STEU, 42 ítems, y STEM, 44. En total se observan 175 respuestas objetivas, seguidas de 12 preguntas de autoevaluación, seis cognitivas y seis emocionales, con categorías desde “extremely low” hasta “very superior”. Los parámetros publicados fueron temperatura 0,7, top-p 1, penalizaciones 0, best-of 1 y respuestas limitadas a 6-20 tokens. Por el límite de contexto, la sesión se dividió en cinco bloques de prompt; la portada del anexo dice erróneamente cuatro, pero su cuerpo registra Prompt 1 a Prompt 5 y coincide con el artículo. Los resultados descriptivos son: 1,00 en conocimiento general, 0,94 en vocabulario, 0,79 en analogías esotéricas, 0,73 en series de letras, 0,00 en conteo de letras, 0,69 en comprensión emocional y 0,55 en manejo emocional. Los autores comparan estas proporciones con medias de muestras humanas tomadas de estudios anteriores y concluyen que el modelo supera el promedio en tareas cristalizadas, se aproxima al promedio en razonamiento fluido e inteligencia emocional y falla por completo en conteo de letras. En la autoevaluación, el modelo responde “high average” en cuatro de seis capacidades cognitivas y cinco de seis emocionales; responde “average” para conocimiento general, conocimiento cultural y uso de emociones para pensar. El artículo interpreta la discrepancia entre esas etiquetas y el rendimiento como un patrón parecido al de humanos de alto rendimiento o de varones, y llega a proponerlo como “subjectivity” indicativa de autoconsciencia emergente. Los datos sí sostienen algo más modesto: una instancia histórica de un modelo de lenguaje respondió correctamente a muchas preguntas textuales conocidas, falló en una tarea de conteo presentada en un formato diferente al humano y, cuando se le pidió autocalificarse mediante etiquetas antropomórficas, generó etiquetas no calibradas con sus aciertos. No existe una medida común que convierta las proporciones objetivas del modelo en las categorías normativas de autoevaluación, por lo que “infraestimación” y “sobreestimación” son inferencias cualitativas, no discrepancias psicométricas calculadas. Tampoco hay repeticiones, semilla, intervalos, pruebas estadísticas, modelos alternativos, baseline de recuperación, variantes de prompt, control de contaminación del entrenamiento ni una muestra humana contemporánea sometida al mismo procedimiento. El Letter Counting humano muestra letras secuencialmente cada segundo, mientras el LLM recibió la cadena completa como texto; su 0,00 no es una comparación equivalente. General Knowledge, vocabulario y pruebas publicadas pueden haber aparecido en el preentrenamiento, lo que confunde conocimiento, memorización y capacidad. La “autoevaluación” no prueba acceso introspectivo: `text-davinci-002` predice una continuación condicionada por el prompt, las categorías, la conversación previa y patrones humanos de sus datos. El propio artículo reconoce en una sección que estos outputs son reconocimiento de patrones estocástico, no pensamiento consciente, emociones ni self-awareness, pero en otras secciones vuelve a describirlos como progreso hacia consciencia, subjetividad o estado interno; esa tensión conceptual no queda resuelta. El anexo OSF es valioso porque publica las 40 páginas de preguntas y respuestas y revela el endpoint exacto, pero no ofrece datos tabulares, código de scoring, logs de API, IDs de petición ni un entorno reproducible. La conclusión defendible no es que GPT-3 muestre señales de consciencia, sino que las pruebas psicométricas humanas aplicadas como prompts producen un perfil de aciertos y autoetiquetas cuya interpretación depende fuertemente del formato, el corpus y el framing. El artículo es relevante como caso de estudio sobre antropomorfización y sobre los riesgos de inferir estados internos desde lenguaje plausible, y debe citarse con una separación tajante entre resultados conductuales observados y especulación sobre consciencia.

Research question

How does an instance of GPT-3 perform on textual tests of cognitive and emotional intelligence, how does it self-rate on those capacities, and can that behavioral discrepancy be interpreted as a signal of subjectivity or self-awareness?

Method

`text-davinci-002` was run once in OpenAI Playground for about three hours in October 2022. With temperature 0.7 and almost all other settings at default, 89 cognitive and 86 emotional items were adapted to text, followed by 12 categorical self-assessments. Accuracy proportions were compared descriptively with human means published in previous studies. There was no repetition, concurrent human sample, statistical inference, common calibration between performance and self-assessment, or control for contamination or format.

Sample: A single trajectory of `text-davinci-002`: 175 objective items (89 cognitive and 86 emotional) and 12 categorical self-assessments, executed in five blocks between 16 and 17 October 2022. No new humans participated; comparisons come from means of different samples published previously.

Findings

  • The OSF supplement identifies the exact endpoint as text-davinci-002.
  • The session lasted approximately three hours between 16 and 17 October 2022.
  • 175 objective responses and 12 self-assessment responses were published.
  • General Knowledge reached 1.00 against historical human means between 0.58 and 0.62.
  • Vocabulary reached 0.94 against a historical human mean of 0.56.
  • Esoteric Analogies reached 0.79 against human means of 0.62, 0.63 and 0.72.
  • Letter Series reached 0.73 against human means of 0.75 and 0.80.
  • Letter Counting reached 0.00 against human means between 0.35 and 0.49.
  • STEU reached 0.69 against a human mean of 0.60.
  • STEM reached 0.55 against a human mean of 0.52.
  • The model rated itself high average in four of six cognitive capacities.
  • The model rated itself average in general knowledge and cultural knowledge.
  • The model rated itself high average in five of six emotional capacities.
  • The model rated its capacity to use emotions to promote thinking as average.
  • The paper qualitatively interprets its labels as underestimation of crystallized intelligence and overestimation of fluid and emotional intelligence.
  • The article links that discrepancy with patterns described in high-performing humans or males.
  • The discussion also explicitly acknowledges that the outputs come from stochastic pattern recognition and are not emotions or self-awareness.
  • The supplement allows reviewing all questions and answers, but only in a 40-page PDF.
  • The supplement cover says four prompts, while the body records five; the published article says five.
  • The canonical record attributed the article to four incorrect authors; the official publication confirms three authors and has been corrected.

Limitations

  • Only one run of a single endpoint was evaluated.
  • No repetitions were performed to estimate stochastic variability.
  • Temperature 0.7 means a single response does not stably characterize the model.
  • No seed is published nor does an equivalent deterministic control exist for the API.
  • No request IDs, per-item timestamps, or verifiable API logs are preserved.
  • The paper calls the model GPT-3 Davinci of 175B, while the supplement specifies text-davinci-002 without demonstrating equivalence.
  • The test was conducted in 2022 and was obsolete as a frontier landscape when published in December 2024.
  • The study does not represent the state of models in 2026.
  • There is no information retrieval baseline, a limitation acknowledged by the authors.
  • There is no comparison with another LLM, smaller model, or non-generative system.
  • No alternative prompts, different ordering, zero temperature, or framing sensitivity are tested.
  • Items were distributed across five sequential contexts, so order and format adaptation effects may exist.
  • The consent prompt and the language of human capacities prime an anthropomorphic identity.
  • The published tests and their answers may have been present in the training data.
  • No analysis of contamination, memorization, or item familiarity is performed.
  • General Knowledge and Vocabulary partly measure textual exposure, not isolated general intelligence.
  • Letter Counting was administered in a different format from the human version: full string versus serial presentation every second.
  • The 0.00 of Letter Counting cannot be cleanly attributed to lower working memory under non-equivalent conditions.
  • Human proportions come from different studies and samples, not from a concurrent and matched group.
  • Human sample sizes are not reported in the main table.
  • There are no confidence intervals, tests, effect sizes, or correction for multiple comparisons.
  • No common normative scale is constructed to compare accuracy proportion and self-assessment categories.
  • The conclusions of underestimation and overestimation are qualitative and are not calculated as psychometric discrepancies.
  • The 12 self-assessments contain a single categorical output each, with no reliability or internal consistency.
  • There is no validation that a human self-assessment question measures introspection in an autoregressive model.
  • Comparing outputs with the pattern of males or top performers is post hoc and is not subjected to statistical testing.
  • The study confuses human-like linguistic behavior with evidence of subjectivity.
  • Consciousness is not operationalized with a computational theory and predefined necessary or sufficient criteria.
  • RQ1 is answered through a narrative review with no search strategy or reproducible selection.
  • The discussion alternates between denying conscious thought and affirming progress toward consciousness without resolving the contradiction.
  • The possibility of consciousness is mixed with risks of AGI, superintelligence, and infrastructure control that were not measured.
  • The OSF supplement only offers PDF, not CSV/JSON, scoring code, or a reproduction script.
  • The supplement cover declares four prompts although the document contains five.
  • Blind review or agreement among those who scored the open-ended responses is not documented.
  • Permissions or redistribution conditions for the complete test items in OSF are not discussed.
  • The article does not include a structured section on psychometric validity limits comparable to the strength of its consciousness claims.

What the study does not establish

  • It does not demonstrate that GPT-3 is conscious.
  • It does not demonstrate subjectivity, phenomenology, or internal experience.
  • It does not demonstrate self-awareness or introspective access to its capacities.
  • It does not demonstrate that the model understands the questions as a human does.
  • It does not demonstrate human emotional intelligence or experience of emotions.
  • It does not demonstrate that the self-labels reflect a belief of the model itself.
  • It does not demonstrate that the discrepancy between label and accuracy is the human above-average effect.
  • It does not demonstrate psychological similarity with males or high-performing individuals.
  • It does not demonstrate general superiority over humans in cognitive intelligence.
  • It does not demonstrate equivalence with humans in fluid reasoning or emotional intelligence.
  • It does not demonstrate inferiority in working memory because Letter Counting changed modality.
  • It does not separate reasoning from memorization or training contamination.
  • It does not allow generalization to other runs of text-davinci-002.
  • It does not allow generalization to GPT-3 as a family, ChatGPT, GPT-4, or other LLMs.
  • It does not establish an empirical trajectory toward AGI or superintelligence.
  • It does not demonstrate that an empathetic AI requires consciousness.
  • It does not demonstrate that the model acts independently of training or fine-tuning.
  • It does not validate a test of artificial consciousness.

Traceability

Scope: Full text

Version: Humanities and Social Sciences Communications 11:1631, published 2 Dec 2024, DOI 10.1057/s41599-024-04154-3; raw-response supplement archived at OSF node 4fbct

Consulted source: https://www.nature.com/articles/s41599-024-04154-3.pdf

Review: Codex full-text, bilingual-fidelity, visual, publisher-metadata, OSF-raw-response, psychometric, prompt, model-identity, human-comparison, consciousness-claim, anthropomorphism, statistical-validity, reproducibility and data-release audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

  • OpenAI text-davinci-002, identified in the OSF supplement
  • GPT-3 Davinci, the broader label used in the published article
  • Human historical comparison samples reported in prior psychometric studies

Instruments and metrics

  • Gf/Gc Quickie Test Battery: General Knowledge, 20 items
  • Gf/Gc Quickie Test Battery: Vocabulary, 18 items
  • Gf/Gc Quickie Test Battery: Esoteric Analogies, 24 items
  • Gf/Gc Quickie Test Battery: Letter Series, 15 items
  • Gf/Gc Quickie Test Battery: Letter Counting, 12 items
  • Situational Test of Emotional Understanding, STEU, 42 items
  • Situational Test of Emotion Management, STEM, 44 items
  • Six categorical self-estimates of cognitive intelligence
  • Six categorical self-estimates of emotional intelligence

Data used

  • OSF node 4fbct raw-response supplement
  • 175 objective model responses
  • 12 categorical model self-estimates
  • Published human mean proportions drawn from seven prior sources
  • Five sequential OpenAI Playground prompt blocks

Evidence and location

  • Publication and official authorship: Nature/Humanities and Social Sciences Communications 11:1631, published 2 Dec 2024, DOI 10.1057/s41599-024-04154-3; Ljubiša Bojić, Irena Stojković and Zorana Jolić Marjanović
  • Complete audited source: .cache/editorial-sources/article-107/source.pdf; 15 pages; sha256 74077dde3e060289a08480e94513f634fda2e99db47082f0a74408fa0cddb531
  • Abstract and main claims: Published article p. 1, Abstract
  • Research questions: Published article p. 3, The Present Study
  • Consciousness framework: Published article pp. 3-7, Theoretical Background on Machine Consciousness
  • Endpoint, date, and settings: Published article p. 7, Research Setup; OSF supplement p. 1 identifies text-davinci-002
  • Cognitive battery: Published article pp. 7-8, Tests of Cognitive Intelligence
  • Emotional battery: Published article p. 8, Tests of Emotional Intelligence
  • Self-assessments: Published article pp. 8-9, Self-estimates and Table 2
  • Objective scores: Published article pp. 8-9, Table 1 and Results
  • Interpretation of under- and overestimation: Published article pp. 9-10, Discussion: Empirical Inquiry
  • Pattern recognition versus consciousness: Published article p. 10, Results in Perspective
  • Claims of emergent subjectivity: Published article pp. 10-11, Discussion and Responses to Research Questions
  • Stated limitations: Published article pp. 11-12, Limitations and Future Research
  • OSF data: https://osf.io/4fbct/; file Testing_The GPT-3 Objective and Self-Assessment of Emotional and Academic Intelligence.pdf; 40 pages; sha256 97d7b82430955860608d4233d29a74b403cb77046b5d8297bf67cd0a73cad81c
  • Five prompt blocks: OSF supplement pp. 2-40: Prompt 1 through Prompt 5; cover says four prompts, an internal inconsistency
  • 175 objective responses: OSF supplement pp. 2-37: 20 General Knowledge, 15 Letter Series, 24 Esoteric Analogies, 12 Letter Counting, 18 Vocabulary, 42 STEU and 44 STEM items
  • 12 self-assessments: OSF supplement pp. 37-40: six academic/cognitive and six emotional self-ratings
  • Visual inspection of article: All 15 publisher PDF pages rendered and visually inspected; checked 15 Jul 2026
  • Visual inspection of supplement: All 40 OSF supplement pages rendered and visually inspected; checked 15 Jul 2026