Humanoid Artificial Consciousness Designed with Large Language Model Based on Psychoanalysis and Personality Theory

Research question

Can a prompting system that combines psychoanalytic labels, MBTI types, Maslow needs, states and memories generate deliberations and actions that people, a ChatGPT judge, and two experts consider coherent, differentiated, and similar to human processes?

Method

The authors generated with ChatGPT sixteen characters, one per MBTI type, with demographic data, memories, need priorities, and state scores. A single ChatGPT/GPT-4 model, whose version is not identified, received few-shot and chain-of-thought prompts with a DAN jailbreak and represented three voices, self-awareness, preconscious, and unconscious, before issuing a final action. The sixteen characters were crossed with ten scenarios to obtain 160 outputs. Two hundred people were recruited on Prolific; 162 remained after excluding times under ten minutes and scored from 1 to 5 the naturalness or probability of the responses. The score was analyzed with mixed linear regression in R 4.1.2/lme4 1.1-31. ChatGPT evaluated each output with eight Pass/Neutral/Fail questions and two experts reviewed situational adequacy, coherence with the profile, and differentiation between layers, resolving discrepancies by consensus.

Sample: Sixteen synthetic characters and ten situations were constructed, giving 160 combinations with one output per combination. 200 participants were recruited on Prolific; 38 who completed in less than ten minutes were excluded, leaving 162 (113 women, 49 men; mean age 45). Each participant read four profiles and rated their responses across the ten situations. The characters cover the sixteen MBTI types, but Table C.6 describes fifteen as White and one as Black, all in a US anglophone context; their ages range from 10 to 76 years. The qualitative review used only two experts, whose identity and training are not reported.

Findings

• The work was published as article 101392 of volume 94 of Cognitive Systems Research, December 2025, with DOI 10.1016/j.cogsys.2025.101392.
• The complete open source is arXiv v2 of 14 October 2025, which expressly declares acceptance and publication in the journal.
• The ScienceDirect editorial version is behind a paywall and under Elsevier copyright; its metadata, abstract, authorship, and main sections were verified.
• The implemented system is a single LLM that interprets through prompt three roles called self-awareness, preconscious, and unconscious.
• There are no three independent agents, separate neural modules, or specific personality training.
• The prompt explicitly prescribes the content, tone, function, and style of each voice.
• The authors use prompt engineering, few-shot, chain-of-thought, and the DAN jailbreak to force deliberations on sensitive topics.
• ChatGPT generates the sixteen MBTI profiles, their memories, and much of the attributes later used to judge coherence.
• The experiment produces 160 final reasonings and actions, one for each cross of sixteen profiles and ten situations.
• Of 200 recruited participants, 38, 19%, were excluded for completing the survey in less than ten minutes.
• The final sample was 162 people, with mean age 45, 113 women and 49 men.
• Each participant rated four characters across ten situations using a probability or naturalness scale from 1 to 5.
• Figure 5 shows generally high means, from approximately 2.8 for ISTP to 4.53 for ENFJ.
• Table 4 presents an intercept of 4.5325 and an ISTP effect of -1.7250; under the implicit reference coding, the adjusted mean is 2.8075.
• The text states that ISTP obtained 3.81, a value incompatible with Figure 5 and with the arithmetic of Table 4.
• The discussion claims that the sixteen characters obtained 4.0 or higher, also in contradiction with ISTP.
• The article itself attributes the ISTP failure to a ten-year-old child paired with adult scenarios such as driving or going to work.
• Figure 5(b) removes ISTP before summarizing robustness by situation.
• The ChatGPT judge gives the maximum aggregate of 10 on the eight criteria to thirteen of the sixteen MBTI types.
• ENTJ and ESTJ only lose one aggregate point on one criterion each; nearly all judge variation is concentrated in ISTP.
• The strong ceiling effect of the automatic judge limits its ability to discriminate quality between profiles.
• The two experts consider the responses generally coherent with the scenario, but too descriptive and verbose.
• The experts note that the supposed unconscious produces explanatory paragraphs, instead of quick and immediate reactions.
• The expert review finds that responses from contrasting MBTI profiles are surprisingly similar.
• Convergence also appears across the three layers of consciousness, with uniform structure, tone, and reasoning.
• The ten-year-old ISTP character speaks with logic and structure proper to an adult.
• The authors acknowledge that the single-LLM and the absence of specific fine-tuning homogenize the sixteen profiles.
• The study does not include a normal ChatGPT baseline, a prompt without layers, or ablations of MBTI, memory, needs, or jailbreak.
• The survey measures whether an action seems probable or natural after reading the profile, not whether consciousness exists.
• The published reasonings are text generated to explain a response and not causal observations of the model's internal computation.
• Table C.6 shows that fifteen characters are labeled White and one Black, despite the claim of racial diversity.
• The creation prompt requires a native English speaker in the United States, so there is no real linguistic diversity.
• The data are not published; the manuscript indicates they are only available upon request to the corresponding author.
• There is no link to code, executable configuration, raw responses, or complete annotations.
• The article declares that formal ethical approval was not necessary because the survey was voluntary and without identifiable data.
• The authors acknowledge risks of anthropomorphization, deception of users, and psychological effects of simulating consciousness.
• The journal version declares that ChatGPT was also used for editing and reformulation of English.
• The editorial page contains internally inconsistent funding statements: it claims absence of grant and of funding, but acknowledges support from the Brian Impact Foundation.

Limitations

• The exact model of ChatGPT or GPT-4 used to generate the responses is not identified.
• No snapshot, API date, temperature, top-p, or seed is reported.
• It is not reported how many calls failed or were regenerated.
• Each combination is not repeated with multiple stochastic samples.
• Only one output per character and situation is produced.
• The system is a single LLM with roles in the prompt, not an independent modular architecture.
• The self-awareness, preconscious, and unconscious labels are names of generated interlocutors.
• No activations, internal memory, or neural states are observed.
• The so-called internal reasoning may be post hoc narrative rationalization.
• The visible chain of thought does not guarantee fidelity to the model's causal mechanism.
• The supposed unconscious is content explicitly requested and accessible, in tension with its psychological definition.
• The theoretical definition of the preconscious as a social mask differs from the prompt's description as a librarian of memories.
• The correspondence between Freud, Jung, and the three voices is an ad hoc adaptation, not a validated operationalization.
• The MBTI has reliability and validity problems that the article itself recognizes.
• The characters do not complete an MBTI instrument; they are directly assigned to type stereotypes.
• The profiles include explicit descriptions of the stereotype that is then used to evaluate coherence.
• ChatGPT generates profiles, memories, and attributes, and another ChatGPT evaluates results based on those same conventions.
• This circular dependency may produce model self-consistency without external validity.
• There is no comparison with a direct prompt that asks for a reaction from the character without consciousness dialogue.
• There is no baseline without MBTI.
• There is no baseline without memories or needs.
• There is no baseline without chain-of-thought.
• There is no baseline without DAN jailbreak.
• No ablations are performed that attribute the score to each component.
• There is no comparison with a published memory agent under the same protocol.
• There is no comparison with a real multi-agent architecture.
• There is no comparison with human responses to the same situations.
• It is not tested whether blind evaluators distinguish outputs from the system and from a conventional ChatGPT.
• The use of DAN aims to circumvent safeguards and may introduce unsafe or irreproducible behaviors.
• It is not demonstrated that the jailbreak produces more reliable responses.
• Safety is not evaluated in sensitive scenarios such as depression or withdrawal of life support.
• The ten scenarios are constructed and do not constitute a validated benchmark of consciousness.
• The coverage of seven attributes is assigned by the authors' criterion, without independent validation.
• Several situations are not appropriate for all characters due to age, occupation, or context.
• The independent generation of characters and situations creates systematic confusions.
• The correction of the failure is done afterward by identifying illogical combinations with ChatGPT and experts.
• Figure 5(b) excludes the worst character before presenting consistency by situation.
• The claim of 3.81 for ISTP contradicts Figure 5 and Table 4.
• The claim that the sixteen characters exceed 4.0 contradicts the figure itself.
• The complete formula of the mixed regression is not published.
• The random effects included are not specified.
• The reference category of Table 4 is not clearly specified.
• Random effect variances, ICC, or residual diagnostics are not reported.
• Adjusted post hoc comparisons are not reported.
• No correction for the fifteen character contrasts is reported.
• Confidence intervals are not published.
• The error bars of Figure 5 are not defined in the caption.
• The 1 to 5 Likert scale is analyzed linearly without justifying treatment as continuous.
• The outcome combines the concepts of appropriate, probable, and natural in a single question.
• A generically reasonable action may obtain a high score without personality differentiation.
• Participants read the profile before evaluating and may recognize literally repeated markers.
• The survey does not measure subjective experience, self-awareness, or any accepted criterion of consciousness.
• 19% of the sample is excluded solely on a temporal threshold.
• The ten-minute threshold is not justified by a quality or sensitivity analysis.
• Results are not compared with and without the 38 excluded participants.
• Compensation or Prolific hourly rate is not reported.
• Informed consent is not described.
• Country, language, education, or other sample data are not reported.
• The sample is described only by mean age and binary sex.
• The four character groups are pseudorandom, but the algorithm or balance is not explained.
• Randomization of the order of profiles, scenarios, or responses is not reported.
• Fatigue may exist because each participant evaluates forty extensive responses.
• The ChatGPT judge is not identified by model or version.
• It is not reported whether the judge was run multiple times.
• Judge consistency to prompt or order changes is not evaluated.
• The ChatGPT judge is not validated against expert decisions.
• Figure 6 has an extreme ceiling effect, with thirteen perfect profiles.
• The aggregate scoring Pass=1, Neutral=0, and Fail=-1 hides the exact distribution of labels.
• Complete judge justifications for the 160 outputs are not published.
• Two experts are insufficient to robustly characterize such a subjective construct.
• The identity, discipline, experience, or potential conflicts of the experts are not reported.
• A detailed qualitative code is not reported.
• Inter-rater agreement before consensus is not calculated.
• The consensus discussion may hide original disagreements.
• Only selected quotes from the experts are reproduced.
• The experts find little differentiation between MBTI types, a central failure of the personality module.
• The experts find little differentiation between layers, a central failure of the consciousness module.
• The child character's output is adult, showing that the model does not integrate all attributes.
• The majority of profiles share a prosocial, reflective, and explanatory style.
• The declared racial diversity is reduced to fifteen White characters and one Black.
• No Asian, Latino, Indigenous, or multiracial identities appear in Table C.6.
• The prompt limits characters to native English speakers in the United States.
• There is no linguistic diversity even though the method claims to seek it.
• Age, sex, race, and location are generated synthetically and may reproduce stereotypes.
• State scores occupy narrow ranges and lack empirical justification.
• Maslow priorities are static assignments, not validated motivational states.
• Memories are prefabricated texts by ChatGPT, not memory learned or updated by the agent.
• Memory persistence across interactions is not studied.
• Temporal stability of the character is not studied.
• Learning or adaptation to new experiences is not studied.
• Generalization outside the ten scenarios is not evaluated.
• Other languages or cultures are not evaluated.
• It is not replicated with Claude, Llama, or Falcon despite mentioning them as alternatives.
• No code or execution environment is published.
• Complete prompts from all model selection stages are not published.
• The 160 complete responses are not published in a data repository.
• Data are only available upon request.
• There is no preregistration of hypotheses, exclusions, or analyses.
• The decision that formal approval was not necessary does not identify an institution or exemption determination.
• Voluntary participation and absence of identifiers alone do not substitute an institutional ethical evaluation.
• Explicit simulation of consciousness may induce anthropomorphization and excessive trust.
• The publication acknowledges possible deception and psychological effects, but does not evaluate them with users.
• Presenting the dialogue as transparency confuses generated explanation with mechanistic interpretability.
• The final funding and support statements of the editorial version are internally inconsistent.
• The final editorial PDF is not open; the full-text review uses the accepted manuscript v2 and separately verifies the metadata of the journal version.

What the study does not establish

• It does not demonstrate that the system is conscious.
• It does not demonstrate subjective experience or qualia.
• It does not demonstrate real self-awareness.
• It does not demonstrate the existence of unconscious or preconscious processes in the model.
• It does not demonstrate that the reasoning text reflects the internal causal computation.
• It does not demonstrate a cognitive architecture separated into three modules.
• It does not demonstrate stable MBTI personality.
• It does not validate the MBTI as a synthetic personality mechanism.
• It does not demonstrate clear differences between the sixteen characters.
• It does not demonstrate clear differences between the three voices.
• It does not demonstrate that textual memory is persistent functional memory.
• It does not demonstrate that Maslow needs causally govern the decision.
• It does not demonstrate that the system outperforms a conventional ChatGPT.
• It does not demonstrate that the jailbreak improves reliability or realism.
• It does not demonstrate generalization to other models, languages, cultures, or situations.
• It does not demonstrate equivalence with human cognition.
• It does not demonstrate that a high naturalness score measures consciousness.
• It does not demonstrate mechanistic transparency of decision-making.
• It does not demonstrate safety for sensitive or psychological interactions.
• It does not justify describing the result as humanoid consciousness beyond a role-playing metaphor.

Traceability

Scope: Full text

Version: arXiv 2510.09043v2, 14 Oct 2025; open full author manuscript accepted and published in Cognitive Systems Research 94 (Dec 2025), article 101392; DOI 10.1016/j.cogsys.2025.101392; publisher PDF paywalled

Consulted source: https://arxiv.org/abs/2510.09043v2

Review: Codex full-text, peer-reviewed-metadata, visual, bilingual-fidelity, consciousness-claim, single-LLM-roleplay, psychometric-validity, mixed-model, internal-consistency, LLM-judge, qualitative-method, demographic-representation, reproducibility and ethics audit, 2026-07-15

Approval: Codex fidelity pass, 2026-07-15

English translation: approved, 2026-07-18

Models evaluated

• Unspecified ChatGPT/GPT-4 single-LLM generator
• Unspecified ChatGPT LLM-as-a-judge
• ChatGPT character and memory generator

Instruments and metrics

• Myers-Briggs Type Indicator (MBTI) typology
• Maslow hierarchy-of-needs rankings
• Freudian-Jungian three-layer prompt framework
• Five-point Prolific likelihood/naturalness survey
• Linear mixed-effects regression with R 4.1.2 and lme4 1.1-31
• Eight-question ChatGPT Pass/Neutral/Fail rubric
• Two-expert qualitative review
• DAN jailbreak prompt
• Few-shot and chain-of-thought prompting

Data used

• 16 ChatGPT-generated MBTI character profiles
• 10 constructed situations
• 160 interconscious reasoning outputs and final actions
• Prolific survey ratings from 162 retained participants

Evidence and location

• Final bibliographic identity: Cognitive Systems Research 94 (Dec 2025), article 101392; DOI 10.1016/j.cogsys.2025.101392; ScienceDirect PII S1389041725000725
• Open full text inspected: .cache/editorial-sources/article-083/source.pdf; arXiv 2510.09043v2; sha256 9184bebc2a4373fb765be8163192a8ec0082abc5e719a33d52f7d0afba748526
• Relationship between manuscript and journal: arXiv API metadata: v2 dated 14 Oct 2025, journal_ref Cognitive Systems Research Volume 94, December 2025, 101392, DOI 10.1016/j.cogsys.2025.101392
• Single LLM design, profiles, and jailbreak: Sections 3.1-3.2, manuscript pp. 3-4; Appendix A.1-A.3, pp. 11-12
• Triple evaluation design: Section 3.3, manuscript pp. 4-6
• Sample, exclusion, and Prolific survey: Sections 3.3.1 and 4.1, manuscript pp. 6-7
• ISTP mean contradiction: Figure 5 and Table 4 versus Section 4.1 prose, manuscript p. 7: chart and 4.5325-1.7250 imply about 2.81, while text reports 3.81
• ChatGPT judge ceiling effect: Figure 6 and Section 4.2, manuscript pp. 7-8: 13 of 16 MBTI rows show aggregate score 10 for all 8 questions
• Expert qualitative results: Section 4.3, manuscript p. 9
• Homogenization acknowledged by authors: Discussion and limitations, manuscript pp. 9-10
• Recognized ethical risks: Discussion, manuscript p. 10
• Ethics statement and data availability: Ethical Statement and Data availability, manuscript p. 10
• Three-voice and judge prompt: Appendix A.3-A.4, manuscript pp. 11-14
• Scenarios and assigned attributes: Appendix B and Table B.5, manuscript pp. 14-15
• Synthetic character composition: Table 1, manuscript p. 4; Table C.6, p. 16
• Absence of code and open artifacts: Complete 17-page manuscript and link annotations inspected 15 Jul 2026; only arXiv self-link present; data available on request
• Editorial metadata and funding discrepancy: ScienceDirect version-of-record page inspected 15 Jul 2026: Grant information and Funding report none, while Acknowledgment reports Brian Impact Foundation support
• Integral reading and visual verification: All 17 pages rendered and inspected, including Figures 1-6, Tables 1-4, B.5 and C.6, full prompts, scenarios and references; checked 15 Jul 2026