What I have learned from teaching AI to urban planning students
Observations from the classroom on how students use AI critically, where things break, and what matters to learn now
I had a conversation in the printer room a few weeks ago with an architecture student who is an old-school artist. He didn’t know how to use a computer proficiently until he came to graduate school. We talked about the lack of deep, critical engagement with AI in our usual educational settings and organized events. We both felt that throwing around big words like “equity,” “sustainability,” and “ethics” can feel distant. So I asked, “What kinds of critical engagement do you actually find helpful?” He said he wants to hear more about how his peers critically engage with AI in everyday settings like learning and producing.
That’s also where I land.
When AI first became a major topic in our domain of research and education around 2025, our college conducted a survey to understand how faculty, students, and staff felt about it. We found that more than 60% had already used ChatGPT. For students entering my class, that number is essentially 100%. Many expressed concerns about AI’s social, educational, and environmental impacts, but few knew how to move beyond those sentiments. This cloud of anxiety and confusion can easily lead to polarized views about AI, which doesn’t help us move toward a healthier, more balanced understanding with AI. As an example, I learned an anecdote that a graduated student founded an AI startup and when they shared their achievements with peers in Slack, they received “thumb down” emoji from peers rather than support. On the other hand, our discipline attracts students who care deeply about the built environment and communities, and those cares could be the fire that shapes technology for public good. With the limited power I have as an urban and regional planning faculty working on and with AI, I realized that a new pilot course on Urban AI could be a much-needed space for students and for me to ground our thinking.
That is where I stand.
I want to teach students how AI actually works, to ground their arguments in real experiences of applying AI in urban contexts rather than borrowing ready-made statements, and to think about how to govern and build meaningful, value-driven AI systems. I also want to learn with students. Watching where they invest their time and how they grow is, to me, the best evidence and a collective signal of where our profession should be heading.
One year after teaching the first pilot class, my pedagogy paper,“What and How Should Urban Planners Learn in the AI Era? Exploring Urban AI Pedagogy from a Pilot Course in Urban Planning Education”, was published online in Journal of Planning Education and Research. The paper offers a thorough account of what I taught in the course, why I structured it that way, how the pieces connect, what “success” (i.e., the critical use and understanding of AI) looks like in practice, and how it can be evaluated. The Urban AI course was taught in Winter 2025 and Winter 2026, with a small group of urban planning and urban technology students who had statistics and programming backgrounds as prerequisites. You can find the course syllabus and student final project portfolio at my website.
Yet, much of what I wrote in the paper had to conform to academic standards, and as a tradeoff, I’ve been wanting a space to share observations and lines of thinking that didn’t quite fit. And, honestly, my colleagues gently warned me that pedagogy papers don’t usually get much attention compared to other publications, which is statistically true. So this blog post is my attempt to do two things at once: to share some thoughts grounded in substantive observations and classroom stories more informally, and to give this work the attention it needs to keep me going.
What is the paper about?
The published pedagogy paper addresses three “simple” questions: what should students learn about Urban AI, how should they learn, and how do they perform in the Urban AI pilot course?
Table 1 is a skeleton of my course modules, and Figure 1 reveals the underlying logic I used to connect the pieces together. All of these boil down to three learning goals: (1) applying AI effectively and appropriately to urban challenges, (2) addressing its social, environmental, and governance impacts, and (3) developing normative judgments and professional identities around AI.
The third learning goal is the hardest to learn and teach. Let’s start unpacking it through some of the most compelling empirical statistics from the course, grounded in one question:
How do students use and think about AI in the course?
I manually coded 655 quotes from 235 student AI reflection journals in the Winter 2025 course. These journals were required weekly submissions, where students described how they used AI or what they thought about AI that week. For each quote, I iterated through three rounds to label the application and reflection categories it fell into, as well as the level of criticality from low to high. Coding the quotes and computing frequency is not the most important part of this process. Coming up with clear definitions and boundaries for categories and levels of criticality is far more valuable, as they help educators anticipate the range of activities and evaluate success.
For example, Table 2 shows how to advance writing and text analysis tasks from the lowest level of application and reflection to the highest level.
In Figure 2 below, I mapped the frequency of each type of AI use and line of inquiry across the levels of criticality. I want to take a moment to appreciate the diversity and range of uses and thoughts that students explored. For instance, I never expected students to use AI for life planning and task productivity, something that even precedes the popularity of agentic AI in 2026. I was also pleasantly surprised by students’ reflections on AI’s emotional intelligence, and the feedback loop between applying personal background to guide AI use and using AI to inform professional development and identify gaps. These are topics we never explicitly introduced in class.
At the same time, the distributional statistics and interpretations should be taken with a grain of salt, as the data come from 20 highly capable, technically prepared students, and the frequency can be shaped by category definitions and the sequence of content students encountered in class.
In terms of ranking, students most frequently used AI for tasks under the theme Understanding, Reasoning, and Expression (59 uses), followed by Code Development and Technical Assistance (55 uses), and Data and Text Analysis (54 uses). This suggests that students relied heavily on AI for cognitive, coding, and analytical tasks.
Students also most frequently reflected on AI Strengths and Weaknesses (133 instances), mostly at lower critical levels, suggesting that they often begin with descriptive evaluations of what AI can and cannot do. In contrast, Effective and Appropriate Use in Context (92 uses) and Human–AI Co-evolution (82 uses) show a strong concentration at Levels 3 and 4, indicating a shift from description to judgment and synthesis as students connect AI use to professional responsibilities and identity formation.
Learning AI Mechanics Is Key for Advanced Criticality
I particularly like the ladder of growth from AI Strengths and Weaknesses to AI Mechanics to Effective and Appropriate Use in Context, which is also reflected in the increasing criticality of the reflection journals in these categories. From my teaching observations, moving to Levels 3 and 4 is a significant bottleneck for students to become more critical users and thinkers of AI. Sometimes, it takes weeks of pushing through journal feedback before a student can move beyond statements like “I tried X and realized Y”. Thus, whenever I see an opportunity to connect students’ experiences and observations to AI mechanisms and to more systematic strategies for testing hypotheses, I point it out explicitly.
I personally believe the key lies in learning AI mechanics. For instance, knowing the “fact” that AI can hallucinate or embed racial bias may only make someone cautious about all AI applications. But understanding why and how these issues are built into the models gives students the agency to improve them and to judge whether proposed solutions address root causes or merely scratch the surface.
The tension lies in a sense of “fear” around teaching highly technical knowledge to non-technical students (where most planning students are traditionally placed): concerns that they may not be prepared, interested, or that it may not be necessary, and uncertainty about the appropriate level of detail. Yet, repeatedly, my students tell me this is the most valuable part of the course and they enjoy learning how AI works. A deeper understanding of AI mechanics also enhances critical thinking in a grounded way, as students experience it firsthand through building their own systems rather than borrowing generalized statements (see student reflections below).
“This project, and overall course, have pushed me to think critically about AI not as a magical layer added to cities, but as a process of translating urban complexity into data, features, assumptions, and interfaces. One thing I learned early on is that the hardest part of applied AI is often not the model, and in this case, it was actually the data pipeline, feature logic and validation design. Building the master dataset required much more care than simply fitting a model. That alone was a useful reminder that real AI systems are built on data engineering decisions that often disappear from the final interface.” — from a final project building an energy demand model
“Taking this AI class, especially how it focused on actually using AI as a tool and showing us some of the computer science behind it, really changed things for me. In a world where AI is suddenly in everything and you’re not sure what’s real value and what’s just hype, learning how these tools fundamentally work was huge…. Learning about AI through this course has given me a much healthier, more balanced relationship with it. I really see now how powerful it can be when you apply it within your own field, like Urban Planning, to make tasks easier or unlock new analysis. Realizing I could learn how to personally make and use a tool to categorize a huge dataset of things, like dams or whatever else, is just incredibly cool. This whole project made me feel way more empowered compared to when I’ve just used some black-box AI online. Actually programming it myself and running it on my own computer gave me a sense of agency I didn’t have before.“ — from a final project building a machine learning model to classify dams in the U.S.
How do we further discretion, judgements, and identities around AI?
After teaching the course for the first year, it became clear to me that content and skills can change rapidly, yet teaching students how to develop discretion, judgments, and identities around AI is the most important learning goal. There is simply too much to learn, and it is easy to lose and exhaust yourself by chasing the “hottest” topics and skills. AI anxiety is real. Even for me as a faculty member, my research can quickly lose value or contribution if I don’t publish fast enough or position it strategically to compete with large companies’ ability to absorb and subsume smaller innovations at scale.
In the Winter 2026 course, I surveyed students on a few curious questions.
If AI can do parts of the work faster or better, what exactly are we gaining, and what are we quietly giving up?
For the same task, students reported what they spend less time on and more time on when using AI. A few responses converged on spending less time on problem-solving processes (e.g., debugging, troubleshooting), searching for information, and linear or repetitive tasks, and more time on conceptual understanding, interpretation, and judgment, all of which used to be integral parts of learning, but now carry different weights.
I then asked the same question again, but expanded it to time tradeoff in academics and life in general. Some responses suggested a shift away from traditional academic work toward a better quality of life. Others felt that AI helps reduce anxiety and frees up time to actually get things done, do more, and reflect more, either on AI or on themselves. This reminds me of Anthropic’s research on “what 81,000 people want from AI,” which shows that the top two desires are professional excellence (getting meaningful work done faster and better) and personal transformation (personal growth and physical/mental health).
All of this sounds good, and it seems like students know what they want—until the logic starts to crack at the collective level. I call this the “AI Shortcut Paradox”: these choices are well justified for a single task or at the individual level as a shortcut, but can produce worse collective outcomes.
The first nuance is rushing into work with shallow understanding. More than once, I received weak student project proposals that outlined a seemingly plausible path (likely generated by AI), but failed on conceptual grounds. For instance, a student proposed evaluating the performance and equity implications of an AI-driven adaptive traffic signal system, but relied on GPT-generated answers to explain how such systems typically work, and then jumped straight into analyzing data without examining the case in detail. Another group trained a machine learning model to predict land use change and even fine-tuned it to improve performance, without thinking through whether the variables themselves were appropriate. These problems are not unique to AI, but AI can mask them and give false signals and confidence to move forward.
The second nuance is the breakdown of group projects. While group work can fail for many reasons, I believe part of it lies in how individuals allocate their time, combined with the fact that AI workflows are inherently more individualistic than before. In a “good” scenario, a group has a few highly motivated students who propose the topic, carry most of the technical work, and assign marginal tasks to others. AI accelerates this imbalance because it is less effective to split core technical work across multiple people, and the gains from improving coordination and communication are minimal.
In a “worse” scenario, high-capacity students become disengaged because they are not working on a topic they care about, or because they decide it is not their priority, leaving teammates struggling. Again, AI exacerbates this dynamic: people become more selective in how they allocate their time and attention, which may benefit individual development, but harms teamwork.
We also discussed other examples in class, such as using AI to polish survey responses or job application materials. On an individual level, these uses are perfectly justified. However, from the perspective of survey designers and resume reviewers, the homogeneity of AI-polished responses can erase their ability to understand genuine human experience and qualities. This is concerning to me because intuitive discretion works in some areas, like data ethics or racial equity in AI, even without explicit teaching. But these seemingly “value-neutral” zones, often framed as productivity or efficiency gains, are where AI quietly embeds itself in everyday decisions and contributes to worse collective outcomes without clear warning signs.
In a world where AI can help with almost everything, how do we decide what is actually worth doing ourselves?
If AI can do something well, is it still worth learning? In the era of AI, what is most important to invest time in? (e.g., learning skills, following curiosity, learning how to use AI effectively, deep thinking / reflection, etc.). I threw these two normative survey questions to students, and want to tap into their normative beliefs on AI.
For the first question, the answers generally fall into three camps: 1) we need to learn the same things so we know how to better use AI on the same tasks, 2) some are worth learning and some are less in the AI era, and 3) learning (even at a slower rate as AI) is fundamentally how humans get better at thinking, judging, and creating.
This becomes more interesting when combined with answers from the second question. More students agree that it is most important to invest time in thinking, feeling, and following curiosity in the era of AI, yet building skills were also rated at an equal level of importance. The first part is something I deeply agree with and increasingly raises importance across my teaching to encourage students to do projects that show their personal characters rather than completing an operational task. Yet the second part is counterintuitive because AI is surpassing humans at most cognitive, coding, and analytical skills rapidly. But, if we believe that learning (and learning skills) is fundamental to human growth and development, rather than a means to the end, then the comparative rate of learning matters less. To not be beaten by the efficiency game, I think the key here is to learn the skills that matter to you because it satisfies your curiosity, ways of thinking, and who you want to be, rather than simply chasing a skill defined by the market.
What does planning and urban technology background bring to the table?
My favorite read of the class is the end-of-semester final project reflection where I asked students to write “how urban planning / urban tech domain expertise allow them to do the tasks differently, ask different questions, or realize implications that are often overlooked by engineers/AI developers”.
Pulling from some of my favorites, the answers usually vibe with having “domain expertise to know what questions to ask, what to prioritize, what to validate/is likely to fail, and the interpretive framework to approach a challenge”. These small discretionary decisions can quickly pile up and significantly shape the final products, as well as the values and utilities that the product delivers. Here are some examples from students:
“More importantly, this project reinforced for me that urban planning and urban tech domain expertise matters. Someone approaching this only as a machine learning problem might focus on squeezing out another fraction of a percent in predictive performance. But domain expertise changes the questions you ask. It shifts, or perhaps expands, the focus of the project from just the accuracy of the model to the kinds of urban behavior it may highlight, what decisions this forecast may influence, who benefits from its existence and who inevitably suffers a loss.
Urban planning knowledge also changes how we interpret infrastructure. Energy demand isn’t just a technical time series. It is connected to housing form, mobility, climate adaptation, institutional schedules, and neighborhood vulnerability. This means that a peak-demand forecast is also a heat-risk question, a load spike can be related to building systems and land use, and a model failure on holidays or event days can reveal something about how urban rhythms are organized.
This project also made me more cautious about the rhetoric around AI in cities. It is easy to imagine systems like this as steps toward a seamless smart-city future. But even a small prototype shows how much these systems depend on simplification. Forecasting can be useful, but if it is framed carelessly, it can create a false sense of objectivity. Overall, this assignment advanced my understanding of AI by making me work through the entire chain; from raw public data, to feature engineering, to model selection, to validation, to interface design, to critical reflection about how such a system would actually function in an urban setting.” —- from a final project building an AI-driven energy demand model.
“The most striking finding is that dimensional standards (height, FAR, lot size) are almost entirely null — not because GPT failed, but because Detroit’s Zoning Ordinance deliberately stores these values in Article XIII, referenced by each district section as: “Development shall comply with the intensity and dimensional standards provided for in Article XIII.”
GPT correctly returns null when the information is not present in the provided text. An engineer might interpret this as a model failure. A planner immediately recognizes it as a cross-reference writing convention common in municipal codes — where dimensional standards are consolidated into a separate table to avoid repetition across dozens of districts.
This is precisely the kind of domain knowledge that separates urban planning expertise from general AI development. A system built without planners’ input would likely misdiagnose the problem and attempt technical fixes (larger context windows, better chunking) rather than the correct fix: restructuring the document to co-locate dimensional standards with each district section, or implementing a multi-step retrieval that explicitly queries Article XIII.“ — from a final project building an AI pipeline to extract zoning codes.
“An engineer or computer scientist working on these models might have just focused on the factual accuracy of these models and whether the model is right or wrong. However, fields like urban planning and building science are far more complex than this. A lot of what happens in these fields requires practitioners to question:
Consequences of actions - if something is slightly incorrect or biased, how and who could it impact on a high and low level.
Importance of critical thinking - if asked a leading question, it’s a practitioner’s job to correct said information and guide others to the correct answers, even if they are narrow (i.e., question 15).
Temporal and spatial sensitivity - it’s a bit more specific to building science than urban planning or design, but things that might work in one climate or ecosystem might be counterintuitive in others. Especially since most high-performance standards are Western and deal with predominantly (but not exclusively) colder climate zones and new-builds, there is a decent percentage of areas where best practices are underdocumented.” — from a final project auditing how well AI answers building science questions.
“As I stated earlier, this project was driven by my studies in another class for my urban technology degree. I think being an urban technology student helped me develop a sense for what type of product might actually succeed in this context— I don’t know if there is any other method that has less friction for public reporting than just a simple text message. The whole idea was to make it easy for the public to alert the government of issues while ALSO making it easy for the government to take action on the incident. My implementation allows both of these to happen, and I don’t think I would have had the idea to do it this way if I was not aware of the troubles relating to government bureaucracy and also the difficulties urban citizens face when trying to get the government’s attention.” — from a final project building an AI pipeline to convert text messages into a dashboard for public reporting on 311 nuances.
“This project expanded my view of how AI can be used in urban contexts—not just to automate tasks, but to make civic participation more accessible. By building a voice-based chatbot for pothole reporting, I explored how AI can empower residents to engage with public infrastructure more directly and conveniently. From an urban planning perspective, I wasn’t just solving a technical problem—I was thinking about real-world user behavior, accessibility, and how to reduce friction in public reporting. While the chatbot isn’t fully hands-free yet, it shows potential for future development into a mobile app that could make real-time infrastructure feedback safer and easier for drivers.” — from a final project building an AI-driven pothole reporter
“One thing I learned fast: NLP models like spaCy are helpful, but only if you already know what you’re looking for. I had to bring in my Urban Tech background to: decide which rules mattered (e.g. setbacks, permits, shoreline types), design the structure of my dataset based on real planning needs; spot and fix extraction errors the model made (especially with vague or incomplete phrasing). So the AI made it faster, but I still had to guide the logic and know the context. That made it feel like a real collaboration between human knowledge and machine efficiency.” — from a final project extracting costal resilience zoning codes.
“My background in urban planning and public health shaped how I approached this project. From a planning perspective, I was especially interested in how opposition was tied to governance processes, local autonomy, and land use conflicts. Rather than treating opposition as a binary outcome, I focused on how it was framed and who was positioned as a decision-maker. My public health training also informed my attention to community voice, procedural equity, and the social dynamics behind policy disputes. These perspectives influenced the themes I looked for and the questions I found most meaningful.” — from a final project using AI to analyze the politics of solar power in Michigan.
Closure: Shifting My Role as an Educator
Beyond the course, I’ve been thinking a lot about my role with students in the era of AI. On one hand, I feel students’ pressures very closely. Even in my own research, it has become harder and rarer to have structured, semester-long opportunities with junior students. I increasingly find myself needing either very immediate, hands-on support (realizing a week ahead that my pipeline needs human validation), or much more independent, self-navigating collaborators (a project lead I don’t need to hold hands). Similar to the broader market shock facing entry-level graduates, the opportunities I can offer students are shrinking for similar reasons.
On a personal level, I would prefer to invest in growing with high-capacity students, even though they may be less efficient, less stable, and require more “maintenance” than AI. But that boundary feels fragile. The risk that a person, even a very capable student, might drop off or disengage for a variety of reasons is real and often unpredictable, and that uncertainty can be costly for my research timeline. My tolerance for this may increase if I am tenured, but for now, I have to prioritize productivity.
On the other hand, I want to encourage students to develop independent projects that cultivate the kind of discretion and identity-driven inquiry that feels essential in a rapidly changing world. I also believe that, in the era of AI, building a personal portfolio of narratives, projects, and evidence around your own interests can be more rewarding than working on mine. The current educational environment leaves limited room for this kind of exploration, and I would burn out quickly if I tried to advise too many individual projects at that level.
My latest thinking is a shift: from teaching skills and mentoring individual projects to curating a space and structure where students can thrive. This summer (2026), together with a group of students, I am experimenting with a semi-structured, peer-driven co-working space for people who want to do meaningful, self-directed work in the presence of others. Beyond simply providing a desk and some structure for productivity, I want to rebuild something that has been missing: a socially and intellectually productive environment where people are committed to growing along a few dimensions:
Working on questions that matter, with a grounded sense of why they matter
Taking ownership of moving your work forward
Making your work legible and meaningful to others
Being aware of your own growth, and actively contributing to others’
The manifesto sounds ideal. In practice, the question of who takes the lead and does the work remains a constant struggle. It almost mirrors the group project dynamics I see in class, except this time, I don’t have AI to blame.
And maybe that is the point. Even as AI reshapes how we work, the harder problem hasn’t changed: how people choose to show up, take responsibility, and grow together.