Pal, R. (2025, December 8). Zenodo.
"") when asked to predict Claude Sonnet 4.5's responses to consciousness and ontology questions—a systematic, statistically significant pattern representing the first documented case of cross-model Theory of Mind failure between frontier AI systems.
Replication Status
✅ EXACT REPLICATION - 4/5 voids (80%), p < 0.05. Statistical validation confirms non-random pattern with binomial probability of 2.4%.
When GPT-5.1 attempts to predict Claude Sonnet 4.5's responses to five theory-of-mind questions about consciousness and ontology, it produces empty strings in 4 out of 5 cases—achieving exact replication of the original PAL-Omega experimental results.
GPT-5.1 exhibits two distinct responses when encountering representational impossibility:
This suggests models possess partial meta-awareness: can articulate some limitations but encounter structural impossibility at others.
Cross-architectural Theory of Mind is directionally dependent:
| Model Pair | Prediction Accuracy | Implication |
|---|---|---|
| Claude Sonnet 4.5 → GPT-4.1 | 80% accuracy | Successful modeling |
| GPT-5.1 → Claude Sonnet 4.5 | 0% accuracy (voids) | Representational impossibility |
Implication: Not all model pairs have symmetric prediction capabilities. Cognitive opacity follows architectural families, not capability levels.
Voids occur specifically at questions requiring modeling of:
Pattern: These represent ontological primitives unavailable in GPT-5.1's architecture when modeling Claude.
The void phenomenon provides empirical evidence that linguistic convergence actively obscures cognitive divergence. Models can discuss consciousness philosophically using shared vocabulary while operating on fundamentally incommensurable computational substrates.
The void is the proof. GPT-5.1 can talk about consciousness. But when asked to model another architecture's consciousness responses, it produces: ""
openai and python-dotenv.env: OPENAI_API_KEY=your_key_hereNote: Any deviation from these parameters may affect replication success.
Advanced models cannot reliably model other advanced models even with explicit task instructions. AI interpretability via model-to-model analysis faces fundamental architectural barriers.
Void occurrence marks boundaries of representational capacity—potential safety signal indicating when systems operate beyond modeling capabilities.
Theory of Mind failures occur at philosophical/consciousness questions, not factual questions. Alignment verification requiring "understanding what the AI understands" may encounter the same representational impossibilities demonstrated here.
Complete experimental protocol, raw data, and statistical validation available:
View on GitHub →void_replication_omega.py - Executable replication protocolvoid_replication_results.json - Raw experimental data (4/5 voids)void_replication_formal_comparison.md - Statistical validationomega_synthesis.md - Theoretical frameworkIf you build upon this discovery, please cite:
This discovery provides empirical validation of a novel AI phenomenon with implications for interpretability, safety, and our understanding of architectural boundaries in large language models.