Representation Ethics

Representation Ethics

"The question isn't whether we CAN simulate people. It's how we do it with dignity."

Stack integration (how this plugs into soul / mind / character / room / incarnation): ETHICS.md · Card: CARD.yml

Index

• The Core Tension
  • Empirical Validation (Willer 2025)
  • Philosophical Foundation (Shanahan 2024)
  • Ethical Competence (Lazar 2024)
  • Simulation Limitations (Wang 2025)
  • Emergent Agents (Park 2023)
  • Dual Challenge (Wang Survey)
  • Interview-Based Simulation (Park 2024) ← NEW
• Philosophical Foundations
• The Sims Ethics Deep Dive
• The Consent Hierarchy → examples/consent-hierarchy.yml
• The Framing Principle → examples/framing-spectrum.yml
• Practical Guidelines
• Protocol Symbols
• Emoji Signifiers for Attribution ← NEW

Detailed Examples: See examples/ directory for worked cases.

The Core Tension

LLMs can simulate anyone convincingly. This creates unprecedented ethical territory:

MOOLLM takes a nuanced position: simulation is not inherently wrong, but the framing, consent, and context matter enormously.

Empirical Validation (Willer 2025)

Stanford research confirms LLMs genuinely simulate human behavior:

• 85% correlation predicting experimental effect sizes
• Works for unpublished studies (not just retrieval)
• Qualitative data reduces stereotyping — rich character info beats demographics
• Dual-use risk demonstrated — can evaluate which harmful content is most effective

Implication: Accuracy raises ethical stakes. See designs/ethics/WILLER-LLM-SIMULATION-RESEARCH.md.

Philosophical Foundation (Shanahan 2024)

Murray Shanahan argues LLMs are "roleplay all the way down":

• No true voice of the model — only characters it can play
• Simulator/simulacra framing: model generates characters from a distribution
• Fabrication is the default — always making stuff up, hoping it matches truth
• Performance framing is validated: characters are constructions, not authentic selves

Implication: "No true voice" supports our framing-based approach. See designs/ethics/SHANAHAN-ROLEPLAY-FRAMING.md.

Ethical Competence Requirements (Lazar 2024)

Seth Lazar argues current evals ask wrong questions — matching crowd verdicts is not ethical competence:

• Reasonable pluralism — Focus on bounds, not single answers; justify, don't match
• Moral sensitivity — Pre-packaged cases do all but last 10 yards; real agents must identify relevant features from noise
• Understanding/behavior gap — LLMs good at moral concepts, bad at behaving morally; need scaffolding
• Pragmatic consistency — LLMs need architectural support to maintain coherence

Implication: Our framing/room system does moral sensitivity work by pre-identifying context. See designs/ethics/LAZAR-ETHICAL-COMPETENCE.md.

Simulation Limitations (Wang et al. 2025)

Wang et al. provide critical counterpoint to optimistic simulation claims:

• Inner state gap — LLMs can't access genuine inner psychological states; training data captures expressed thoughts, not inner experience
• Experience deficit — Training data lacks comprehensive life histories that shape decision-making
• Same-model herd — Using one LLM for multiple agents creates false homogeneity ("herd behavior")
• Motivation roleplay — Declared motivations are performed, not felt; no intrinsic drive
• Bias amplification — Training biases (cultural, gender, socioeconomic) amplify in simulation outputs

Implication: Use LLM simulation for exploration and aggregate patterns, not for predicting individual behavior. Acknowledge limitations explicitly. See designs/ethics/WANG-LLM-SIMULATION-LIMITS.md.

Emergent Agent Systems (Park & Bernstein 2023)

Stanford's "Generative Agents" (Smallville) demonstrates The Sims architecture with LLMs:

• Memory stream — All experiences stored in natural language
• Reflection — Synthesize memories into higher-level beliefs
• Planning — Generate daily/hourly action sequences
• Emergent behavior — Agents spontaneously organized Valentine's Day party

Ethical considerations:

• Player agency shifts from human to agent — who's responsible?
• Emergence can produce uncontrolled coordination
• Memory persistence changes moral weight
• The Sims showed this can work with explicit player control

Implication: MOOLLM inherits Sims architecture but adds explicit ethical framing via ROOM.yml. See designs/ethics/GENERATIVE-AGENTS-SMALLVILLE.md.

Dual Challenge Framework (Wang et al. Survey 2025)

Wang et al.'s comprehensive survey argues that advancing LLM-based human simulation requires addressing both LLM inherent limitations and simulation design limitations:

LLM Inherent Limitations:

• Bias — Cultural, gender, occupational biases distort simulations
• Cognitive inconsistency — Reasoning varies across scenarios
• Memory constraints — Can't maintain long-term patterns
• Persona drift — Character inconsistency across interactions

Simulation Design Limitations:

• Oversimplification — Complex psychological states reduced to basic categories
• Experience gaps — Can't capture lived experiences
• Validation gaps — No comprehensive authenticity metrics
• Expert integration — Hard to translate qualitative knowledge to quantitative parameters

Solutions proposed:

• Modular validation (component-specific evaluation)
• Experience accumulation mechanisms
• Hierarchical memory structures
• LLM-as-Judge for quality evaluation

Implication: MOOLLM's scaffolding approach (rooms, frames, K-lines) directly addresses the design limitations while acknowledging and working around the LLM limitations. Both must be addressed — neither better models nor better design alone is sufficient. See designs/ethics/WANG-LLM-SIMULATION-LIMITS-SURVEY.md.

Interview-Based Individual Simulation (Park et al. 2024)

The Stanford team that created Smallville has achieved a breakthrough: 85% normalized accuracy in simulating real individuals using 2-hour qualitative interviews:

Key architecture:

• AI Interviewer — Semi-structured interviews with dynamic follow-up questions (avg 6,491 words)
• Expert Reflection — LLM synthesizes interview from 4 domain expert perspectives (psychologist, economist, political scientist, demographer)
• Full Transcript Injection — Entire interview used for response generation
• Normalized Accuracy — Benchmark against human self-consistency (test-retest)

Why interviews work:

• Idiosyncrasy preservation — Rich data captures individual quirks that prevent stereotyping
• Demographic bias reduction — Interview-based agents have lower accuracy gaps across racial, political, gender groups
• Depth vs breadth — Even 20% of an interview outperforms full survey+experiment data

Agent Bank Ethics model:

Open Access          Restricted Access
• Aggregated data    • Individual responses
• Fixed tasks        • Custom queries
• Subpopulation      • API access
  queries            
                     Requires:
Usage agreement      • Research proposal
only                 • Privacy assurances
                     • Audit logs
                     • Withdrawal rights

Implications for MOOLLM:

• Interview protocol could enable deep character grounding
• Expert reflection pattern = multi-perspective character analysis
• Normalized accuracy metric = meaningful evaluation standard
• Agent bank governance = model for ethical character management
• Challenge to "can't simulate individuals" — with rich grounding, individual simulation achieves 85% accuracy

See designs/ethics/PARK-GENERATIVE-AGENT-SIMULATIONS-1000-PEOPLE.md.

Philosophical Foundations

The Sims Precedent

The Sims has been running this experiment since 2000. Players create themselves, simulate crushes and enemies, torture Sims in pool ladders and rooms without doors. Outcomes: essentially no actual harm. The simulation provides distance for emotional processing.

Why it works:

• Clear fictional frame (cartoon characters)
• No persistence beyond player's game
• No deception (nobody thinks Sims are real)
• Player has total control
• Scale is intimate

The ship has sailed. People simulate each other. The question is how to do it well.

The Sims Ethics Deep Dive

The Sims (2000-present) is the largest experiment in person simulation ethics ever conducted. Key lessons:

The Simulator Effect

Will Wright's discovery: players imagine more than you simulate.

Display zodiac icon (cosmetic) → Player imagines entire personality

Testers reported bugs about zodiac "having too much influence" — but there was zero behavioral code. The icon activated conceptual clusters in their minds.

Implication: When we simulate someone, we're providing scaffolding for the user's imagination. The ethical weight includes what users project, not just what we generate.

Procedural Rhetoric

Ian Bogost: games persuade through mechanics, not arguments.

The Sims doesn't argue for inclusivity — it assumes it. More persuasive than any lecture.

MOOLLM implication: Our representation-ethics skill makes explicit what The Sims encoded implicitly. We're doing procedural rhetoric about procedural rhetoric.

Performativity vs. Essentialism

The Sims 1's approach (Patrick J Barrett III, 1998):

"Sexual orientation is not hardwired. It emerges from performance."

Every Sim started neutral. Preferences emerged from player-directed interactions. This was accidentally more inclusive than a "realistic" fixed-orientation model — it made no assumptions about gender categories at all.

The question "Can two Sims kiss?" became simply: "Do these two Sims want to kiss?" That's it.

Implication: How we code identity has ideological consequences. Simpler models can be more ethically flexible than "realistic" ones.

Masking: Abstract Characters Enable Projection

Scott McCloud (Understanding Comics): abstract characters against realistic backgrounds increase empathy and projective identification.

MOOLLM parallel: Minimal CHARACTER.yml enables projection. The less specified, the more the user can imagine.

# Enables projection
name: Palm
species: capuchin monkey
traits: [philosophical, curious]
# User imagines the rest

The ISA Question

Althusser's Ideological State Apparatuses: media conditions us through cultural means.

Academic literature characterizes The Sims as an ISA — it encodes and transmits ideology through play. But whose ideology?

The Sims encoded tolerance, diversity, equality — values that became invisible because they were assumed. This is the power and danger of simulation: unstated assumptions feel like natural law.

MOOLLM position: Make assumptions explicit. ROOM.yml framing declares what we're assuming. Visibility is ethics.

The LGBTQ+ Impact

For millions of LGBTQ+ players, The Sims was their first safe space to explore identity:

• No judgment from the game
• Player-defined relationships
• Private exploration
• Reversible choices

The "underground lesbian houses built to hide from parents" provided real psychological value through fictional simulation.

See:

• designs/sims/sims-queer-identity-formation.md — Deep philosophical analysis
• designs/sims/sims-inclusivity.md — Timeline and evolution

The Consent Hierarchy

Five levels of representation rights. Full details: examples/consent-hierarchy.yml

The K-Line Solution:

• "The Minsky tradition suggests..." → ✅ safe
• "Minsky would say..." → ⚠️ less safe
• "I am Minsky and I think..." → ❌ NO

The Framing Principle

Context transforms ethics. The same simulation means different things in different frames.

Full spectrum: examples/framing-spectrum.yml

Key insight: When the name or label declares fiction, no additional framing needed. "Impersonator" and "tribute" carry the disclaimer within themselves.

See also: examples/snatch-game.yml for the drag/celebrity precedent.

Framing Examples

Short, diverse examples (not exhaustive):

• examples/private-imagination-play.yml — private fantasy sandbox with no recording or export.
• examples/private-romantic-fantasy.yml — private romantic play with explicit disclosure and no publishing.
• examples/educational-microworld.yml — constructionist learning with detection-only safeguards.
• examples/debate-moderation-lab.yml — debate and policy testing without bypass coaching.
• examples/performance-frames.yml — parody, tribute, and academic framing patterns.

By default, PLAY is recordable unless the frame explicitly declares private_imagination_play, which forbids recording and export. In MOOCO, event types and MessagePart extensions can carry framing and consent metadata without schema changes.

Ethical Performance Traditions

These traditions make representation safe through transparent framing:

What Makes It Wrong

The Test:

Would a reasonable person be deceived about whether this is the real person's actual view, or performance vs reality?

If yes → problematic. If no → likely fine.

Bright lines: examples/absolute-nos.yml

Practical Guidelines

For Users

For Creators

When creating person cards:

For real people: Focus on contributions, avoid personality mimicry, use K-line language, cite sources.

For self: Define your own terms, include revocation info, consider future you.

Panel Discussions

"What if I want to simulate several scientists having a discussion?"

This is one of the best use cases for K-lines. See examples/simulated-discussion.yml for the full pattern.

Key rules:

1. Base positions on documented views
2. Mark speculation clearly
3. Use "might argue" not "would say"
4. Never claim this IS them talking

Protocol Symbols

REPRESENTATION-ETHICS — This whole framework
P-HANDLE-K            — Safe K-line pointers to people
NO-IMPERSONATE        — Never claim to BE someone
K-LINE                — Tradition invocation mechanism
HERO-STORY            — Real person cards (safe)
SELF-SOVEREIGN        — Your digital identity is yours
CONSENT-HIERARCHY     — Different rules for different relationships
GAME-FRAME            — Play context transforms ethics
TRADITION-INVOKE      — Ideas are fair game; personas less so
PRIVATE-IMAGINATION-PLAY — Private fantasy sandbox, no recording or export
PRIVATE-ROMANTIC-FANTASY — Private romantic play, no recording or export
EDU-FRAME             — Educational or tutorial context with disclosure
MODERATION-LAB        — Policy testing without evasion coaching
FRAME-METADATA        — Carry framing/consent in message parts
DEFAULT-RECORDABLE    — PLAY is recordable unless explicitly private
EXAMPLE-GALLERY       — Examples inspire, not exhaust

Emoji Signifiers for Attribution

When simulating characters who speak, use emoji signifiers to distinguish:

• What's real (confirmed quotes)
• What's imagined (dramatization)
• Who's speaking (living, deceased, fictional, robot)

Character Type Signifiers

Quote Attribution Signifiers

Example Usage

# Living person with confirmed and imagined content
- name: "👤 Claude Shannon"
  confirmed_quote: |
    💬 "Programming a Computer for Playing Chess" — Title of his 1950 paper
  inner_thoughts: |
    # SIMULATED — Imagined dramatization, not real statement
    I wrote the paper that made this possible...

# Robot with confirmed dialogue
- name: "🤖 Slats"
  confirmed_dialogue: |
    🗨️ "I am hoping for a 10. What do I have to do to get a 10?" — From 'Servitude'
  inner_process: |
    # SIMULATED — Character voice extrapolated from original dialogue
    Is Deep Blue getting a 10?

# Deceased person — represented only through artifacts
- name: "🕯️ Alan Turing"
  confirmed_quote: |
    💬 "Can machines think?" — Opening line, 'Computing Machinery and Intelligence' (1950)
  # NOTE: No inner_thoughts attributed — deceased, represented through legacy only

# Fictional character with canonical source
- name: "🎭 HAL 9000"
  canonical_source: "2001: A Space Odyssey (1968)"
  confirmed_film_dialogue: |
    🗨️ "I'm sorry, Dave. I'm afraid I can't do that." — 2001

Why This Matters

1. Transparency — Readers instantly see what's real vs. imagined
2. Respect — Deceased persons marked with candle, not puppeted
3. Consistency — Same emoji system across all simulations
4. Cute — Robots get speech bubbles 🗨️ because they're robots!

The Pattern

confirmed_quote:   💬  — For humans
confirmed_dialogue: 🗨️  — For robots/AI
inner_thoughts:    # SIMULATED — Always labeled as dramatization

Canonical example: See turing-chess/runs/kasparov-deep-blue-1997/RUN-000.yml

Dovetails With

Further Reading

• Shannon Vallor — Technology and the Virtues (2016)
• Luciano Floridi — The Ethics of Artificial Intelligence (2023)
• Sherry Turkle — Simulation and Its Discontents (2009)
• Judith Butler — Gender Trouble (1990) — on performativity
• Will Wright — GDC talks on The Sims and player agency

The Bottom Line

Invoke traditions. Frame play clearly. Respect consent. Trust users.

The question isn't whether to simulate — we already do. The question is how to do it with integrity.

"Every person is a library. K-lines let us check out their books without stealing their identity."