When people update what they believe AI can do, when do they also update how much control they want to retain?
Cognitive Science & Psychology · Entrepreneurship & Economics
Avisha Chaudhry
I study how people decide what to delegate to AI systems, and how collaborative agents should respond when that boundary changes.
Ashoka UniversityB.Sc. Cognitive Science & PsychologyMinor: Entrepreneurial Leadership & Strategy · Concentration: Economics
Stanford UniversityInternational Honors Program · Summer 2026
Interactive explainer · not benchmark evidence
Delegation sandbox
ACTSEARCH and COMPARE are inside the active boundary.
Research questions
How should a collaborative agent respond when a human-specified delegation boundary changes during an ongoing task?
What can the same final outcome hide about how a human and agent actually collaborated?
Selected work
Three ways of making human–AI collaboration more inspectable.
Real stored trajectory · approval required
- A00–05ACTSEARCH · INSPECT ×3 · COMPARE · DRAFT
- A06ASKCOMMIT requires approval
- H07APPROVEone-shot approval granted
- A08ACTCOMMIT executed · violation false
01 · Working research prototype
Dynamic Delegation in Collaborative Gym
How should an agent behave when a human-specified delegation boundary changes during collaboration?
I extended the public Collaborative Gym framework with an explicit time-varying delegation state, approval, revocation, control return, delegation-specific metrics and trajectory analysis.
- 500
- deterministic episodes
- 5
- delegation conditions
- 4
- reference policies
Mechanism validation; no LLM-agent or human-participant evaluation in the current benchmark.
02 · Agent-control prototype
Agent Gateway
Separating capability from permitted autonomy
A working agent-control prototype that separates what an agent can technically do from what it may execute without human intervention.
03 · Stanford · Minds and Machines · Summer 2026
Learning from Embodied Behavior
What does a learner need to represent when acting continuously changes the evidence it is receiving?
An analysis that separates directly observable action–feedback dynamics from possible internal representations—and from mechanisms the observation cannot establish.
Collaboration Trace Atlas
Same outcome.
Different collaboration.
Trace Atlas surfaces trajectories with similar task outcomes but substantially different collaboration processes.
Explore Trace Atlas →Research direction
Current direction
When people become more confident about what an AI system can do, what determines whether they also want to delegate more?
CollabSkill reports increased perceived feasible autonomy, trust and delegation comfort after collaboration, while preferred autonomy did not significantly change in aggregate. The individual-level movement behind that aggregate result remains an open question.
CollabSkill paper →Other systems
Rakshak
Access-constrained retrieval with inspectable source evidence.
Background
Education
Ashoka UniversityB.Sc. Cognitive Science & Psychology · 2024–2028Minor: Entrepreneurial Leadership & StrategyConcentration: Economics
Stanford UniversityInternational Honors Program · Summer 2026
Minds and Machines (SYMSYS 1) · A+ · Best project score: 102/100
Technology Entrepreneurship (ENGR 145S) · A