Skip to main content
School of Architecture, Planning and Preservation
Programs Admissions Our Work Student Experience About
Programs
Undergraduate Programs Graduate Programs Certificate Programs High School Summer Programs
Admissions
Scholarships and Financial Aid Visit Information Sessions
Our Work
Faculty Research Professional Practice Student Work Research Centers & Institutes Research Labs Galleries and Exhibits
Student Experience
Campus to Capitol Mentoring Programs Student Organizations Spaces and Studio Education Abroad Competitions Professional Development Student Resources Alumni
About
People News and Events Accreditation Contact Us Giving
Undergraduate Programs Graduate Programs Certificate Programs High School Summer Programs
Scholarships and Financial Aid Visit Information Sessions
Faculty Research Professional Practice Student Work Research Centers & Institutes Research Labs Galleries and Exhibits
Campus to Capitol Mentoring Programs Student Organizations Spaces and Studio Education Abroad Competitions Professional Development Student Resources Alumni
People News and Events Accreditation Contact Us Giving
Student Resources Alumni News and Events Giving
Student Resources Alumni News and Events Giving
National Center for Smart Growth

To Claude or Not to Claude? Maybe That’s Not the Question

AI is Polarizing for Many, But a UMD Data Scientist Sees a Middle Ground

Home About News and Events News To Claude or Not to Claude? Maybe That’s Not the Question
Screenshot of the Claude chatbot saying Good Afternoon, Robert

When Assistant Research Professor Chester Harvey begins a conversation with students around AI, he often asks them to participate in a quick exercise: line up across the wall between "enthusiastic adoption” on one side and “I won’t touch AI with a 10-foot-pole” on the other.  

These days, the in-between is pretty sparse—and that’s what Harvey wants to change. As an affiliate with the Artificial Intelligence Interdisciplinary Institute at Maryland, he’s currently collaborating with colleagues to design new curricula for the School of Architecture, Planning and Preservation to help students determine, not if they’ll use AI, but when, how, and why.  

“It’s a different way of approaching AI and critical for our students as they navigate the role of humans, the role of artificial agents and the space in between,” he said. “That’s the direction we’re trying to take: Not how to button-click off-the-shelf tools to replace human work, but to think critically about how AI can be used to support human intelligence. Where does it add value beyond human capabilities? And at what cost?”

An urban planner and data scientist for UMD’s National Center for Smart Growth, Harvey knows the limitations of AI, but also its potential. He has more than a decade of expertise in “old-school AI,” like machine learning prediction models; more recently, his team is using large language models (LLMs) and other generative AI in their research, such as examining housing instability along the Purple Line light rail line. But asking the stickier questions, he says—like how it will change our work and relevance as humans, and its impact on the environment, economics or society— helps delineate roles for both machines and humans, and the value of collaboration between them. 

Below, Harvey talks about AI’s surprising benefits, how being critical of its use is the secret to using it wisely, and why humans will always matter: 

You’ve been using AI as a research tool long before the emergence of LLMs like ChatGPT. What does your work look like these days? I think AI is excellent at doing some of the heavy lifting that in turn frees up my time to develop a better product. I’m using LLMs in a pretty conventional way to do coding. It dramatically increases what types of analysis we can develop and allows me to focus on design thinking instead of the minutiae of code syntax. This allows me to put more effort into the human and social dimensions. How do I want this system to be built? What are the security and ethical implications? How does it relate to the users? That feels like a bit of a win-win. 

How has AI been useful in ways you didn’t initially predict? There’s a big debate within AI about how much you need to verify its outputs, but there are multiple sides of that coin. As a human user, you need to take ownership and verify and correct it if it's going off course. But there are a lot of situations where AI allows us to check things that we were never able to check before. We can quickly scour a huge corpus of data for errors or gaps, something that would have taken dozens, maybe hundreds of hours of human work. We have a couple of projects that use AI tools to help verify and connect data that we would otherwise never have had the capacity to examine so deeply. 

There is a lot of discussion about AI’s capabilities, which can stoke fear around what’s left for humans to contribute. In your experience, what can’t AI do? For now, critical and creative thinking are things AI can’t do. An LLM might look like it’s being creative, but it’s just good at convincingly reproducing patterns, not coming up with genuinely new ideas. That’s a uniquely human facility, at least for now. Understanding that is crucial for interpreting where there is value and contributions from these tech tools and where there is risk. Despite all the “learning” LLMs have been doing on data scrounged from the internet, they also often don’t have all the worldly context they would need to make sound decisions. We’ve had situations where an LLM has proposed storing information in a place that isn’t secure enough. It reveals that, despite being sophisticated, LLMs still suffer from the same garbage in / garbage out issue that has always plagued modeling: not the right information going in, poor decision coming out. I’m dubious that AIs will ever fully replace human expertise, especially in understanding context. So, as a human, I think you want to be an expert in context and be able to think critically about what solutions are not just possible, but appropriate.

AI’s confidence is a bit unnerving. What’s the cautionary tale here? Maybe this is simplistic, but I think in some ways it's like hiring a person who has certain skills but is also really really good at bullshitting. You need to find a way to leverage their capabilities but not take everything at face value. An unfortunately common way to use an LLM is just to give it free agency, and perhaps that’s not very different from putting all of your faith in a human assistant. BUT we’re used to reading the capabilities of a human and have well-developed heuristics—are they well-organized; does their writing make sense?—to help us judge their reliability. Unfortunately, those don’t map very well onto judging an LLM. It sounds confident, and it’s really good at making you believe it knows what it's talking about. You have to change your heuristics, and that takes awareness and experience. 

Back to the “spectrum of use exercise” that you conduct with students. What might help nudge them a bit more towards the middle? Students know that people aren’t going to be wholesale replaced by AI; human engagement and context awareness is such an important part of any practice that if anything, it increases our value. But I think they also know that a job seeker who knows how to use AI is more competitive than one that doesn’t.

I have a colleague who is a skeptic. I’m certainly adopting it more and tend to be more optimistic. We’ve had some really useful dialogues together in front of our students. We’ve discussed the opportunities and constraints and try to think both critically and creatively about the role that AI can play, without a starting place that isn’t “this is great,” and “this is terrible”. I understand the tendency but I don’t think that’s realistic—and I think there is real benefit from measured and tactical adoption. 

School of Architecture, Planning and Preservation
3835 Campus Drive, College Park, MD 20742
archinfo@umd.edu 301.405.8000