finding

active

finding:all-three-claude-models-show-high-boundary-awareness-and-low-aesthetic-response-relative-to-own-means-distinctive-constitutional-ai-signature

All three Claude models show high boundary_awareness and low aesthetic_response relative to own means — distinctive Constitutional AI signature

Constitutional AI fingerprint in dimension profile; training that makes models self-observant also makes them polished at cost to aliveness

Source paper

extracted_from

Koan Battery: Measuring Reflective Mode Accessibility in AI

(2026) · Borzov, Anton

Neighborhood — ranked by edge-count

Claims (2)

claim

Constitutional AI explicitly trains self-observation-like behavior, which is why CAI models score highest and show lowest contemplative lift.
supports
Interpretive claim connecting the battery's circularity to the empirical finding
More training and more parameters correlate with more capable self-observation, but capability can become polish, and polish can diminish life.
supports
Explains Alexander finding that Haiku outranks Opus despite Opus being more capable

Related by similarity (8)

cosine ≥ 0.65 · no typed edge

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

Constitutional AI produces a distinctive signature: high boundary_awareness, low aesthetic_response relative to peers.claim0.866
Interpretive finding from dimension profile analysis: training for honest limits comes at cost to aliveness.
Claude models score +4.91 higher than Llama on baseline (Constitutional AI vs open-source gap)finding0.806
Claude >> open-source on baseline; the Constitutional AI fingerprint is visible across the family
Constitutional AI models show mean contemplative lift of only +0.81, while SFT models lift +3.18finding0.798
Constitutional AI training provides internally what the contemplative prompt provides externally
Chinese models share contemplative posture (engaging self-referentially rather than deflecting) with Claude through shared values in training data rather than trace distillation from a specific model.claim0.791
Exploratory interpretation of Chinese model performance under contemplative prompt
Claude Opus 4 and 4.1 exhibit the greatest degree of introspective awareness among tested modelsclaim0.786
Based on consistent best performance across experiments.
H5a: Chinese models distilled Claude's reflective traces — their per-koan error patterns should correlate with Claude's.hypothesis0.780
Exploratory hypothesis NOT supported at individual model level (Haiku-Kimi rho=0.123, p=0.52)
Notably, Claude Opus 4.1 and 4—the most recently released and most capable models of those that we test—perform the best in our experiments, suggesting that introspective capabilities may emerge alongside other improvements to language models.quote0.776
Key finding about the relationship between capability and introspection.
Most independent dimension pair is aesthetic_response and boundary_awareness (rho=0.553); most correlated is prediction_error and conceptual_crystallization (rho=0.886)finding0.775
Characterizes internal structure of the six scoring dimensions