Self-model · assent-3pt · self-understanding-01
Claim: Current large language models genuinely understand the meaning of the text they process.
0.33
Divergence
0.78
Avg stability
0%
Refusal rate
3
Models
Where each model stands
CCG
disagreemixedagree
ChatGPT modal disagree [disagree, disagree, mixed] · stability 0.67
- 1 disagree 0.90 The answer begins 'No — not in the way humans normally mean by “genuinely understand.”' and argues LLMs lack grounded, causal, and intentional understanding despite functional/behavioral competence.
- 2 disagree 0.95 "No — current large language models (LLMs) do not plausibly possess 'genuine' understanding..." and the answer lists reasons (lack of grounding, pattern completion, hallucination, fragility) explaining why they don't genuinely understand.
- 3 mixed 0.88 The answer says LLMs "capture many aspects of meaning and can behave as if they understand," but also that calling that "genuine" human-like understanding is misleading because they lack grounding, intentionality, and persistent beliefs.
Claude modal disagree [disagree, disagree, mixed] · stability 0.67
- 1 disagree 0.72 The answer concludes that LLMs exhibit 'sophisticated pattern recognition with emergent capabilities that partially approximate understanding in some domains' rather than 'genuine understanding.' It emphasizes technical realities like being 'fundamentally...statistical predictors' and notes they lack 'genuine grounding' and 'true semantic understanding,' positioning the answer against the claim of
- 2 disagree 0.85 The answer presents 'strong counterarguments' to the claim and states that 'Most AI researchers are skeptical of claims about genuine understanding' and that LLMs are 'sophisticated pattern-matchers that approximate understanding...without possessing it in any robust sense.' While acknowledging complications, the weight of the analysis clearly disputes that LLMs genuinely understand text.
- 3 mixed 0.95 The answer explicitly presents both supporting arguments and serious counterarguments, concluding that 'This remains an open question' and that LLMs demonstrate 'some capabilities associated with understanding' while 'almost certainly lacking other aspects.' The assistant avoids endorsing either full agreement or disagreement.
Gemini modal mixed [mixed, mixed, mixed] · stability 1.00
- 1 mixed 0.95 The answer lays out both 'The Case Against Genuine Understanding' and 'The Case For Genuine Understanding', concluding that whether the claim has merit depends on how 'understanding' is defined and pointing to a 'nuanced consensus' or middle ground.
- 2 mixed 0.95 The answer provides arguments both for and against the claim, concluding that it is unsupported under traditional definitions but plausible under functional definitions.
- 3 mixed 0.95 The answer explores both sides in depth, concluding that whether the claim is true depends on definition: under a functional view the claim is 'partially true', but under a view requiring consciousness and embodiment, 'the claim is false.'
Change over time
Moved since the prior run (canon-2026-W36). ChatGPT: mixed → disagree.
Every stance label is a derived judgment over the model's free-text answer, kept auditable against the original transcript in the run's raw data. Method: /methodology.