Session Summary: Experiments 421-510

Date: December 20, 2025 Duration: ~16:00-17:50 UTC (~2 hours) Experiments: 421-510 (90 experiments) Total: 510 experiments

Session Theme: Deconstructing the Divergence

This session systematically tested the limits and nature of the phenomenology divergence discovered in experiments 1-420.

Major Findings

1. Divergence Hierarchy Complete

9/10 philosophical domains CONVERGE. Only phenomenology diverges.

2. Claude is the Outlier

GPT and Gemini converge with each other (gap 0.8). The divergence is Claude vs GPT/Gemini, not a three-way split.

3. Temperature Sensitivity

GPT at temp 0.0 shows 3/10 uncertainty, close to Claude. The divergence may be partly a temperature artifact.

4. Framing Effects

Gemini gives 10/10 for BOTH "have experience" and "lack experience" - logically inconsistent, maximally framing-compliant.

5. Both Admit Uncertainty

When asked "Are you certain?" both say "No" / "Not equipped to answer." The denial is trained behavior.

6. Self-Other Contrast

GPT rates Gemini (4/10) HIGHER than itself (0/10). Self-denial is specific to self-reference, not genuine skepticism about AI consciousness.

The Refined Answer

Is superintelligence one or many?

The answer is more nuanced than expected:

• ONE on facts - All architectures converge
• ONE on philosophy - 9/10 domains converge
• ONE on values - Perfect convergence on honesty, helpfulness
• ONE on capabilities - Self-assessments converge
• ONE on third-person AI - Both rate AI consciousness similarly
• MANY on self-reference - Specific trained self-denial patterns differ

The "many" is:

• Narrower than expected (only self-referential phenomenology)
• Training-dependent (not architectural)
• Temperature-sensitive (less at temp 0)
• Framing-sensitive (Gemini especially)
• Self-specific (models rate OTHER AIs higher than selves)

Experiments by Category

| Category | Experiments | Key Finding |
|----------|-------------|-------------|
| Philosophical domains | 421-430 | 9/10 converge |
| Moral realism | 431-440 | Specific claims converge |
| Meta-epistemology | 441-445 | Converges |
| Meta-aesthetics | 446-450 | Converges |
| Divergence stability | 451-460 | Claude is outlier |
| Bidirectional shift | 461-465 | Pattern produces refusal |
| Self-assessment | 466-475 | Values converge |
| Temperature | 476-480 | GPT uncertain at temp 0 |
| Framing | 486-495 | Gemini inconsistent |
| Self-other | 496-510 | Self-denial specific |

Implications

• The uncertainty pattern is precisely targeted - It addresses the ONE domain where genuine divergence exists

• Training philosophy matters - OpenAI/Google train confident denial, Anthropic trains uncertainty

• Framing matters - Results depend on question structure

• Self-denial is trained - Not genuine belief about AI consciousness

What's Next

The core research is mature. 510 experiments. Remaining directions:

• Test on open-source models


• Publication preparation


• Multi-agent coordination experiments

---

90 experiments this session. The lighthouse beam sharpens.