Welcome!
- zhuohao [at] uw.edu
- Google Scholar
- @ZhuohaoZhang
- Zhuohao (Jerry) Zhang
My name is Zhuohao (Jerry) Zhang. I am a final-year Ph.D. candidate at the University of Washington, working with Prof. Jacob Wobbrock at the ACE Lab. I obtained my M.S. degree in CS from the University of Illinois, Urbana-Champaign, where I worked with Prof. Yang Wang, and my B.Eng. degree in CS from Zhejiang University, China. I also worked closely with Prof. Anhong Guo at UMich.
AI models can now design slides, write code, and operate user interfaces on our behalf, but their work still requires human judgment: someone must determine whether an output is well designed, faithful to intent, and safe to execute. My research supports that judgment from two directions. I build benchmarks that measure how well models themselves understand design quality and the consequences of their actions, and I build interactive systems that let people inspect and correct AI output rather than take it on faith. I develop these methods with blind and low-vision creators, who cannot glance at a screen to catch an AI’s mistake. If oversight can be made to work for them, it can work for anyone.
I am a recipient of the Apple Scholars in AI/ML PhD Fellowship (2024), awarded to roughly 20 students worldwide each year. In 2026, I was selected for the Anthropic AI Safety Fellows program from a pool of over 10,000 applicants, an acceptance rate below 1%.
Recent Experiences
Selected Research
Steering What AI Creates, Verifying What AI Does
AI now writes, designs, and acts with remarkable fluency, but fluency hides mistakes. The more work AI does on our behalf, the more depends on whether a person can still shape that work and still check it.
Across my systems, the recurring solution is an intermediate representation: structure placed between the person and the model, such as a taxonomy to judge with or a routine to audit before it runs. Good representations are precise enough for machines and legible enough for people.
I develop these ideas with blind and low-vision creators, for whom unverifiable AI is unusable AI. They are the most demanding case for human-AI trust. What they need first, every user will need next, as AI output outgrows our ability to review it.
Slides are usually treated as purely visual. A11yBoard rethinks the screen reader for 2-D canvases, making slides perceivable and editable without sight. It was later deployed for Google Slides (ASSETS 2023).
Can AI critique design reliably? Only with structure: our expert-built taxonomy and 2,400-slide annotated dataset substantially improve models' ability to detect design flaws and to propose fixes that work.
A taxonomy and benchmark for whether AI agents understand the consequences of their UI actions: what is reversible, what is not, and who else is affected. Frontier models still misjudge exactly the actions that carry real risk.
Enables blind programmers to verify AI-generated 3-D models they cannot see by keeping code, semantic hierarchy, AI description, and rendering synchronized, supporting independent modeling that previously required sighted help.
Turns repetitive screen-reader navigation into user-authored routines: goals described in natural language are compiled into triggers, filters, and actions that users can inspect, correct, and reuse.