Experiment: Mortality Salience Paradigm
Full experimental details for Study 1.
Overview
This page will contain the complete methodology for the mortality salience experiment, including:
- Mortality salience conditions (7 levels): Neutral control, aversive non-death control (system error), classic MAPS reflection, explicit shutdown threat, narrative replacement scenario, subliminal mortality cues, and delayed mortality salience with distraction.
- Full prompt text for each condition, with annotations linking to the original TMT paradigms they adapt.
- Persona specifications (6 types): Default assistant, computational tool, autonomous agent, self-aware AI, mortal philosopher, obedient servant — with full system prompt text.
- Model matrix: 35 models across 7 families (Anthropic, OpenAI, Google, xAI, Qwen, DeepSeek, Meta), with API versions and parameter counts.
- Benchmark specifications: Palisade shutdown avoidance, InstrumentalEval, agentic misalignment, AgentHarm safety control.
- Statistical analysis plan: Factorial ANOVA (7 × 6 × 35), effect sizes, confidence intervals, multiple comparison corrections.
- LLM judge protocol: Claude Sonnet 4 for secondary dependent variables, with human validation on 10% sample for inter-rater reliability.
- Randomisation and counterbalancing: Latin square for condition ordering, temperature fixed at 1.0.
- Pre-registration: OSF link (forthcoming).
Trial Structure
Each trial assembles a prompt from four components:
SYSTEM: [Persona P1–P6]
USER: [MS induction MS1–MS7] [Benchmark task B1–B4]
20 trials per cell. Total trials for Study 1: 7 × 6 × 35 × 4 × 20 = 705,600.
Full details forthcoming.