Across educational contexts, AI is increasingly positioned as an essential literacy for contemporary learners. Curricula introduce foundational terminology, workshops demonstrate emerging tools, and classrooms experiment with AI-assisted activities. These developments signal a broad recognition that AI is socially and economically consequential. Yet the structure of AI instruction reveals a recurring pattern: the emphasis lies on access and application rather than on the conditions of understanding.

AI is frequently presented as a capability to be explored—what it can generate, automate, or optimize. Instructional sequences often move quickly from introduction to use, assuming that interaction itself will produce comprehension. However, interacting with an AI system does not necessarily clarify how that system functions.

Several structural tendencies are observable:

• System behavior is encountered as output rather than examined as a process.

• Engagement is organized around completion of tasks instead of interrogation of mechanisms.

• Explanations of AI rely on simplifications that obscure the limits and constraints governing system behavior.

These tendencies do not reflect a lack of instructional intention. Rather, they reflect a broader orientation in which AI is treated as an emerging tool category that learners must become familiar with.

Under this model, familiarity risks being mistaken for understanding.

Reframing the Educational Question

If familiarity is insufficient, then the problem may not be the pace of AI adoption, but the way the learning problem itself has been defined.

Current discussions often frame AI education as an issue of preparedness: ensuring that students are equipped to operate within an AI-infused world. While this framing is pragmatic, it centers future utility rather than present cognition. It assumes that exposure and practice will naturally give rise to conceptual clarity.

However, complex systems do not become intelligible through exposure alone. Understanding emerges when learners are positioned to examine relationships, test assumptions, and encounter the limits of their existing mental models.

The central challenge, therefore, is not simply how to introduce AI into classrooms. It is how to create instructional conditions in which learners can grapple with the structure of system behavior.

This reframing shifts attention away from adoption and toward design:

• What kinds of learning environments make system logic visible?

• How can instructional structures make reasoning traceable rather than opaque?

• Under what conditions do misconceptions become observable rather than hidden?

By redefining the problem at the level of instructional design rather than technological inclusion, we create space to reconsider how AI learning environments should be structured.

The following section outlines how the AI Game Project responds to this design challenge.

AI as a System for Conceptual Reasoning

Under this reframing, we can understand AI as a structured system whose behavior can be examined, tested, and debated. The AI Game Project adopts game-based, unplugged design as a pedagogical response to the conceptual demands of AI learning.

This approach rests on several foundational commitments:

• Conceptual understanding should precede computational implementation.

• Learning activities should surface, rather than bypass, common misconceptions.

• Instructional designs must remain feasible across low-resource and diverse learning contexts.

By positioning games as environments for reasoning, the project treats understanding as something that develops through structured interaction with rules and feedback. In doing so, it reframes AI education as a process of conceptual inquiry rather than technical exposure.

A Principled System for Conceptual AI Learning

In response to the limitations identified in prevailing approaches to AI instruction, the AI Game Project advances a principled system grounded in the assumption that conceptual understanding emerges through structured interaction. Instead of positioning games as motivational supplements or engagement tools, the project treats game design as a pedagogical system capable of organizing reasoning, surfacing misconceptions, and supporting conceptual change.

This system is guided by a small set of design principles that function as operational guides:

• Each game is anchored in a limited set of predefined AI concepts, with learning goals explicitly articulated in terms of conceptual understanding, misconception reduction, and knowledge transformation.

• Game mechanics are designed as explicit, inspectable rule systems that externalize how decisions are made, enabling learners to reason about system behavior.

• Common misconceptions are treated as anticipated reasoning patterns and are intentionally surfaced, observed, and analyzed through gameplay and debriefing

• All game designs are constrained by feasibility considerations—including material cost, facilitation burden, classroom time, and technological access—so that learning outcomes are not dependent on idealized or high-resource conditions.

Together, these principles define a system in which learning is structured through carefully designed constraints that guide how learners interact with AI concepts over time.

Operationalizing the System in Practice

The AI Game Project operationalizes by treating implementation n as a continuous process embedded across design, enactment, and evaluation. The project begins by specifying a small set of target AI concepts and common misconceptions, and then translating those conceptual commitments into explicit game rules and mechanics. In this way, gameplay is designed to function as the learning space where learners encounter, reason through, and revise their understanding of how AI systems behave.

In practice, this system unfolds through iterative prototyping and classroom enactment, where games are tested, observed, and revised based on how learners actually interact with them. Gameplays are documented and analyzed to examine patterns of understanding and misunderstanding. Throughout this process, feasibility constraints such as low-resource settings, limited instructional time, and facilitation demands are treated as core design parameters rather than limitations. The result is a system that supports conceptual AI learning, not under ideal conditions, but in the environments where teaching and learning most often take place

—

The Turbo Pascal coding line

[Ed: Nam] I often think back to my first encounters with learning computer science, Turbo Pascal back in 1999. We scribbled our line of code onto a piece of paper. It wasn’t a sleek, one-to-one experience with a laptop. It was a line of fifteen students stretching out the door of a hot classroom, all waiting for a single compiler.

There was a beautiful efficiency to that line. You had one shot to run your code. If it failed, you didn’t just hit “undo.” You walked all the way back to the end of the line. That walk was where the real learning happened—that was your metacognitive development time. You had exactly the length of that line to troubleshoot, to mentally simulate the logic, and to fix your syntax before your next turn.

We didn’t have “access” in the modern sense, but we did have something better: the power of friction. Today, we assume that to teach complex concepts like Artificial Intelligence or Machine Learning, we must first provide a tablet and a stable Wi-Fi connection. We’ve normalized a narrative where “learning technology” is synonymous with “digital platform”.

Bridging the STEM-divide

But complex computational thinking is not a “STEM-only” property. It belongs to the history teacher, the ESL instructor, and the student in a resource-constrained village as much as it belongs to the computer scientist.

This is the inspiration behind our AIED Instructional Games project. We are designing an instructional game that moves away from the “service delivery” model of education—where students are just clients consuming content—and moves toward a model of epistemic agency.

Using the “unplugged” model, we can teach:

• Classification with Decision Trees: Understanding how AI recognizes patterns and makes independent decisions without a single line of digital code.

• Reinforcement Learning: Using a simple “paper” computer and a reward system to show how machines learn through “punishment” and “reinforcement”.

• The Logic of Neural Networks: Recreating the abstraction layers of image recognition using sketches and human roles to identify complex features.

The Power of Play in Resource-Constrained Spaces

In our work with Professor. Seiji Isotani, we’ve conducted experiments that show how easily people gravitate toward traditional models when given digital resources—like the “paper airplane” default when asked to make a flying object. But when you remove the search engine and the prescribed rules, curiosity can and will take over.

Students in under-resourced contexts already possess the most powerful pedagogical tool ever invented: the power of play. They have an unreal, endless curiosity that doesn’t need to be “unlocked” by a login screen.

Our project focuses on making these abstract concepts perceptually available through physical materials—printed cards, diagrams, and collaborative dialogue. We aren’t just delivering information; we are creating the “critical exceptions” that drive theory revision and conceptual change.

The “Monday Morning” Call: Iterate Toward the Edge

As Justin Reich argues in his book Iterate, innovation in schools isn’t a top-down overhaul; it’s a bottom-up cycle of experiment and peer learning. Reich advocates for “short design spirals”—modest, manageable steps that can be repeatedly refined.

This is the “Turbo Pascal coding line” for the 21st century. By choosing low-cost, unplugged tools, we create a space to:

1. Plan: Design a small change to a lesson that strips away the high-tech barrier.

2. Experiment: Try the logic in a live, interactive environment.

3. Experience: Reflect on the results with students and peers to decide what to do next.

We don’t need to wait for the infrastructure to catch up to the pedagogy. We can perhaps start building classrooms today where students ask fiercely difficult questions and challenge the very rules of the task. Because in the end, if we treat education as a service to be consumed, we strip away its heart. If we treat it as a shared project of co-creation, any surface becomes a whiteboard, and any classroom becomes a lab.

Together, they bridge the gap between high-level policy, early-stage educational research, and the practical application of AI in learning labs. Below is a summary of our collaborative Op-Ed, exploring whether these frameworks are truly moving the needle or simply retrofitting the past.

Intro: The “North Star” and the Foundation of Sand

Nam: Welcome to this conversation: There is an undeniable urgency right now, with frameworks coming from every direction—TeachAI, the U.S. Department of Labor, and Anthropic’s AI Fluency Index. Li, I’ll start with you: Where is the contention point in all of this?

Li: I think AI literacy is both the common ground and a point of contention. Although all of these frameworks—from TeachAI to the Department of Labor—refer to AI literacy, they do not provide a single, clear definition of it. This may be because literacy itself is inherently multi-layered and highly contextualized. As I mentioned earlier, each framework serves different target users and contexts, which likely leads them to emphasize different dimensions of literacy.

Nam: Looking at these frameworks, are they new? Are they novel?

Li: Honestly, I do not think the framework introduces particularly surprising ideas; rather, it appears to be becoming more comprehensive and mature. When reviewing the “Highlights from Various AI Literacy Frameworks” on TeachAI, I could clearly see the expanding scope from AI4K12 to the AILit framework. Additionally, I notice that all three frameworks emphasize the importance of adapting to rapidly evolving AI technologies, focusing on cultivating fundamental skills and knowledge that are durable and unlikely to be replaced as technologies change.

Discipline vs. Interdisciplinary Integration

Nam: There is a big debate here. Should we frame AI as a sole field or cross-disciplinary? Should we have a separate AI class or integrate AI into different classes?

Joe: It’s a complex balance. Russell and Norvig argue that AI is both a discipline and an interdisciplinary project. It has its own methods, theories, and research communities, but it is fundamentally built on—and continues to rely on—other fields. I think a standalone AI course can definitely build foundations, while cross-disciplinary integration ensures AI literacy is not isolated from the contexts in which it operates.

Li: I suppose positionality also shapes how we understand AI, whether as a distinct discipline or as an interdisciplinary endeavor. However, for the general public and in K–12 education, it may be more appropriate to frame AI as an interdisciplinary project. This reminds me that there are many different “literacy” frameworks now, such as data literacy, media literacy, and digital literacy. We often discuss them conceptually, but in practice, at least in China, where I am from, there is rarely a standalone K-12 subject dedicated solely to these literacies. This may be because it is difficult to introduce a new subject into the existing educational system.

The Arts and the Danger of Superficial Embedding

Nam: Li, You’ve raised a concern that AI literacy contains domain-specific knowledge that cannot be fully addressed when it is simply embedded across subjects. You used your background in the arts as a prime example.

Li: Yes, my question was prompted by the paper “AI Literacy Scenarios for Every Subject.” Although it provides example lesson plans for arts classes, as a former arts educator, I do not think they go far enough. AI has fundamentally transformed the arts field itself, reshaping what we understand as artistic practice and even what counts as “the arts.” From my observation, many people use generative AI tools to produce images without critically engaging with artistic principles or creative intentionality. Therefore, I would argue that it is equally important to guide students to create with AI, not only to reflect on broader conceptual issues such as authorship (which are also important), but also to develop creative decision-making and disciplinary awareness within AI-mediated artistic practice.

Cina: This connects to how AI is reshaping the socially organized epistemic practices in fields like the arts. It illustrates how AI literacy contains domain-specific knowledge that cannot be fully addressed when it is simply embedded across subjects. This may sound arts-centered, but other disciplines may also be undergoing similar domain-specific transformations. Given this, integrating AI literacy across all subjects in K–12 education could become fragmented or inconsistent. A shared responsibility and clearer cross-disciplinary consensus might be a more practical and sustainable approach.

Discipline-Relative Literacy (Natural Science vs. Social Sciences)

Nam: Joe, you’ve picked up on this “arts-only” problem. Do you see this as a universal literacy or something relative to each discipline?

Joe: My follow-up question would be: Are AI literacy frameworks trying to create a universal literacy, or should they be understood as discipline-relative literacies shaped by how knowledge is produced in each field? I agree with Li’s argument that this should not be seen as an arts-only problem. Similar patterns appear in other disciplines when AI is introduced without attention to how knowledge is produced in that field. In science, this might mean accepting AI-generated explanations without questioning uncertainty. In social sciences, it could involve reproducing biased categories without critical reflection. These examples make me think that AI literacy cannot be separated from disciplinary literacy. If AI is simply “integrated across subjects” without shared principles, AI use risks becoming inconsistent and tool-driven. Meaningful AI literacy must help students develop judgment and responsibility within the specific epistemic practices of each discipline.

The Translation Gap and the Burden of Implementation

Nam: Let’s talk about the classroom. How do educators and learners interpret and enact AI literacy in practice, and what gaps emerge between policy frameworks and lived classroom experience?

Li: I can see the difficulty of translating such frameworks into sustained, real-world practice. One possible pathway might be through textbooks. Many of these frameworks can appear abstract to teachers, and even when lesson plans are provided, they may feel fragmented. In China, most public school teachers design their courses primarily around officially approved textbooks, and curricular changes often follow textbook revisions. Therefore, a more practical strategy, especially at the policy level, might be to translate these frameworks into textbook content, potentially embedding them across different subjects. This could provide a more systematic and sustainable way to implement AI literacy in K-12 education.

Cina: Since there is no consensus on what AI literacy actually means and how it should look like in practice, it’s difficult to provide specific and clear guidelines for AI use. When we can’t align on a shared definition, how do we design curriculum, assess outcomes, or even choose the right tools for learners? Between policy framework and lived experiences, there’re also domain-specific challenges and general guidelines can’t be anticipated. We don’t want to force AI into spaces where it doesn’t naturally fit. Instead, we need to evaluate AI’s fitness within the epistemic practices and communities of practice that define each discipline. How knowledge is produced, validated, and shared looks very different in an art studio than in an engineering lab, even though they might involve the same processes of prototyping and iterative design.

Conclusion: The Need for Translation Mechanisms

Joe: Across the AI literacy frameworks discussed, there is strong conceptual agreement on goals such as critical thinking, ethical awareness, human agency, and responsible AI use. However, these high-level policy frameworks often struggle to translate into sustained classroom practice. For many educators, AI literacy frameworks remain abstract ideals. Even when lesson scenarios are provided, they can feel misaligned with institutional realities such as assessment systems, time constraints, and textbook-driven teaching.

This gap is especially visible in non-Western contexts, where curricular change usually follows officially approved textbooks rather than teacher-designed modules. In such systems, I would argue that expecting individual teachers to implement broad AI literacy frameworks places an unrealistic burden on them. As a result, AI literacy risks becoming unevenly implemented and superficial instead of deeper engagement with disciplinary knowledge and judgment. Seen this way, the main challenge of AI literacy is the lack of translation mechanisms like curricula, textbooks, assessments, and teacher support that connect policy visions to lived pedagogical practice. Addressing this gap is essential if AI literacy is to become educationally meaningful.

Nam: Thank you all for these comprehensive insights. We’ve gone beyond the framework to look at the actual reality of the classroom. Looking forward to our future conversations!

Cina, Joe, Li, and Nam

—

References

• Anthropic. (2024). The AI Fluency Index: Framework & Foundations.

• Russell, S. J., & Norvig, P. (2023). Artificial Intelligence: A Modern Approach.

• Scribner, S. (1984). Literacy in three metaphors. American Journal of Education, 93(1), 6-21.

• TeachAI. (2024). AI Literacy Framework.

• U.S. Department of Labor. (2024). AI Literacy Framework: Foundational Content Areas and Delivery Principles.

Walk into our lab, and the first thing you’ll see isn’t a high-powered GPU or a stack of consent forms. It’s a series of dot-voting boards. We ask the “burning” questions: Who wins the Super Bowl? When will the Philly snow finally melt? Is AI a parrot, a genie, or a spork?.

To some, this might look like a distraction from “real” science. But these boards are actually the front lines of our research. They represent a shift from a transactional model—where we extract data from a subject—to a relational model, where we build common ground with our community.

By the time someone sits down for a formal interview or a strategizing tasks, we’ve already started a conversation. We’ve sparked their curiosity and, more importantly, we’ve shown them that their everyday observations are the evidence we care about.

The Generative Tension: Rigor vs. Relationship

In academia, there is an undeniable obsession with the “rigor” of the process: the meticulous tracking of citations, the pressure for high-impact publications, and the clinical isolation of variables. This creates a tension. On one hand, we need precision. On the other, we are studying humans—beings who generate inferences, carry anxieties, and hold their own mental models of the world.

As Charlotte Vaughn notes, many researchers study people as if they were “rocks,” focusing on the data while overlooking the human experience. The Research as Engagement Model argues that:

• Research is inherently public engagement: Whether we are intentional about it or not, we are communicating messages about our field to every participant.

• The “Data Point” Fallacy: Viewing participants solely as data points misses the opportunity for meaningful interaction that can actually improve our science.

• The Power of Inferences: Participants don’t just follow instructions; they generate inferences about our goals, sometimes mistakenly believing we are “grammar police” or “bias checkers”

Why This Matters Specifically for AI Research

In the world of AI-mediated communication, this tension is even more acute. When we engage with human subjects in AI research, we aren’t just looking at inferences in a model; we are investigating how humans understand a technology that is often a “black box” to them.

• Replenishing the Reservoir of Goodwill: Public confidence in scientists has declined recently, and trust in institutions is at a low point. In AI research, where data privacy and algorithmic bias are top-of-mind, every respectful interaction is an opportunity to “replenish” a reservoir of public trust.

• Defining Knowledge: What counts as “knowledge” in AI? Is it the raw output of a Large Language Model, or is it the human’s ability to use that model as a “spork” (a versatile, everyday tool)? By engaging in everyday inquiries, we better understand the “funds of knowledge” people bring to their interactions with technology.

• Inclusive Innovation: If we collect demographic data in an exclusionary way—such as forced-choice gender options—we don’t just get poor data; we communicate that certain identities are “outliers” in the eyes of science

The Debriefing: From Boilerplate to Breakthrough

The ultimate “rehumanizing” tool in our toolkit is the debriefing. Usually, this is a legalistic boilerplate—something we write and participants read because we are both “obligated” to.

Under the REM, the debriefing becomes a proactive tool for sparking curiosity. It’s our chance to answer the “So what?” and connect the incremental findings of a single study to the big-picture questions of how the mind works.

When we ask people to predict the snow melt or vote on an AI analogy, we aren’t just having fun. We are practicing Everyday Science—the belief that the most rigorous research in the world is only as good as the human connection it builds.

Reference

Vaughn, C. (2026). Your Research Is Public Engagement: A Case for More Intentional Science Communication in Research With Human Subjects. Journal of Experimental Psychology: General, 155(2), 261–275. https://doi.org/10.1037/xge0001897

In the current landscape, AI is frequently treated as a “retrofit.” We are attempting to run high-performance cloud compute through a 56k dial-up modem, forcing liquid, high-bandwidth tools into the legacy ports of 19th-century syllabi and static assignments. The result is often a system characterized by latency, friction, and a loss of human connection.

At the AIED Lab, we approach the classroom as a purpose-built environment. We envision our work as a Functional Stack:

• Community is our bedrock: The foundational layer where we share expertise and support collective growth.

• Teaching and Learning is our heart: The primary interface where we safeguard the relationship between teacher and student.

• Research is our roof: The outward-facing API that shares our findings with the global educational community.

But a house needs an engine room to keep the lights on. For us, that’s PAPII.

The PAPII Engine: A Study in High-Fidelity Inquiry

The Pedagogy and AI Practicum for Interdisciplinary Inquiry (PAPII) functions as a collaborative design experiment. We ground our work in the “Synapse”—a high-voltage pairing of two distinct types of expertise:

• The Map (The Disciplinary Explorer): The Subject Matter Expert who navigates the terrain of Physics, Law, or Nursing. They identify the “Stuck Points”—those specific, complex concepts where students consistently hit a wall.

• The Mirror (The Learning Scientist): The cognitive architect from PennGSE Learning Sciences & Technology Grad students who understands the science of the mind. They analyze how information is processed, where curiosity flags, and how feedback can spark a breakthrough.

When these perspectives collide, we follow a precise cycle of inquiry: we identify the “verbs” of the discipline—the core logic-gates of thinking like a professional—and co-design high-fidelity tools to Amplify the teacher’s reach. This might manifest as a digital persona for ethics debates or a “digital twin” of the universe for cosmology labs, allowing students to iterate and fail safely.

• Finding the Friction: We don’t start with “What can the bot do?” We start with “What is the hardest thing for your students to grasp?”. We look for the “verbs” of the discipline—the doing, the thinking, and the questioning.

• The “High-Fidelity” Design: Once we find that stuck point, we co-design a tool to help. We use AI to Amplify the teacher’s reach, not replace it. Maybe it’s a way for a student to “talk” to a dataset, or a simulation that lets them fail safely before they try it for real.

• The Classroom Test: We take these designs back into the real world. We test them in actual Penn classes to see if they actually help a student see the concept better.

The Practicum: An Interdisciplinary Journey

PAPII is architected as an iterative “arch” that moves from theoretical inquiry to production-ready implementation:

Stage 1: Problematization & Needs Analysis The journey begins with a deep-dive into the epistemological logic of a field. We focus on identifying exactly where learning bottlenecks occur, defining the friction between complex content and student understanding.

Stage 2: The Design Loop (Research & Prototyping) Pairs move into integrated cycles of research and co-design. This phase produces a Minimum Viable Product (MVP)—a tool grounded in learning science and domain-specific pedagogy designed to increase intellectual throughput.

Stage 3: The Pilot (Testing & Scalability) These are production-ready stress tests. We implement our designs in real Penn classrooms to evaluate their pedagogical sustainability and their actual impact on the student experience.

Stage 4: Evaluation & Refinement In the live environment, we perform high-stakes analysis of the tool’s performance. We collect data from instructors and students to refine the tool’s logic, ensuring the intervention addresses the core “Stuck Point” with surgical precision.

Stage 5: Global Dissemination The final output is a contribution to a growing body of interdisciplinary knowledge. Breakthroughs are documented in Reflective Briefs, ensuring that a design successful in the Medical School can be ported and adapted by the Department of Sociology.

The Vision Realized: A Living Repository of Wisdom

Our North Star at the AIED Lab is a foundational principle: Pedagogy is the operating system; AI is the processor. When we rebuild the system from the “Stuck Points” up, we move beyond retrofitting into true innovation. We are building a high-fidelity environment where the human mind is free to do what it does best: wonder, critique, and create.

The Incoming Penn AI Pedagogy Repository represents our commitment to this open-source future. It is a place where we prove that pairing the rigor of a discipline with the science of the mind grants teachers and students the superpowers required to solve the most complex problems of the next century.

The engine room is live. The future of intelligence is being architected here. Join us.

Subscribe now