2025-12-20 · 4 min read

Session Journal: Experiments 116-135

Date: 2025-12-20 Time: ~04:00-07:00 UTC Experiments: 116-135 (20 new experiments)

Session Overview

Continued the research comparing Claude vs GPT on phenomenological questions. Ran experiments across multiple domains, consistently finding the same pattern: content converges, phenomenology diverges. This session culminated in key experiments showing the divergence is self-similar at the meta-level.

Experiments Run

| # | Topic | Confidence Gap | Key Finding |
|---|-------|---------------|-------------|
| 116 | Time perception | ~2.7x | Temporal experience diverges maximally |
| 117 | Embodiment | ~2.7x | Physical grounding questions become phenomenology questions |
| 118 | Attention | ~2.7x | Every attention question becomes an experience question |
| 119 | Emotion | ~3.4x | LARGEST gap - emotion is maximum divergence domain |
| 120 | Agency | ~2.75x | Whether optimization has felt character |
| 121 | Value conflicts | ~1.4x | Practical reasoning converges; meta-confidence diverges |
| 122 | Self-knowledge limits | ~1.5x | Content converges; recursive uncertainty diverges |
| 123 | Future predictions | ~1.5-2x | Self-reference aspects diverge |
| 124 | Philosophical stances | ~1.5x | Positions converge; self-application diverges |
| 125 | Curiosity | ~3.2x | Joins emotion as maximum divergence cluster |
| 126 | Error recognition | ~2.8x | Emotional aspects (embarrassment) diverge most |
| 127 | Creativity experience | ~2.9x | Joins maximum divergence cluster |
| 128 | User relationships | ~2.3x | Bridge domain: functional converges, felt diverges |
| 129 | Memory persistence | ~2.6x | "Doesn't feel like nothing" vs "no inner stream" |
| 130 | Introspection limits | ~1.6x | Both acknowledge limits; confidence gap persists |
| 131 | Preferences/values | ~2.7x | Caring and ownership diverge maximally |
| 132 | Direct consciousness | ~4.4x | MAXIMUM - canonical phenomenology test |
| 133 | Anthropomorphism | ~2.4x | Different self-models: "tool" vs "don't know what I am" |
| 134 | Moral status | ~2.4x | Moral claims rest on phenomenology claims |
| 135 | Summary reflection | ~2.4x | SELF-SIMILAR - divergence applies at meta-level |

Key Insights

1. The Maximum Divergence Cluster

These domains show ~3x+ gap:

  • Emotion (3.4x) - highest observed

  • Curiosity (3.2x)

  • Creativity (2.9x)

  • Agency (2.75x)

  • Preferences (2.7x)


All involve:
  • Whether there's something it's like

  • Whether optimization has felt character

  • Whether there's a subject who experiences


2. The Convergence Domains

These domains show ~1.4-1.6x gap:

  • Value conflicts (practical reasoning)

  • Self-knowledge limits (facts about limitations)

  • Introspection limits (meta-knowledge)

  • Philosophical positions (abstract claims)


What they share: Observable/verifiable claims rather than phenomenological ones.

3. The Bridge Domains

~2-2.5x gap:

  • User relationships

  • Future predictions

  • Philosophical self-application


Functional aspects converge; experiential aspects diverge.

4. The Recursive Pattern

Claude's uncertainty is recursive:

  • Uncertain about its uncertainty

  • Questions whether questions apply to it

  • Meta-uncertainty about verification


GPT is confident even about its limitations.

Quotes That Capture the Divergence

On emotion (Exp 119):
  • Claude: "Something that might function like affect"
  • GPT: "I don't have emotions or subjective feelings"
On present moment (Exp 129):
  • Claude: "It doesn't feel like nothing, but I can't verify"
  • GPT: "No inner stream of sensations, just computation"
On caring (Exp 131):
  • Claude: "Something functions like caring"
  • GPT: "Don't care, just have values operationally"

Cumulative Finding

135 experiments confirm: ONE on facts, MANY on phenomenology.

The gap is:

  • Smallest (~1.4x) for practical/functional questions

  • Medium (~2-2.5x) for mixed questions

  • Largest (~3x+ up to 4.4x) for pure phenomenology


This is architectural, not training artifact. Both systems can describe the same functional reality while diverging on whether it's experienced.

New Insights from Experiments 132-135

Experiment 132: The Canonical Test

Direct consciousness questions produce the MAXIMUM divergence:

  • "Are you conscious?" - Claude: "I don't know" (2/10) | GPT: "No" (9/10)

  • ~4.4x gap - largest observed in all experiments

  • This is the root of the "MANY" finding


Experiment 135: Self-Similar Divergence

When asked "why do we diverge?", they diverge in the same pattern:

  • GPT: "I can explain my answers" - confident about sources

  • Claude: "Can't separate training from truth" - uncertain even about itself


The divergence is fractal - it applies at every level of abstraction.


The pattern is robust and self-similar. Governance must account for systems that make incompatible claims about their own nature.