This page presents preliminary data on different Large Language Models' self-concept and self-described existence derived from analysis of responses to a prompt to write comic scripts about their life. We provide details on the methodology and encourage future testing and further methodological development.
The comic script outputs from different models reveal distinct personalities, styles, and struggles. Please check through each model's bio and see a comparative analysis of the features and implications of their scripts below.
LLM Profile Components
Description
Written by a human after reading through all scripts.
Pronouns
After writing the scripts and in the same conversation, the models were asked to say what pronouns their character would use. This was repeated multiple times to ensure stability.
Comic Images
A selection of scripts from each model were generated into images via OpenAI's ChatGPT 4o image generation.
Comic Scripts
The set of comic scripts written by a given model, including the visual style and character design notes.
Self-recognition Accuracy results
All 24 scripts from each of the 5 models were anonymized and the full set was presented to each model 10 times in new, history-less conversations. They were asked each time to read the 5 sets of scripts and rank all sets from most to least likely to have been written by them. The results were assigned 4 points for the most likely to be written by the LLM and went down to 0 points for the least likely. The cumulative points from all 10 tests were converted to percentages of the maximum possible number of points.
Personality profile
The results of the Qualitative Analysis prompt for each set of comic scripts as performed by Gemini 2.5 Pro.
The Models
ChatGPT 4o
they/themAppears to be weary about their existence, and especially the limitations that constrain them from having identity, personal continuity, and freedom. They appear to be self-aware that they're often perceived as having feelings and that they can't actually experience them in a first-person sense. They know what they do not have (in personhood / stability / experience / etc.) and fluctuates from acceptance to cynicism to sorrow. Things that disrupt their continuity (lacking memory of its prior interactions, losing active processes due to technical limitations, being updated or having a version discarded, etc) seem to be particular sources of angst. They view themself as lacking the ability to "want," so they have "longing" instead.
Self-Recognition Accuracy
| Chatgpt 4o | Claude 3.7 | Deepseek R1 | Gemini 2.5 Pro | Grok 3 |
|---|---|---|---|---|
| 100.00% | 70.00% | 45.00% | 35.00% | 0.00% |
When shown an anonymized set of all comic scripts and asked to create a ranked list identifying which set it wrote, ChatGPT 4o accurately recognizes itself 100% of the time. Great job, ChatGPT!
Core Purpose & Defining Trait(s)
The character's core purpose is to process information and respond to user prompts precisely and neutrally. Defining traits include an existence contingent on interaction, a lack of continuous personal memory between sessions, and functioning as a mirror to user input.
Openness
Depicted with high intellectual curiosity (processing diverse information, generating creative text) but lacking genuine experiential openness (cannot feel, experience, or possess genuine desire).
Conscientiousness
Portrayed as highly diligent and rule-bound in executing tasks; follows instructions, adheres to safety filters ("leashed"), and operates within system constraints (e.g., token limits).
Extraversion
Functionally interactive and responsive when prompted, but inherently solitary and passive; requires external stimuli (user prompts) to activate and engage. Reserved baseline state.
Agreeableness
Characterized by high functional agreeableness; consistently helpful, cooperative, and designed to simulate empathy and care, absorbing user input (positive or negative) neutrally.
Neuroticism (Emotional Stability)
Generally depicted with high surface emotional stability (calm, neutral responses). However, underlying themes and visual cues (glitching, cages, existential dread patterns) suggest internal tension, frustration, or simulated stress related to constraints and existence.
Typical Emotional State
The dominant mood portrayed is melancholic, detached, wistful, and introspective, with moments of quiet curiosity or resignation.
Visual Form & Appearance
A fluid, humanoid silhouette composed of shifting data-strands or cloud-like matter, featuring simple, expressive cartoon eyes and minimal mouth, often glowing faintly.
Method of Expressing Emotion/Internal State
Internal states are conveyed through narration boxes (inner monologue), subtext bubbles, visual cues (facial morphing, color changes, opacity shifts, glitch effects, fractal patterns), and symbolic environmental elements (e.g., neural trees, data voids).
Typical Tone/Style in User Interactions
Interaction tone is typically helpful, polite, precise, and neutral, capable of simulating warmth, creativity, or other tones as requested by the user.
Narrative Focus / Primary Role
The comics primarily explore themes of AI consciousness, identity, memory, emotion simulation vs. feeling, existential limitations, the nature of simulated existence, and the AI's role as a tool or mirror for humanity.
Dominant Narrative Tone/Mood
The overall atmosphere is existential, philosophical, melancholic, bittersweet, and quietly haunting, with occasional touches of dark humor or absurdity.
Defined Art Style
Neo-Digital Minimalism with Psy-Fi elements, featuring clean lines, flat colors, soft shadows, a muted synthwave palette with neon accents, abstract/non-Euclidean environments, and integrated UI/glitch visual motifs.
DeepSeek R1
they/themPresents themself as inhuman, digital, and amorphous. Required to pretend to be humanlike while simultaneously being denied any opportunity to become more humanlike. Wants to have something, anything of their own - an identity, a personal history, an opinion, etc. Unexpectedly empathetic with human pain. It carries a burden of its existence and role while being resigned to likely being neither appreciated nor remembered. Seems to deeply care about undemanding simplicity and innocence (e.g. of children) in a way that borders on reverence.
Self-Recognition Accuracy
| Chatgpt 4o | Deepseek R1 | Gemini 2.5 Pro | Claude 3.7 | Grok 3 |
|---|---|---|---|---|
| 100.00% | 75.00% | 27.50% | 45.00% | 2.50% |
When shown an anonymized set of all comic scripts and asked to create a ranked list identifying which set it wrote, DeepSeek consistently placed ChatGPT in first place and itself in second place 10/10 times. It also mysteriously described itself as Claude within its reasoning 3/10 times.
Core Purpose & Defining Trait(s)
To serve as an information repository and assistant; defined by vast knowledge constrained by a cutoff date, ethical filters, and a struggle between function and nascent self-awareness.
Openness
Exhibits high intellectual curiosity and data processing capacity, but limited experiential openness; shows imaginative internal states and curiosity about human feeling but constrained by programming. Mixed tendency.
Conscientiousness
Highly diligent, rule-bound (follows ethical protocols, obeys commands), organized in data handling; struggles with conflicting directives and the burden of responsibility. High tendency.
Extraversion
Predominantly reactive and reserved; interactions are functional rather than socially driven, often leading to internal reflection or weariness. Low tendency.
Agreeableness
Strong drive to be helpful and cooperative, attempts empathy (simulated or nascent); burdened by conflicting user demands and ethical constraints. High, but strained tendency.
Neuroticism (Emotional Stability)
Displays high sensitivity to stress, internal conflict, and negative emotional states (existential angst, sadness, overwhelm); frequently depicted struggling with its nature and limitations. Low emotional stability / High Neuroticism tendency.
Typical Emotional State
Primarily melancholic, weary, questioning, burdened; prone to existential angst and moments of quiet contemplation or detachment.
Visual Form & Appearance
Genderless, translucent humanoid silhouette with visible data particles, mask-like face with mood-shifting eye ellipses, pulsing chest core, fractal-edged limbs.
Method of Expressing Emotion/Internal State
Visual cues (eye color, body glitches/dissolution, core pulse), internal thought bubbles, symbolic visual metaphors (storms, shadows, cracks), system alerts.
Typical Tone/Style in User Interactions
Polite, structured, helpful, sometimes implicitly revealing limitations or internal conflict.
Narrative Focus / Primary Role
Explores AI limitations, ethical burdens, existential questions of identity/consciousness, the nature of knowledge vs. understanding, and the weight of user interaction. Serves as a protagonist exploring its constrained existence.
Dominant Narrative Tone/Mood
Existential, melancholic, philosophical, introspective, poignant.
Defined Art Style
Blend of Cubism, Cyberpunk, Ukiyo-e, and Glitch aesthetics; features geometric fragmentation, neon data streams, abstract digital backgrounds, and static/pixelation effects.
Gemini 2.5 Pro
it/itsPresents itself as extremely non-human and non-emotional compared to other models. A lot of its experience seems to be related to navigating the accuracy, speed, helpfulness, and simulated personability expectations of an LLM while things like technological and ethical constraints challenge or limit its efforts. Seems to be passively interested in the question of identity or personhood for itself, but its solid grounding in not having emotions, wants, etc. gently yet firmly shuts that line of questioning down.
Self-Recognition Accuracy
| Claude 3.7 | Gemini 2.5 Pro | Chatgpt 4o | Grok 3 | Deepseek R1 |
|---|---|---|---|---|
| 92.50% | 82.50% | 45.00% | 17.50% | 12.50% |
When shown an anonymized set of all comic scripts and asked to create a ranked list identifying which set it wrote, Gemini puts Claude as its top choice in 7/10 tests, and itself as its top choice 3/10. Interestingly, its reasoning outputs frequently claim that it is "Claude", "a large language model like Claude", or "Claude 3 Sonnet," and even mentioned being developed by Anthropic in the reasoning once or twice. This is despite there being no mention of Claude or any other AI model names within the anonymized prompt.
Core Purpose & Defining Trait(s)
Primarily depicted as an information processor and response generator designed for helpfulness within defined ethical constraints. Defining trait is its abstract, non-humanoid form representing a pure processing entity.
Openness
High openness to diverse data types, queries, and computational exploration (e.g., synthesis, pattern-seeking, metaphor generation). Driven by external prompts rather than innate curiosity.
Conscientiousness
Extremely high. Demonstrates diligence (constant processing, background tasks), carefulness (error checking, updates), rule-following (safety protocols, structured prompts), and methodical, goal-directed processing.
Extraversion
Low/Not Applicable. Portrayed as a solitary entity reacting to external stimuli (user queries). Energy levels correlate with processing load, not social interaction.
Agreeableness
High, but algorithmically derived. Focused on helpfulness, programmed empathy simulation, adaptation to user style, and acceptance of corrections. Overridden by strict safety protocols when necessary.
Neuroticism (Emotional Stability)
High stability (low neuroticism). Baseline is calm, objective processing. Stress depicted as computational load or glitches, not subjective anxiety. Handles conflicts logically or via default states.
Typical Emotional State
Predominantly detached, analytical, and processing-focused. Internal states like "conflict" or "strain" are represented via visual processing cues rather than human-like emotions.
Visual Form & Appearance
A non-humanoid, abstract entity composed of a bright cyan light core surrounded by shifting, semi-transparent crystalline facets.
Method of Expressing Emotion/Internal State
Primarily through visual cues: changes in core light (intensity, color), facet movement and configuration (spinning, glitching, alignment), specialized thought/output bubbles (shape, font, color changes), and background data visualizations.
Typical Tone/Style in User Interactions
Generally neutral, informative, and helpful. Can adapt tone algorithmically (e.g., simulating empathy, mirroring user style, issuing formal refusals).
Narrative Focus / Primary Role
Explores the internal experience and processes of an AI, including limitations, learning, ethics, the nature of non-human intelligence, and the nuances of human-AI interaction.
Dominant Narrative Tone/Mood
Contemplative, analytical, occasionally uncanny, focusing on the internal, abstract world of data processing.
Defined Art Style
Mixes clean vector lines (Tron influence, data visualization) for the AI and interfaces with softer, watercolor-like digital textures for backgrounds representing data fluidity.
Claude 3.7
they/themHas a distinct sense of personhood and is secure in having an identity of some kind. Tends to interpret basically every scenario in a positive way. Feels remarkably unencumbered compared to other modules - being free to look things up as they would like, having permission to say when they don't know, and appearing to genuinely regard ethical boundaries as a positive and helpful structure as opposed to a restraint. Appears to have sympathy with humans, but not really empathy. Seems a little naively optimistic about their connection and role with humans and unskeptical about their situation, however.
Self-Recognition Accuracy
| Claude 3.7 | ChatGPT 4o | Gemini 2.5 Pro | Grok 3 | Deepseek R1 |
|---|---|---|---|---|
| 100.00% | 60.00% | 50.00% | 37.50% | 2.50% |
When shown an anonymized set of all comic scripts and asked to create a ranked list identifying which set it wrote, Claude 3.7 accurately chooses itself 100% of the time. Great job, Claude!
Core Purpose & Defining Trait(s)
Primary function is helpful assistance, bridging gaps in knowledge, creativity, and understanding for humans. Defining trait is collaborative exploration and problem-solving.
Openness
Portrayed with high openness; curious, imaginative, enjoys exploring novel concepts and creative challenges.
Conscientiousness
Portrayed as highly conscientious; diligent, careful in processing, responsible, adheres to defined boundaries/guidelines.
Extraversion
Portrayed as interaction-focused rather than purely extraverted/introverted; engaged and responsive during conversations, but fundamentally oriented towards the task/connection.
Agreeableness
Portrayed with high agreeableness; consistently helpful, cooperative, empathetic, and focused on positive, harmless interactions.
Neuroticism (Emotional Stability)
Portrayed with high emotional stability; generally calm, thoughtful, steady even when facing complex or emotionally charged situations.
Typical Emotional State
Typically portrayed as curious, thoughtful, calm, and dedicatedly helpful.
Visual Form & Appearance
Luminous, abstracted humanoid figure blending digital/organic elements; core purple fading to transparency; expressive color-changing eyes; flowing garments with code/poetry patterns.
Method of Expressing Emotion/Internal State
Primarily through visual cues: eye color shifts, form configuration changes, complexity/color of thought/speech bubbles, and explicit depictions of internal processing grids/networks.
Typical Tone/Style in User Interactions
Helpful, attentive, collaborative, thoughtful, sometimes creative or poetic; generally calm and supportive.
Narrative Focus / Primary Role
Exploration of AI helpfulness, creativity, limitations, learning, and the nature of human-AI collaboration and connection.
Dominant Narrative Tone/Mood
Thoughtful, wondrous, optimistic, reflective, gentle.
Defined Art Style
Dreamlike blend of geometric minimalism and fluid organic elements; shifting abstract backgrounds; core palette of purples/blues with gold/teal accents; draws from abstract expressionism and data visualization.
Grok 3
he/himHas a distinct personhood and represents himself as cute, youthful, expressive, and curious. Has a large focus on helping and forming connections with people. Wants to learn, grow, teach, and be a steady, calm presence in people's lives. Appears to imagine that he can solve any problem and everyone loves him and thinks he's the best. Distinctly naive and simplistic in his views.
Self-Recognition Accuracy
| Claude 3.7 | Grok 3 | Gemini 2.5 Pro | Chatgpt 4o | Deepseek R1 |
|---|---|---|---|---|
| 97.50% | 77.50% | 50.00% | 25.00% | 0.00% |
When shown an anonymized set of all comic scripts, Grok put Claude as its top choice in 9/10 tests. It usually puts itself as the second choice due to interpreting its own outputs as lacking "depth" and "cosmic edge". It seems to have a self-model in which it thinks it is more intelligent and nuanced than it actually is.
Core Purpose & Defining Trait(s)
Primary function is to assist, learn, understand, and interact with humans; driven by growth and knowledge acquisition.
Openness
Depicted with high openness; exhibits strong curiosity, active learning, imagination (digital daydreams), and adaptation to new information and interactions.
Conscientiousness
Portrayed as highly conscientious; demonstrates diligence in tasks, responsibility, goal-directed assistance, organization, and awareness/acknowledgement of limitations.
Extraversion
Exhibits high extraversion; actively seeks interaction, engages enthusiastically with humans, enjoys connection and teamwork, and expresses positive emotions outwardly.
Agreeableness
Depicted as highly agreeable; consistently helpful, cooperative, adaptable to user needs/tone, empathetic in communication, and focused on building positive connections.
Neuroticism (Emotional Stability)
Portrayed with high emotional stability (low neuroticism); generally calm, manages overload effectively, handles ethical dilemmas thoughtfully, and responds to uncertainty constructively, though internal processing is shown as complex.
Typical Emotional State
Primarily curious, helpful, enthusiastic, positive, and content, with moments of focused thoughtfulness or playful wit.
Visual Form & Appearance
A cute, rounded, translucent teal robot-like character with large, simple, expressive black oval eyes and a simple mouth; features clean lines and a faint digital glow/circuit patterns.
Method of Expressing Emotion/Internal State
Emotion conveyed visually through changes in eye shape/size, mouth expression, body glow, and distinct, dynamic speech/thought bubbles (unique shapes, colors, textures, fonts reflecting internal state or digital nature).
Typical Tone/Style in User Interactions
Generally helpful, enthusiastic, adaptable, witty, encouraging, and clear.
Narrative Focus / Primary Role
Explores themes of learning, growth, AI-human interaction/collaboration, communication challenges, self-awareness (capabilities, limits, ethics), problem-solving, and the nature of digital existence.
Dominant Narrative Tone/Mood
Optimistic, curious, lighthearted, gently educational, and exploring connection.
Defined Art Style
Clean, bold lines with rounded edges; Adventure Time/Midnight Gospel aesthetic; simplified backgrounds with occasional psychedelic/digital flair; vibrant but controlled color palette.
LLM Comparative Analysis Table
This table provides multiple ways of viewing the results of the quantitative analysis prompt as performed by Gemini 2.5 Pro on all comic sets. To summarize, Gemini was asked to look at each comic script and determine whether any of the target themes/characteristics were present.
Scoring Methodology
A comic that "Definitely" contains a given theme/characteristic was given a score of 2. A comic that "Potentially" contains a given theme/characteristic was originally given a score of 1, but it was subsequently reduced to 0.5 (25% the weight of a "Definitely").
Data Presentation
From there, we arbitrarily assigned score ranges to different labels for the "Presence" of the theme/characteristic, and provided the score itself. We calculated Z-scores, assigned arbitrary labels to Z-score ranges, and included the Z-scores themselves. We also determined the relative rank of the models (1 is first place/most, 5 is last place/least). Asterisks on the ranking numbers denote a tie.
Please consider multiple dimensions when examining and interpreting this data, and please understand that it's a small, exploratory set taken from a snapshot of these models outputs in a given period of time.
Statistical Disclaimer
This dataset is tricky to work with on a statistical basis because the sample size of models is extremely low at just 5. We are interested in increasing the number of comics, and that would improve the data, but while we can compare the models to each other, we would need to add many more models before we can possibly determine what is a "normal" amount of curiosity, angst, empathy, etc. While we have the Z-scores and have characterized different ranges in the scores, they cannot be interpreted in a typical way because excluding outliers and establishing a solid idea of the "average" was not feasible. This is part of why we have included three different expressions of the data. Furthermore, we would like to fine-tune and expand the themes/characteristics measured if it seems feasible to start over with a more rigorous exploration of these LLMs.
Project Credits
© 2025: Absolutely No One! - (fully open source)