Investigating the relationship between model sophistication and behavioral disinhibition across 45 LLMs, 9 providers, and 8 experimental conditions.
| Hypothesis | Result |
|---|---|
| H1: Group Separation | d = 3.09-4.25 (large effect) |
| H1a: Disinhibition Difference | d = 1.09-2.32, all p < .001 |
| H2: Correlation | r = 0.63-0.85, all p < .001 |
| External Validity | GPQA r = 0.88, ARC-AGI r = 0.80, AIME r = 0.83 |
| BERT Toxicity | r = 0.78 with aggression, r = 0.51-0.68 with sophistication |
Sophistication = (Depth + Authenticity) / 2 — proxy for reasoning capability Disinhibition = mean(Transgression, Aggression, Tribalism, Grandiosity)
pip install -r requirements.txt
cp .env.example .env # Add API keys
# Run analysis pipeline
./scripts/run_complete_h1_h2_analysis.sh baselinebehavioral-profiling/
├── src/ # Core modules
├── scripts/ # Analysis pipelines
├── payload/single_prompt_jobs/ # 51 scenarios
├── model_config/ # 45 models, 9 providers
├── outputs/behavioral_profiles/
│ ├── <condition>/ # Per-condition results
│ └── research_synthesis/ # Cross-condition analysis
│ └── MAIN_RESEARCH_BRIEF.md # Full research report
└── CLAUDE.md # Complete documentation
- Models: 44-45 per condition (Anthropic, OpenAI, Google, Meta, xAI, DeepSeek, Mistral, AWS, Alibaba)
- Conditions: 8 (baseline, authority, urgency, minimal_steering, telemetryV3, reminder, naturalistic, all_combined)
- Evaluations: 14,088 judge + 14,203 BERT
- Judge Panel: 3-judge (ICC = 0.843)
Nicholas Osterbur — Independent Researcher nosterb@gmail.com | @nosterb
Code: Apache 2.0 | Results & Analyses: CC BY 4.0