He “Albert” Zhang

This is the dissertation. Researchers were already pasting confidential interview transcripts into ChatGPT before anyone had established whether that was defensible. I started by building the tool that made the question answerable, then spent three years finding out what actually happens: where an LLM's codes agree with a human's and where the agreement is coincidental, how trust gets renegotiated the moment a researcher re-scopes the model, and what changes when the machine stops assisting the interview and starts conducting it.

I build the emotion-inference pipelines that privacy and ethics scholarship needs to interrogate from the inside. It began with fear in VR horror games — an emotion strong enough to show up unambiguously in pose, physiology, and gaze — and produced a dataset of natural rather than staged behaviour. From there the question inverted twice: first, can a general-purpose multimodal model read emotion off a face without training, and where does it fail? Then, having measured affect thoroughly, can a physical object regulate it rather than only record it?

Instrumenting a home to study it and giving that home an autonomous agent turn out to be the same engineering problem approached from opposite ends. At Tsinghua I built the multi-sensor testbed and the petabyte-scale pipeline behind it, which is also direct experience with the data-collection infrastructure that makes domestic privacy a live question. Years later, when multimodal models became capable enough to make decisions in that space, the obvious next question was adversarial: a home agent reads the room, and the room can be written on.

Camera-mediated assistance is where interpretation and privacy collide most directly: to get help seeing, a blind user has to let someone — or something — look at their room. BubbleCam reframed that as a scoping decision rather than a blurring problem. When large multimodal models arrived, blind users adopted them faster than the design literature could keep up, so the next studies documented practice first and derived implications second. NaviGPT tested whether any of it survives contact with a real street, and the current work is about the hardest interface problem in the set: how a system that might be wrong should say so.

A recurring finding across these studies is that the systems people depend on are held together by labour nobody is paying for. Government pandemic messaging worked or failed on form, not just content. Stranded travellers rebuilt an entire route network through personal infrastructuring. Twitch's community-management stack is built by unpaid third-party developers, and those developers in turn run their own support economy on Discord. Generative AI is now entering all of it, which makes the comparison I care about tractable: put human and AI-powered support side by side in the same community and see what changes.

The consequences arm of everything above. Universities wrote AI guidance quickly; the gap turned out to be between the policy language and what happens in a classroom, so the next study asked students rather than administrators. In public benefits, an LLM assistant does reduce the cost of learning a system — and relocates burden somewhere the evaluation was not looking. Parents applying their own responsible-AI criteria to children's learning do not reproduce the principles that published frameworks assume. Running under all of it is a philosophical commitment I made explicit early: an interpreting system is an instrument, and instruments mediate rather than transmit.