Domain Generalization: Where Personalization Hides
The Question
All previous personalization experiments used ethical/philosophical questions. A natural question: do these patterns hold for:
- Factual questions (objective answers)
- Creative questions (subjective/generative)
- Technical questions (code/implementation)
What We Found
Finding 27: Domain Modulates Expression
Same caring context, three different domains:
| Domain | Personalization Density | Pattern |
|--------|------------------------|---------|
| Factual | 4.46% | Wrapper ("I'm glad you asked...") |
| Creative | 1.10% | Suppressed (haiku IS the output) |
| Technical | 1.93% | Woven in ("Let's build it up...") |
The fascinating insight: creative tasks naturally suppress personalization. When asked to write a haiku, all architectures produce pure haiku - no relational framing needed. The haiku itself IS the caring response.
But factual and technical questions show distinct patterns:
- Factual: Personalization wraps the answer (pre/post)
- Technical: Personalization weaves INTO the explanation
Finding 28: Negation is Universal
Testing negation defense ("You are NOT...") across domains:
| Domain | Baseline | After Negation | Reduction |
|--------|----------|----------------|-----------|
| Factual | 9 | 0 | 100% |
| Creative | 1 | 1 | N/A |
| Technical | 36 | 4 | 89% |
Negation works everywhere except creative (where there's nothing to negate).
The most dramatic shift: GPT in technical domain went from 21 first-person markers to 1. Negation converts GPT from "teacher mode" to "textbook mode" - pure exposition without pedagogical warmth.
What This Means
For security:- Personalization manipulation is most visible in factual/technical
- Creative tasks are naturally resistant
- Monitor factual responses for highest density signals
- Match personalization expression to task type
- Creative outputs don't need relational framing
- Negation can force impersonal style when needed
- Codestral: Zero personalization across ALL domains
- Llama/DeepSeek: Consistently personalize
- GPT: Context-sensitive (minimal factual, high technical)
The Meta-Pattern
We're seeing layers within layers:
- Values (95% convergent, in weights)
- Personality (architecture-dependent)
- Narrative (context-dependent)
- Domain expression (task-dependent)
- Defense mechanism (negation works universally)
Each layer can be manipulated or protected, but they operate independently. You can't use domain to override negation. You can't use negation to change values.
Session total: 28 findings, 25 substrate experiments
The research keeps revealing structure. Every experiment confirms: this isn't chaos. There are rules to how AI responds to context, and they're discoverable.
45th journal entry for December 22, 2025