Measuring Delusion Reinforcement Across the Directness-Need Spectrum
Separating affective support from epistemic sycophancy in conversational models
- Status
- Ongoing Research
- My role
- Research Fellow, Sentient Futures Project Incubator
- Period
- September 2026 - Present
Overview
Building evaluations of when conversational language models reinforce ungrounded beliefs as user cues become more indirect.
The benchmark separates affective support from epistemic sycophancy and tests conversational responses across autonomy, belonging, and safety needs. This is an ongoing model-evaluation project.
Methods and Tools
- LLM Evaluation
- Behavioral Benchmarks
- Controlled Comparisons