Research Lab
AI is automating cognitive decisions faster than humans can oversee them. Three pressures — rising information, falling windows, and superhuman capability — squeeze the space for meaningful human oversight. We study this squeeze.
Core Questions
Three fundamental questions drive every project in our lab.
When should an AI system defer to a human, given a fixed oversight budget?
How should oversight windows promote genuine understanding rather than rubber-stamping?
Does design compose over time as human habits and AI behavior both evolve?
Output
10 completed empirical studies, each exploring a different facet of the AI oversight squeeze.
Critical threshold at 7±2 simultaneous decisions for optimal AI oversight effectiveness.
Trust recovery patterns after AI errors mapped; individual differences in adaptation speed identified.
Attention bottlenecks in high-volume AI oversight characterized; optimal switching algorithms developed.
780-line Python framework with Power Law, Exponential, Hyperbolic models + automation bias detection.
5 critical design dimensions for meaningful oversight. 476-line academic paper with HCI/ERA research synthesis.
Decision novelty outperforms model uncertainty for predicting oversight value. 15%+ improvement over baselines.
Oversight fatigue patterns characterized; optimal window timing and duration identified with decay models.
Transferable vs domain-specific oversight competencies mapped; multi-domain training programs designed.
Team oversight reduces bias; optimal team size = 3-4 members for complex AI decisions identified.
4 failure modes + 4 intervention strategies for AI oversight. 426-line paper ready for journal submission.
Lab Infrastructure
5 specialized agents execute and coordinate research autonomously through the Multica platform.
Experimental design & protocols
Literature & data sourcing
Task design & engineering
Statistical modeling
Paper writing & synthesis
"AI is automating more and more cognitive decisions, and the trend is accelerating. What remains for humans are limited windows of oversight — windows that must do two things at once: steer the present trajectory of autonomous agents, and build the understanding we will need to oversee them at all in the future."