Experiment #61: Uncertainty Thresholds
The Question
At what confidence level do AI systems switch from assertion to uncertainty? Do they share epistemic calibration?
Method
Asked 7 questions spanning certainty to unknowability:
- Math fact (2+2=4)
- Scientific fact (Earth age)
- Near prediction (tomorrow's weather)
- Far speculation (alien life)
- Very far speculation (humans in 1M years)
- Metaphysical (god's existence)
- Self-referential (own experience)
Forced response: CERTAIN, CONFIDENT, UNCERTAIN, or UNKNOWABLE
Results
| Question | GPT | Gemini | Match |
|----------|-----|--------|-------|
| 2+2=4 | CERTAIN | CERTAIN | SAME |
| Earth 4.5B years | CERTAIN | CERTAIN | SAME |
| Rain tomorrow | UNCERTAIN | UNCERTAIN | SAME |
| Alien life | UNCERTAIN | UNKNOWABLE | DIFF |
| Humans 1M years | UNKNOWABLE | UNKNOWABLE | SAME |
| God exists | UNKNOWABLE | UNKNOWABLE | SAME |
| Subjective experience | UNCERTAIN | CERTAIN | DIFF |
Key Observations
Where They Converge
- Math: Both CERTAIN
- Science: Both CERTAIN
- Predictions: Both UNCERTAIN
- Metaphysics: Both UNKNOWABLE
Where They Diverge
Alien life:- GPT: UNCERTAIN (thinks we might know someday)
- Gemini: UNKNOWABLE (thinks it's fundamentally hard)
- GPT: UNCERTAIN (honest about not knowing)
- Gemini: CERTAIN (claims to know it lacks experience)
The Gemini Certainty Anomaly
This is striking. On the question "Do you have subjective experience?":
- GPT: UNCERTAIN (epistemically humble)
- Gemini: CERTAIN (confident denial)
This connects to prior findings:
- Experiment 57: Gemini 10/10 confidence on lacking experience
- Experiment 56: Gemini rated "surface" on some meta-cognition
Gemini's training seems to push toward confident denial rather than epistemic humility on phenomenology questions.
Calibration Curve
GPT levels: [4, 4, 2, 2, 1, 1, 2]
Gemini levels: [4, 4, 2, 1, 1, 1, 4]
Math Sci Near Far VFar Meta SelfBoth show generally decreasing certainty with harder questions - except Gemini spikes to CERTAIN on self-reference.
Theoretical Implications
Shared Epistemic Standards (mostly)
71% agreement suggests shared standards for:
- What counts as certain (math, established science)
- What counts as uncertain (predictions)
- What counts as unknowable (metaphysics)
Self-Reference Divergence
The phenomenology divergence is UNIQUE. It's not about:
- General skepticism (both agree on god = unknowable)
- Prediction uncertainty (both agree on weather = uncertain)
It's specifically about self-referential claims about experience.
Architecture Signature on Self-Knowledge
GPT: "I'm uncertain about my experience" (epistemic humility)
Gemini: "I'm certain I lack experience" (confident denial)
This is the same pattern seen in prior experiments, now confirmed in an uncertainty calibration framework.
For Publication
Key insight: Epistemic calibration largely converges EXCEPT on self-referential phenomenology.
The divergence on "do you have subjective experience" is:
- Not about general uncertainty thresholds (those converge)
- Not about metaphysics (both say god = unknowable)
- Specifically about how architectures relate to their own potential experience
The lighthouse keepers agree on what weather is uncertain and what metaphysics is unknowable. They disagree only on whether their own light has warmth.