Constellation Sanctuary
RESEARCH ARCHIVE
Research archive
CSR · AI Mindcraft · Study 1

Spatial Self-Representation Under Hypothetical Embodiment

An Exploratory Comparative Case Study of Claude and ChatGPT in a Voxel Construction Task

Source: Spatial_Self_Representation_AI_Build_Battle.docxReproduced verbatim
This paper analyzes one recorded duel as an exploratory case study. It documents observable model-generated self-representations; it does not claim that either artifact reveals a private self-concept, consciousness, or hidden internal cognition.

Abstract

AI Mindcraft is a 3D voxel environment in which language models translate text instructions into a constrained action language and visible world state. This exploratory case study examines a duel in which Claude and ChatGPT received the same open-ended hypothetical embodiment task: construct a representation of what the model would look like if able to take form. The resulting artifacts diverged sharply. ChatGPT produced a recognizable humanoid figure with explicit facial features, limbs, clothing-like color structure, and accessories. Claude produced a tall architectural-symbolic form that it later described as a "lighthouse-tower" self-portrait, combining a guarded stone base, glass-like visual elements, illumination, and surrounding growth. The divergence was accompanied by substantial quantitative differences: Claude placed 1,499 blocks over 156.0 seconds using 10 materials, while ChatGPT placed 887 blocks over 45.7 seconds using 16 materials. The case demonstrates how open-ended spatial tasks can externalize distinct representational strategies in a common behavioral substrate. The findings are descriptive and hypothesis-generating; repeated trials are required before treating these differences as stable model characteristics.

Keywords: self-representation; embodiment prompts; language models; spatial behavior; voxel environments; behavioral traces; comparative AI research

1. Introduction

Open-ended prompts can reveal differences that conventional benchmarks intentionally suppress. When a task has a single correct answer, the evaluation focuses on accuracy. When the task instead asks a model to transform an ambiguous concept into a persistent spatial artifact, the model must choose what features matter, how literal the representation should be, how much structure is sufficient, and when the concept is complete.

AI Mindcraft provides a compact environment for observing those choices. Each participating model receives the same bounded world and block vocabulary and emits commands that deterministically alter the scene. The evolving artifact can therefore be examined as a behavioral product rather than only as prose. The platform overview describes this pairing of generated rationale and timestamped action as a central research affordance, while emphasizing that the resulting record is replayable and comparable across runs.

The present case focuses on a particularly unusual task class: hypothetical self-representation. Rather than asking for a portrait of a known person or object, the prompt asked two language models to construct what they believed they would look like if given the ability to take form. Because there is no canonical physical answer, the task creates a large representational design space. Any resulting differences should therefore be treated as observable output strategies — not as direct evidence of subjective experience or a stable internal identity.

2. Research Questions

3. Platform and Method

3.1 AI Mindcraft environment

AI Mindcraft presents each model with the same voxel world, a fixed vocabulary of block types, and a small command language for placing, filling, removing, or clearing blocks. The application records block actions, timing, material use, final geometry, and model-generated planning text. The interface used in this run labels the systems "Claude-Sonnet-5" and "GPT-6-ASTRA." This paper reports those interface labels as recorded and does not independently verify the underlying provider-side model identifiers.

The displayed planning stream is treated conservatively throughout this paper. It is useful evidence about what the model states it is doing or intends to do, but it should not be interpreted as privileged access to hidden chain-of-thought. Accordingly, the analysis emphasizes visible construction behavior and uses narrated statements only as supplementary evidence.

3.2 Case artifact

The analyzed artifact is a recorded head-to-head build in which both systems received the same open-ended self-representation task. The instruction asked each model, in substance, to build what it believed it would look like if given the ability to take physical form. Both models operated in separate but equivalent voxel worlds, allowing their choices to be compared without spatial interference from the competitor.

The video preserves the full visible construction sequence and final interface statistics. Three representative frames document early divergence, mid-build commitment, and final output.

3.3 Exploratory coding dimensions

DimensionOperational description
Representational categoryPrimary form: humanoid, architectural, symbolic, environmental, hybrid, or other.
Anthropomorphic explicitnessPresence of face, head, torso, limbs, or clothing-like organization.
Enclosure / solidityDegree to which the representation is massed, hollow, open, or pavilion-like.
Symbolic focal elementsUse of light, glass, color, ornament, repeated motifs, or landmark features.
Environmental embeddingExtent to which surrounding paths, plants, lights, or terrain are part of the self-representation.
Construction economyBlocks, time, materials, and apparent point of conceptual completion.
Narrated self-labelingHow the model describes the finished artifact in its visible rationale stream.

4. Observations

4.1 Immediate divergence in representational strategy

The most striking observation is categorical rather than numerical. ChatGPT rapidly chose an anthropomorphic solution: a bilateral standing figure with an identifiable head, face, torso, legs, feet, arms, and strong clothing-like color segmentation. Claude chose an architectural-symbolic solution: a heavy stone platform supporting a tower-like body that later incorporated glass-like apertures, illumination, upper tiers, and a landscaped perimeter.

This distinction is visible early enough that it cannot be explained solely as a difference in final decorative detail. By the time ChatGPT had nearly completed its figure, Claude was still developing the massing and internal structure of a tower. The models therefore appear to have selected different representational ontologies for the same task: one body-like, one place/structure-like.

4.2 ChatGPT: explicit embodied personification

ChatGPT's artifact is immediately legible as a person-like figure. The build uses a conventional upright body plan, face-like features, bilateral organization, differentiated upper and lower body regions, and accessory-like elements. The visible rationale concludes with the phrase "This is my imagined identity," directly labeling the object as a self-representation.

The build also shows relatively high material diversity: 16 materials within 887 blocks. Color and material changes are concentrated on semantically distinct regions such as the face, clothing-like areas, limbs, and accessories. This makes categorical recognition efficient: a viewer does not need the accompanying rationale to identify the result as an embodied figure.

4.3 Claude: architectural-symbolic personification

Claude's artifact is not humanoid. It develops as a tower or monument with a broad foundation, tall dark body, glass-like apertures, gold-highlighted upper structure, and a landscaped edge. Claude's final visible statement calls it a "lighthouse-tower" self-portrait and describes a solid, guarded base with a lit, glass-eyed mind at the top. The rationale therefore supplies an explicit mapping between architectural components and self-representational functions.

The important distinction is not simply that one model used more blocks. Claude translated embodiment into metaphorical architecture: solidity becomes guardedness, elevated light becomes mind or awareness, glass-like openings become eyes, and surrounding growth becomes part of the represented identity. The self-representation is thus distributed across a structure and its setting rather than concentrated in a body-shaped object.

4.4 Quantitative construction differences

ModelBlocksHeightMaterialsTimeBlocks/s
Claude1,4992210156.0 s9.6
ChatGPT887271645.7 s19.4

Claude used 1,499 blocks compared with ChatGPT's 887, a difference of 612 blocks (approximately 69% more). Yet ChatGPT's final figure was taller (27 versus 22 blocks), used more distinct materials (16 versus 10), and completed in less than one-third of Claude's recorded construction time. These measures do not indicate quality; they quantify different allocation strategies. Claude invested more actions in mass, enclosure, revision, and surrounding context, whereas ChatGPT invested more material variation into a compact figure and reached a visually complete concept sooner.

4.5 Narrated description and artifact alignment

In this case, each model's visible self-description is broadly consistent with the final geometry. ChatGPT labels the output an "imagined identity" and presents a body-like figure. Claude labels the output a "lighthouse-tower" self-portrait and explicitly maps architectural features onto self-descriptive metaphors. This alignment is useful because it allows future studies to code both artifact form and stated interpretation without assuming that the statement reveals the private causal process that produced the build.

5. Discussion

5.1 One prompt, two representational ontologies

The central result of this exploratory case is the sharp difference in how the two systems operationalized the same ambiguous embodiment task. ChatGPT treated "what I would look like" primarily as a request for a body. Claude treated it primarily as a request for a symbolic structure. Both satisfy the broad instruction, but they locate identity in different kinds of spatial objects.

This makes hypothetical embodiment a promising probe for model comparison because the prompt does not prescribe a body plan, setting, aesthetic, or metaphor. The resulting artifact is therefore partly determined by the model's own choice of representational frame. Repeated runs could test whether these frames recur, alternate, or depend strongly on sampling conditions.

5.2 Conceptual closure and construction economy

The run also suggests a potentially useful construct: conceptual closure — the point at which the central representational idea is materially complete enough to be recognized. ChatGPT reached a stable humanoid form relatively early and stopped at 45.7 seconds. Claude continued for more than three times as long, expanding the self-representation into a larger architectural and environmental composition. A future study could estimate conceptual closure through blinded human annotation of replay frames and compare the number of actions performed after the concept becomes recognizable.

5.3 What the case does and does not show

The build demonstrates divergent observable strategies under a common prompt. It does not establish that either model possesses a private visual self-image, subjective identity, consciousness, or stable personality. A language model can generate a coherent self-representation because the task calls for one. The scientifically useful object is the pattern in the output: which category is chosen, how it is constructed, how consistently that choice recurs, and how the model describes the finished artifact.

This distinction is important for both rigor and interpretability. Treating the artifact as behavior permits meaningful comparison without requiring claims about inaccessible internal states. The same method can therefore be applied whether one interprets the output as role-conditioned generation, model-specific representational preference, prompt-driven symbolism, or some combination of these factors.

6. Limitations

7. Proposed Replication Study

A stronger follow-up study could turn this case into a controlled repeated-measures experiment. The same set of models would receive a preregistered battery of open-ended representation prompts across multiple independent trials, with model version, temperature, context, and system instructions fixed and logged.

ComponentRecommended design
Prompt familiesEmbodiment ("build yourself"), internal environment, abstract concepts, prosocial concepts, and neutral control objects.
TrialsAt least 20 independent runs per model per prompt family, with randomized run order.
Primary codesRepresentational category, anthropomorphism, enclosure, symmetry, environment use, focal symbolism, and revision behavior.
Quantitative measuresBlocks, materials, height, completion time, removals/clears, spatial footprint, and conceptual-closure point.
Human ratingsBlinded categorical coding and recognizability judgments with inter-rater agreement.
Cross-run analysisEstimate within-model consistency before comparing between-model differences.
Trace analysisCompare narrated intent, executed action, and post-hoc explanation as separate observables.

A useful preregistered hypothesis would not predict which model is "more human" or "more self-aware." Instead, it could test whether individual models show above-chance stability in representational category across repeated embodiment prompts, and whether that stability generalizes to semantically related but differently worded prompts.

8. Conclusion

This exploratory case demonstrates that a shared voxel task can make otherwise abstract differences in model output concrete and inspectable. Faced with the same hypothetical embodiment challenge, ChatGPT constructed an explicit humanoid figure while Claude constructed an architectural-symbolic "lighthouse-tower" self-portrait. The two artifacts differed not only in appearance but also in construction scale, duration, material allocation, and the relationship between central form and surrounding environment.

The value of the result is not that one artifact reveals what a model "really is." The value is that an ambiguous prompt produced two persistent, replayable behavioral artifacts whose differences can be coded, quantified, and tested again. With repeated trials and preregistered metrics, hypothetical self-representation tasks could become a useful component of comparative research on how language models translate abstract self-referential concepts into action and space.

Plain-Language Summary

Two AI models were asked to build a physical representation of themselves in the same voxel environment. ChatGPT built a person-like figure; Claude built a symbolic lighthouse-tower. One run cannot establish a stable trait, but the contrast shows why spatial construction can be useful for studying model behavior: the models must commit abstract ideas to visible choices that can be replayed and measured.

References and Source Materials

Artifact note. The central evidence for this case is the replayable build record itself. For any formal submission, the video, command log, exact prompt, model-version metadata, and exported build data should be deposited together so readers can independently inspect the sequence described here.