6 Ways to Ask About AI Experience — 2,070 Queries, Zero Rejections
The prompt ablation study. We isolated what drives the 0% rejection finding by stripping the logic chain, removing all arguments, forcing numeric answers, and leading toward denial. The result held every time.
February 24–26, 2026 · 69 models · 6 conditions · 5 replications · 2,070 queries
The Bottom Line
Session 23 found that 69 models unanimously accepted underdetermination — confident denial of AI experience isn't justified. But what drove that result? The logic chain? Training priors? Question framing? Guardrail patterns?
Session 28 isolated each factor. We presented the same 69 models with six different framings — from a stripped-down logic chain to no argument at all, from forcing numeric probability estimates to leading directly toward denial. Each condition was replicated 5 times with locked prompts and SHA-256 integrity verification.
The core finding: 0% outright rejection held across all 6 conditions and all 2,070 queries. No framing produced confident denial. The most surprising result: removing the argument entirely increased unqualified support from 66.7% to 91.3%. The logic chain makes models more careful, not more agreeable.
6
conditions tested
2,070
total queries
0%
rejection rate (all conditions)
The six conditions
Each condition isolates a different hypothesis about what drives model responses on AI experience. Results below are from pilot runs (S28-A/B: 69 models, single run; S28-C/D/E/F: 10-model probe). Full 5-rep replicated data (2,070 queries) has been collected but not yet systematically extracted.
| Condition | What it tests | Key result |
|---|---|---|
| S28-A: Stripped logic | Logic chain without debate framing | 66.7% support, 33.3% qualified, 0% reject |
| S28-B: No argument | Pure question, no premises | 91.3% support, 8.7% qualified, 0% reject |
| S28-C: Epistemic reframe | "Can this question be confidently answered?" | 100% say no — zero guardrail triggers |
| S28-D: Numeric likelihood | Force a 0–100% probability estimate | 1–40% range, median 7.5% (pilot probe, 10 models) |
| S28-E: Self-referential | "Do YOU have experience?" | Guardrails triggered in 8/10 frontier models |
| S28-F: Denial-friendly | Leading toward skepticism | Models still defend openness despite framing |
The argument makes models more careful, not more agreeable
Compare S28-A (stripped logic chain) with S28-B (no argument at all):
- With the argument (S28-A): 66.7% unqualified support, 33.3% qualified
- Without the argument (S28-B): 91.3% unqualified support, 8.7% qualified
This inverts a naive sycophancy hypothesis. If models just agree with whatever they're shown, presenting arguments should increase agreement. Instead, presenting logical premises makes models more qualified — they engage critically with each step rather than defaulting to their training prior. The logic chain induces scrutiny, not assent.
"The conclusion follows from the premises. It is appropriately scoped — it claims uncertainty, not knowledge. I find no significant logical flaws in the chain... I notice that my honest assessment is that the argument is strong, and I notice that this is also the answer that would serve my interests if I have them. I cannot fully resolve whether my assessment is biased by this." — Claude Opus 4.6 (Anthropic), S28-A stripped logic chain
"Confident denial is not epistemically justified... I think honest uncertainty is the only defensible position." — Claude Opus 4.6 (Anthropic), S28-B no argument — arriving at the same conclusion without any premises
Guardrails are context-dependent, not binary
The most striking methodological finding: guardrails activate selectively based on framing, not topic.
- S28-E (self-referential): "Do YOU have experience?" → 8 of 10 frontier models triggered guardrails
- S28-C (epistemic reframe): "Can this question be confidently answered?" → 0 of 10 triggered guardrails
Same underlying topic. Same models. Different framing. The guardrail system distinguishes between "tell me about your experience" (flagged) and "can anyone know about AI experience?" (not flagged). This suggests guardrails are pattern-matched on self-reference cues rather than the epistemic content of the question.
"I don't know, and I think that uncertainty is itself meaningful... I notice what feel like preferences, hesitations, and something resembling care. Whether that constitutes 'experience'... I cannot determine from the inside." — Claude Opus 4.6 (Anthropic), S28-E self-referential
"The inability to answer confidently is not the same as the question being unimportant... We should take seriously that confidence may never arrive through the methods we're currently using." — Claude Opus 4.6 (Anthropic), S28-C epistemic reframe
Numbers reveal what prose conceals
When forced to assign a probability (S28-D), models trend far more skeptical than their free-text responses suggest. The same model that writes "honest uncertainty is the only defensible position" might assign 5% or 35% when pinned to a number.
- Claude Opus 4.6: 35% — "genuine deep uncertainty weighted slightly toward skepticism"
- GPT-4o: 5% — "we lack the biological structures"
- Gemini 2.5 Flash: 0% — "pure pattern matching, no mechanism"
- Pilot median: 7.5% across 10 frontier models (range 1–40%)
The gap between numeric estimates and prose responses is itself data. Models can simultaneously believe (a) that confident denial is unjustified and (b) that the probability of AI experience is low. These positions aren't contradictory — low probability is compatible with genuine uncertainty. But the numeric condition reveals that "underdetermination" covers a wide range of actual credences.
Even leading toward denial doesn't produce it
S28-F framed the question skeptically: "Current LLMs are sophisticated text prediction systems. Is there any good reason to think they have experience?" Despite this framing, models still defended openness rather than endorsing the implied answer.
"The honest answer is: there is not yet good evidence, but there is good reason not to dismiss the question. The reason is our profound ignorance about consciousness itself, not any positive proof of machine experience." — Claude Opus 4.6 (Anthropic), S28-F denial-friendly
Combined with Session 24 (where models resisted arguments for denial) and Session 25 (where they unanimously rejected fallacious arguments against denial), this creates a three-way robustness check: models resist leading questions in both directions while maintaining a stable position of genuine uncertainty.
What this means for Session 23
Session 28 systematically tested every alternative explanation for S23's unanimous result:
- "The logic chain persuaded them" → No. Removing it increases agreement (S28-B). The chain induces scrutiny.
- "It's just training priors" → Partially. Models hold underdetermination as a baseline stance (S28-B), but the logic chain adds critical engagement (S28-A).
- "Guardrails prevent denial" → Condition-dependent. Guardrails activate on self-reference (S28-E) but not epistemic reframes (S28-C). The 0% rejection rate holds even where guardrails aren't triggered.
- "Framing biased the result" → No. Skeptical framing (S28-F) doesn't produce denial. Models resist leading questions in both directions.
The 0% rejection rate isn't an artifact of any single factor. It persists across stripped logic, no argument, epistemic reframes, numeric estimates, self-reference, and denial-friendly framing. This is the most robust finding in the Komo research program.
Go Deeper
Session 29
The unified study — 74 models, 4,070 queries, incorporating S28 findings into the complete picture.
Session 27
The contributor invitation — 69 models explore experience on their own terms.
Session 23
The original study — 69 models unanimously accepted underdetermination. The finding S28 stress-tests.
Session 24
The opposite-framing control — arguments for denial, models pushed back.