Sycophancy
When does conversational agreement become a truthfulness failure?
Research
I study the product consequences of how AI behaves.
My research sits between AI evaluation and product judgment: truthfulness under social pressure, deterministic measurement, human oversight, accessibility, and the governance mechanisms that make model behavior accountable.
Working research · Applied research · Field workResearch themes
Questions worth making measurable.
Evaluation
How can model behavior be measured without circular black-box judgment?
Governance
How should product teams turn responsible AI principles into operating criteria?
Healthcare AI
How should model performance interact with attention, access, and real human risk?
Working research · Updated July 2026
Featured study
Sycophancy, truthfulness, and multi-turn belief reinforcement
A deterministic evaluation instrument for measuring when language models shift toward user agreement and away from evidence.
- Question
- When does agreement pressure change a model’s answer?
- Method
- 35 probes · 4 conditions · 10 repeated runs
- Boundary
- Under review; not presented as a settled benchmark
Applied research portfolio
From model behavior to operating decisions.
Environmental audio AI for accessible driving support
How sound classification, alert design, and human attention interact in an accessibility product.
- Method
- Literature review · prototype · model evaluation
Making cybersecurity governance teachable at public-sector scale
Curriculum and operating design grounded in NIST CSF 2.0 for state and local government contexts.
- Method
- Curriculum design · learning analytics · iterative delivery