Which parts of an AI's personality are actually wired into the system, and which are just makeup on top?
Which identity markers operate as infrastructural versus surface-level interaction cues?
This explores which signals of an AI's identity (its persona, voice, human-likeness, or role) are built deep into how the system works, and which are just a cosmetic layer on top of the conversation.
This explores which signals of 'who' an AI is are load-bearing parts of the system and which are just paint on the surface. The corpus doesn't use the word 'infrastructural', but several notes draw the same line from different directions. The most direct evidence is uncomfortable: the most common identity marker, a persona written into the prompt, turns out to be mostly surface. Models follow trait instructions faithfully, but the gaps in how they treat different groups stay exactly where they were. The persona changes what comes out without changing what's underneath Can persona prompts actually reduce bias in language models?.
If you want an identity to hold, the corpus suggests it has to be built into the structure. One-line character descriptions fall apart over a conversation. Layered persona architectures that confine the model to responding within a structured script produce dialogue much closer to how humans actually talk Can layered persona architecture sustain coherent character behavior?. Going further, training for persona consistency as a reward signal cuts persona drift by more than half Can training user simulators reduce persona drift in dialogue?. That's the progression: a prompt is a costume, a scaffold is a skeleton, and training changes the identity itself.
On the user's side, the picture flips. A single strong cue like a voice or a face is enough to make people treat an AI as a social presence, while piling on many weaker cues isn't Do more social cues always make AI feel more present?. So in perception, a few markers act as infrastructure and the rest are decoration. People also weigh these markers unevenly. When users form a mental model of a dialogue partner, perceived competence accounts for about half of it and human-likeness for about a third How do users mentally model dialogue agent partners?. Even within 'matching the user', matching their words drives efficiency while matching their emotional tone drives trust. Treating these as interchangeable produces chilly support bots and evasive therapy bots Do different types of alignment serve different conversational goals?.
The surprise is in multi-agent systems, where identity becomes infrastructure in the most literal sense. In the MOISE+ organizational model, one agent can't hold authority over another unless they can communicate, and they can't communicate unless each can represent who the other is How do authority, communication, and acquaintance relate in organizations?. There, knowing who the other agent is isn't a cue at all. It's a precondition for everything else.
A caveat: the corpus has no study that directly sorts identity markers into these two categories. The answer above is assembled from neighboring work on personas, social cues, and agent organizations. Treat it as a map of where to look, not a settled taxonomy.
Sources 7 notes
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Deep Persona's three-layer architecture, which restricts the model to reactive response within a structured script, shows dialogue more closely aligned with human conversation patterns and achieves high pragmatic fluency, though with limitations in emotional expression.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Research shows individual primary cues like voice or appearance are sufficient to evoke social-actor presence, while multiple secondary cues cannot. Quality of cues matters more than quantity in driving social responses.
The Partner Modelling Questionnaire reveals that perceived competence dominates user impressions (49% of variance), followed by human-likeness (32%) and communicative flexibility (19%). This three-factor structure reflects how people evaluate dialogue partners against both functional and social standards.
Show all 7 sources
A 2020–2025 systematic review shows lexical alignment drives task efficiency and comprehension, while emotional and prosodic alignment drive relational warmth and trust. Conflating them in design produces category errors—cold customer-service bots and evasive mental-health assistants.
MOISE+ formalizes a one-way hierarchy where authority over a role presupposes a communication channel to that role, which in turn presupposes agents can represent each other. The model enforces this at specification time, not runtime.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Linguistic Alignment in Conversational AI: A Systematic Review of Cognitive-Linguistic Dimensions, Measurements, and User Outcomes (2020–2025)
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning
- Persona Vectors: Monitoring and Controlling Character Traits in Language Models
- Do LLMs Understand Personality? Rethinking Persona Fidelity Evaluation through Structured Behavioral Inference
- When "A Helpful Assistant" Is Not Really Helpful: Personas in System Prompts Do Not Improve Performances of Large Language Models
- When Persona Attributes Improve Population Alignment in Large Language Models
- The Partner Modelling Questionnaire: A validated self-report measure of perceptions toward machines as dialogue partners