AI Policy Wiki
Dashboard
⚲
AI Policy Wiki › People & Organizations

Kevin Feng

medium confidence · updated 2026-06-09

University of Washington PhD candidate (HCI / Allen School); lead author of Levels of Autonomy for AI Agents (Knight Columbia, 2025), proposing a five-level user-centered framework for AI agent autonomy.

Contents

Kevin Feng is a PhD candidate at the University of Washington (Paul G. Allen School of Computer Science & Engineering / Human-Computer Interaction) and the lead author of Levels of Autonomy for AI Agents (Knight Columbia, 2025), a five-level user-centered framework for AI agent autonomy.

Background

Feng works on human-AI interaction, agent autonomy design, and the user-centered evaluation of AI agents. He is advised by Amy X. Zhang and David W. McDonald at the Paul G. Allen School (Uw).

Feng describes himself as a final-year PhD student in the Human Centered Design & Engineering (HCDE) department at the University of Washington, co-advised by David W. McDonald (HCDE) and Amy X. Zhang (Allen School of Computer Science & Engineering) (Source: https://kjfeng.me/). According to his Knight First Amendment Institute biography, his research interests lie at the intersection of social computing and human-centered AI, with a focus on designing sociotechnical methods and interactive systems to advance AI usability, safety, and governance (Source: https://knightcolumbia.org/bios/view/kevin-feng). He states that his broader interest is in how advanced AI can diffuse productively and safely throughout society, including human-centered approaches to alignment and scalable oversight (Source: https://kjfeng.me/).

Before UW, Feng received a B.S.E. in computer science from Princeton University, where he conducted research at the Center for Information Technology Policy (CITP) advised by Arvind Narayanan, and completed a minor in visual arts (Source: https://kjfeng.me/).

Roles and affiliations

Feng's work has been supported by an OpenAI Democratic Inputs to AI grant (2023) and a UW Herbold Fellowship (2022) (Source: https://kjfeng.me/). He was a summer fellow at the Centre for the Governance of AI (GovAI) in 2025, interned at the Allen Institute for AI (Ai2) in 2024 and at Microsoft Research in 2023, and has held a student researcher role at Google Research (Source: https://kjfeng.me/) (Source: https://knightcolumbia.org/bios/view/kevin-feng).

Work on agent autonomy

Feng is the lead author, with David W. McDonald and Amy X. Zhang, of Levels of Autonomy for AI Agents (Knight Columbia, July 2025; republished in the May 2026 symposium) — see Levels of Autonomy for AI Agents (Feng-McDonald-Zhang framework) and Levels of Autonomy for AI Agents (Feng + McDonald + Zhang, Knight Columbia, 2026). The framework treats autonomy as a design decision separable from capability, operationalized through five user-role-centered levels and an autonomy certificates governance mechanism (Autonomy Certificates).

The paper was posted to arXiv as arXiv:2506.12469, submitted 14 June 2025 and revised 28 July 2025, and was published in the Knight First Amendment Institute's "AI and Democratic Freedoms" essay series (Source: https://arxiv.org/abs/2506.12469). The authors argue that an agent's level of autonomy can be treated as a deliberate design decision, separate from its capability and operational environment, and define five levels of escalating autonomy characterized by the role a user takes when interacting with an agent: operator, collaborator, consultant, approver, and observer (Source: https://arxiv.org/abs/2506.12469). The paper describes a potential application toward AI autonomy certificates for governing agent behavior in single- and multi-agent systems, and proposes early ideas for evaluating agents' autonomy (Source: https://arxiv.org/abs/2506.12469).

His research engages with Agentic AI, Agent Architecture Patterns, Agent Autonomy Spectrum (5 Levels), and Principal-Agent Problem Applied to AI.

Other research

Feng's other projects span tools and techniques for human-AI interaction, AI policy elicitation, and agent steerability:

  • Cocoa ("Co-Planning and Co-Execution with AI Agents"), on techniques for human-agent collaboration, received a Best Paper award at CHI 2026 (Source: https://kjfeng.me/) (arXiv:2412.10999).
  • PolicyPad ("Collaborative Prototyping of LLM Policies"), a system for collaboratively prototyping LLM behavioral policies, was published at CHI 2026 (Source: https://kjfeng.me/) (arXiv:2509.19680).
  • Canvil ("Designerly Adaptation for LLM-Powered User Experiences"), aimed at helping designers craft AI-powered user experiences, was published at CHI 2025 (Source: https://kjfeng.me/) (arXiv:2401.09051).
  • Work on responsible AI-generated legal advice, engaging legal experts toward responsible LLM policies for legal advice, was published at FAccT 2024 (Source: https://kjfeng.me/) (arXiv:2402.01864).
  • Work on case-based reasoning for AI alignment ("Democratic Inputs to AI") was presented at a NeurIPS 2023 workshop and connects to his OpenAI Democratic Inputs to AI grant (Source: https://kjfeng.me/).

Relationships

Sources

Categories: People & Organizations | academic | framework | technical | frontier-models
What links here (9)