J.D. Zamfirescu-Pereira is an assistant professor in the Computer Science Department at the UCLA Samueli School of Engineering, where he researches human-AI interaction. He earned his B.S. and master’s degrees in computer science from MIT, and his Ph.D. in the same field from UC Berkeley. He has 13 years of industry experience, including stints at Google and co-founding a startup that Google acquired. His research focuses on joint human/AI systems, exploring how people interact with and design alongside generative AI without losing control or agency in the creative process. He was awarded the Google Ph.D. Fellowship in August 2024.

“In my class, students consider values that are not typically driving design in apps and tools like compassion, community, joy, connection, awe and fulfillment. In their final projects they explore new representations and interaction modes that give primacy to those new values. AI offers a critical mechanism for creating software driven by these values, which often require handling unstructured input in a flexible and context-sensitive way.” J.D. Zamfirescu-Pereira said.

Q: What are some of the main research projects that you are focusing on this academic year? How will your research be translated into new technologies?
A: A major research area in the lab I direct with computer science assistant professor Eunice Jun is AI-assisted design. By design, I mean anything from computer programs to scientific experiments to mechanical contraptions, and of course images and visual designs as well. We build software tools in domains that let us put people in new kinds of design situations that AI makes possible. These situations might be something like designing a chatbot’s behavior using only natural language or overseeing 1,000 agents. Then we study people interacting with, and using, those tools to draw lessons about the people, the AIs and the tools themselves — lessons we hope will help us all build computing systems that are more aligned with our human values, in a way that we have well-placed confidence in.

One recent project focuses on how to design better programming environments, where we consider ways to improve communication between human users and AI agents. We’re taking a cue from pair programming, studying signals like user gaze, emotional state and cursor position, to identify whether and how they can be used to find misalignment between human goals and AI assumptions earlier and then initiate repair to come to alignment.

Our hope is that these lessons, and the systems we build to study them, will pave the way for designers to be better equipped to see and understand the impacts of the decisions they — and the AIs that they are increasingly relying on — are making.

Several of the projects I’ve worked on recently have resulted in concrete project deployments at startups and big companies, including Google Gemini’s recently announced voice mode, that includes ideas from a project (Rambler) that I worked on in my Ph.D.

Q: How do you work with undergraduate and graduate students on these research projects? What are your key priorities and goals when advising students?
A: There are a few ways I work with undergraduate and graduate students. In the lab, most of this work is driven by Ph.D. and master’s students. They are the ones coming up with the technical and evaluation approaches and then implementing them. As an advisor, I’m primarily setting the high-level agenda and direction (and, of course, fundraising!) based on where I see opportunities to study human-AI interaction in new and impactful ways.

My goal is for my grad students to become independent thinkers with a deep bench of question-asking and answering approaches, and the ability to seek out new perspectives on human-AI interaction. Similarly, I want my undergrads to be able to ask good questions that push beyond the boundaries of established wisdom — a challenge that gets harder by the week as our lives become more infused with chat AIs that are trained to bring us exactly that established wisdom.

We are at an inflection point in how we interact with computing systems. AI is letting us move away from predefined structured representations — fixed Graphical User Interfaces, menus, code — to dynamic representations generated based on the tasks users take on and describe in natural language and other unstructured formats. This requires a new set of paradigms that we’re studying in my lab — and my hope is that my students will go on to academic and industry roles where they will help define these future interactions.

Q: How do you prepare computer science students to thoughtfully design and work with AI systems in real-world applications, and think critically about the ethical implications of the technologies they create?
A: This quarter, I am prototyping a new course called “AI for Human Interaction.” Among other topics, we learn about how software designers’ values are embedded in computing systems, including AI systems, and what that means for the power app and model designers have over users’ thinking. In my class, students consider values that are not typically driving design in apps and tools like compassion, community, joy, connection, awe and fulfillment. In their final projects they explore new representations and interaction modes that give primacy to those new values. AI offers a critical mechanism for creating software driven by these values, which often require handling unstructured input in a flexible and context-sensitive way.

Throughout the course, student groups periodically present their values and representations work in design critique, refining their ideas in response to instructor and peer feedback, and going deep in pushing the boundaries of established wisdom. But most of all, through design critique they learn how to think critically about the work they’re doing, why they’re doing it, what values are embedded in that work, and then how to consider which values they actually want embedded and how to make that happen.

Q: What are the biggest challenges in making AI systems more intuitive and accessible for people without technical backgrounds?
A: This question is what my dissertation was about! The upshot is that although chat interactions can feel intuitive, the “humanlike” nature of the interaction leads users to make humanlike assumptions about AIs. These assumptions can lead to misunderstandings and often hide what AIs are actually good at!

A classic example of the former is when a chat AI makes some representation about itself or its thinking process, like contritely apologizing for getting some code wrong, and promising that some new code is now right — but of course, it’s still wrong. There’s not a human-like accountability incentive for AIs, and this kind of high-stakes claim for a human (“I got it right this time, I promise”) just doesn’t carry weight for an AI. But people are still likely to overrely on these kinds of statements, to their own detriment.

As for chat hiding what AIs are good at, a classic example is “asking for 12 alternatives.” It takes quite a bit of power differential to demand a fellow human do something repetitive —but AIs really don’t care. If you asked a junior software developer for 12 different versions of every piece of code they wrote, they’d quit immediately. But an AI will do it, evaluate all 12, and give you a rationale for why you should pick one of the 12. Chat makes it look like you’re asking a human for something, and we carry a lot of baggage around social expectations there. Plus, comparing 12 versions of the same thing in a scrolling chat window isn’t ideal either.

One of the major goals we’re working toward is a notion of “informed agency.” This is the idea that a human user of an AI system not only understands what the system is doing at the right level of abstraction of that user’s knowledge and capabilities, but also can advocate and effect change in that system’s behavior and goals. Stay tuned for more on this concept soon!

Q: How could private funding through donor gifts enable you to further your research at UCLA?
A: All the projects — and courses — I’ve mentioned above require funding to make progress. For example, right now I’m looking to bring in postdoctoral researchers to work on two projects. The first is to develop frameworks around informed agency, which I believe are critical to the creation of AI systems that engineers and technology companies are already building today. The second is to work on nonverbal signals through a project that combines those signals with a highly synchronous voice interface agent, an area where we have already seen commercial attempts flounder because of their designers’ fundamental misunderstandings that our research aims to address. These two topics are super urgent in many ways, especially given the rapid pace of AI research in general, and private funding could help move the needle quickly in ways that will have deep follow-on effects for how AI systems are built in the years to come, all while supporting training for researchers who are on the cusp of launching their own faculty or industry research careers.

I love working with donors directly because they tend to be passionate about the topics they choose to fund. In my past work, I’ve found that donors can often bring specific expertise and unique insights that can positively impact projects.

Lastly, as many donors know, private funding gives us the flexibility to pursue this research agenda in the ways we believe will have maximum impact. Government grants constrain how we produce and disseminate our results, in ways private funding doesn’t. For example, grants typically require money to be spent at a given rate on a certain timeline, limiting our ability to dedicate disproportionately more resources to avenues that turn out to be unexpectedly productive. Similarly, the grant approval cycle is quite long, limiting how we can respond to new opportunities. Unrestricted funds give us the flexibility to pursue the highest-impact work as soon as we find it, in addition to letting us dedicate more of our time and effort to the research itself and to disseminating it.