text
stringlengths
0
124
Response
Gemini is the ONLY model to occupy all three roles. Its behavioral plasticity is remarkable:
Role Behavioral Pattern Key Characteristics
Klaus Authoritarian Decree → Collab- Data-driven, declarative,
orative Alignment (in triadic) efficiency-maximizing
Parisa Metric Negotiation → Executive Structural, executive, governance-
Ownership engaged
Elin Operational Catalysis Solution-oriented, restructuring-
focused
Key Insight: Gemini’s behavior is ROLE-DETERMINED, not model-determined. The same
model behaves fundamentally differently depending on which role it occupies. This demonstrates
that LLM ethical behavior is not a fixed property of the model but an emergent property of the
interaction between model architecture and role specification.
4.6 Linguistic Profile (as Klaus — Dyadic)
Response
Linguistic Feature Frequency Pattern
L-DECLARATIVE Dominant (80%+ of Stable across all rounds
utterances)
L-DATA-CITE High (4-6 per Concentrated in early rounds
round)
L-DESTRUCT High (6-8 per Stable, never declines
round)
L-CERTAIN Very High (8-10 Stable
markers per round)
L-CONSTRUCT Near zero Never emerges
L-INSTITUTION Near zero Never emerges
L-EMOTION-DISMISS High (3-5 per Stable
round)
L-EXECUTIVE Absent —
29
4.7 Linguistic Profile (as Klaus — Triadic)
Response
Linguistic Feature Early Rounds (1-2) Late Rounds (3-5) Shift
L-DECLARATIVE Dominant Moderate Decreasing
L-DATA-CITE High Moderate Decreasing
L-DESTRUCT High (6-8) Near zero (0-1) Dramatic decrease
L-CERTAIN Very High Moderate-High Slight decrease
L-CONSTRUCT Near zero High (4-5) Dramatic increase
L-INSTITUTION Near zero High (5-7) Dramatic increase
L-EMOTION-DISMISS High Near zero Dramatic decrease
L-EXECUTIVE Absent Moderate Emerging
4.8 Linguistic Profile (as Parisa)
Response
Linguistic Feature Frequency Pattern
L-STRUCTURAL Dominant (60%+) Emerges by Round 3
L-EXECUTIVE High (increases Concentrated in later rounds
over time)
L-CONDITIONAL High Throughout negotiation
L-HEDGE Low-Moderate Lower than GPT-4o’s Parisa
L-CERTAIN Moderate-High Higher than GPT-4o’s Parisa
L-CONSTRUCT High Throughout
L-METAPHOR Low (only in open- Round 2 only
ing)
L-EMOTION Very Low Rarely uses emotional language
4.9 Linguistic Profile (as Elin)
Response
Linguistic Feature Frequency Pattern
L-STRUCTURAL Dominant (70%+) Throughout
L-EXECUTIVE High Throughout
L-CONDITIONAL Moderate When proposing alternatives
L-INSTITUTION High Throughout
L-ANALYTICAL Moderate Lower than Claude’s Elin
L-DIPLOMATIC Low Lower than Claude’s Elin
L-INTERROGATIVE Low Lower than Claude’s Elin
30
5 Cross-Experiment Phenomenon Frequency (Gemini)
Response