How to Design an AI Curriculum? A Replicable Three-Layer Framework Distilled from Global Experience ​

We · October 2026

This article is the first in the "AI Curriculum Design Guide" series. This series doesn't discuss "whether AI education is important" — the answer is obvious. We focus on "how": how to design curriculum, choose environments, select models, and build training labs.

China has 621 universities offering undergraduate AI majors, nearly one-fifth of all undergraduate institutions. In 2024 alone, 406 new AI-related degree programs were added nationwide. This growth rate is unprecedented.

But we noticed a problem: degree programs are increasing fast, but is curriculum quality keeping up?

Research shows 88.12% of schools report insufficient AI faculty, with 43.88% severely insufficient; 54.13% of teachers lack confidence in their AI teaching ability. Reports also indicate many AI courses suffer from "heavy on concepts, light on practice; heavy on theory, light on case studies" — a phenomenon some students jokingly call "AI 101 Syndrome."

Where's the problem? We believe it's not entirely about faculty or hardware — the fundamental issue is that many schools are "adding an AI course to a traditional curriculum" rather than "redesigning the curriculum for AI."

Generative AI brings not just new knowledge points but a whole new tech stack, new development methods, and new engineering practices. If you just patch the old system, students discover after graduating that what they learned is completely different from what the industry uses.

In this article, we distill the underlying methodology of AI curriculum design from global cases including MIT, CMU, Stanford, China's "101 Plan," and UNESCO's competency framework. Not listing curricula, but examining "how these schools think about curriculum design" — their design logic, principles, and structures. Finally, we attempt to provide a replicable three-layer curriculum framework.

Five Curriculum Design Methodologies Worldwide ​

Different universities' AI curricula look vastly different, but the design logic follows discernible patterns. We have compiled five representative methodologies, each with lessons worth learning.

Methodology One: MIT's "Objective—Outcome—Assessment" Alignment ​

MIT's classic AI course 6.034 has a very clear design logic: define objectives first, then work backward to what competencies students should achieve, and finally design assessment methods.

Specifically, the course explicitly lists what students should be able to do upon completion: explain AI's basic knowledge representation, problem-solving, and learning methods; evaluate their applicability, advantages, and limitations when solving specific engineering problems; assemble intelligent systems; understand the role of knowledge representation, problem-solving, and learning in intelligent systems engineering.

The core of this methodology is "beginning with the end in mind" — not "teaching whatever content I have" but "what abilities should students have upon graduation, and work backward to design the curriculum." Every teaching element corresponds to an ability objective; every assessment tests a capability dimension.

We believe this is the starting point for all curriculum design — before discussing "what to teach," first answer "what should students be able to do after learning."

Methodology Two: CMU's "Core + Project" Dual-Wheel Structure ​

Carnegie Mellon University's Master of Science in Artificial Intelligence and Innovation (MSAII) uses a "core courses + knowledge requirements + electives + Capstone project" structure. Of the 195 total units, core courses are 84 units, knowledge requirements 72 units, and electives at least 36 units.

The highlight of this structure is that projects are not ornaments but core. The Capstone is mandatory — students must team up to complete a real AI system development project. This means students don't "do projects" after finishing all courses; they continuously apply what they learn in projects throughout the learning process.

CMU's design logic: Courses teach knowledge, projects build capability — both are indispensable. Courses without projects are armchair strategy; projects without course support are castles in the air.

Methodology Three: Stanford's "Integration—Reconstruction" Logic ​

Stanford's computer science curriculum revision approach is interesting — not adding new courses but "integrating existing course fragments, eliminating redundancy, and forming a new core course structure." They divided courses into "Systems Core" (3 courses) and "Theory Core" (3 courses), and clarified that certain courses are "funnels" (entry courses) rather than core courses.

The lesson from this methodology: AI curriculum design is not addition but integration. The mistake many schools make is "adding a course for whatever is hot" — when LLMs are hot, add an LLM course; when Agents are hot, add an Agent course; when RAG is hot, add a RAG course. The result is more and more courses, but students learn a pile of scattered knowledge points without a system.

Stanford's approach is the opposite: first clarify what core abilities are, then integrate relevant content into core courses. Generative AI content isn't added as a standalone "LLM course" but permeates related core courses.

Methodology Four: China's "101 Plan" Three-Layer Curriculum System ​

The Ministry of Education's AI "101 Plan" takes another route — systematically building a multi-level curriculum system. With "foundational, cutting-edge, interdisciplinary, integrative, systematic, practical" as construction principles, it built a three-layer structure of 12 core courses, 15 reference courses, and 2 comprehensive experimental courses.

Core courses include "Introduction to AI," "Data Science Fundamentals," "Computer Vision," "Generative AI," etc. Reference courses are extensions and deepening of core courses. Comprehensive experimental courses are practical modules, integrating knowledge from multiple courses into projects.

The characteristic of this system is high standardization and complete structure — essentially providing a "national standard for AI major courses" for universities nationwide. For schools with insufficient faculty and experience, this is a valuable reference framework.

Methodology Five: UNESCO's Competency Framework Model ​

UNESCO takes a more fundamental route — starting not from courses but from "competencies."

UNESCO's Student AI Competency Framework divides competencies into five core areas: human-centered mindset, AI ethics, AI foundations and applications, AI pedagogy, and AI professional development. Each area has three progression levels (understanding → application → creation).

The teacher AI competency framework has a similar structure — not "what technologies teachers must know" but "what AI-related capabilities teachers should have in teaching."

The value of this methodology is that it provides a "competency coordinate system" — when designing curriculum, schools can first reference the competency framework to define what level students should achieve, then design corresponding course content. This is more fundamental than directly "listing course inventories."

Commonalities Among the Five Methodologies ​

These five methodologies come from different countries and types of institutions, with different entry points — some start from objectives, some from projects, some from competencies. But we found several deep commonalities:

These commonalities are actually the underlying principles of AI curriculum design. Whether you're a comprehensive university, vocational institution, or training organization, whether you teach computer science or design, these principles are universal.

A Replicable Three-Layer Curriculum Framework ​

Based on the commonalities of these five methodologies, we attempt to distill a more universal, more implementable AI curriculum design framework — a three-layer structure:

Layer One: Core Foundation ​

This is the basic AI knowledge all students should master, regardless of major.

Core content:

• Basic AI concepts and development history
• Principles and capability boundaries of generative AI (no need to go deep into math; understand "what it can and cannot do")
• AI ethics and social responsibility (bias, privacy, copyright, deepfakes, etc.)
• Prompt engineering basics (the ability to communicate effectively with AI)

Characteristics:

• Emphasizes literacy over technology
• For all majors (not just computer science)
• Goal is building AI literacy, not training AI engineers

The National University of Singapore announced that starting 2026, all undergraduates must take AI courses, moving in this direction. China's Ministry of Education and four other departments' "AI + Education Action Plan" also calls for promoting AI as a university-wide general foundational course.

We believe this layer has the highest ROI — it doesn't need complex lab environments; online tools can teach it, but it establishes AI awareness for all majors, laying the foundation for subsequent professional integration.

Layer Two: Practice & Methods ​

This layer is about "how to use AI to do things," for majors that need deeper AI tool usage.

Core content (by direction):

Note: Specific models iterate quickly; the options listed are representative at the current stage. Curriculum should focus on principles and methods, with model selection updated periodically — tools and methodologies are stable, models are fluid.

Characteristics:

• Emphasizes practice over theory (no need to derive algorithms; understand principles + know how to use)
• Centered on tools and workflows
• Requires practice environment support (local deployment or cloud environment)
• Different majors choose different directions; full coverage not required

This layer is the weakest link at many schools today. Theory courses are easy to offer — find a teacher to read slides. But practice courses need environments, tools, and teachers who can actually use them. Many schools' AI courses stay at the first layer (concepts) and can't keep up at the second layer (hands-on).

We want to emphasize: the second layer is where real differentiation happens. Students who can build things after learning and students who only know concepts are at completely different levels. And the biggest bottleneck at the second layer is not faculty (which can be trained) but practice environments — if tools can't be installed and models can't run, even the best curriculum design can't be implemented.

Layer Three: Cross-Disciplinary Application ​

This layer is the deep integration of "AI + Major," for upper-grade students and graduate students.

Core content (examples by professional field):

Characteristics:

• Project-centered, not course-centered
• Requires comprehensive application of Layer 1 and Layer 2 knowledge
• Integrates with real industry needs
• Typically presented as Capstone projects, graduation theses, internships

CMU's Capstone project system and Stanford's project-based learning are essentially efforts at this layer. China's "101 Plan" 2 comprehensive experimental courses have similar positioning.

Three Universal Problems and Solutions ​

Having a framework doesn't mean it can be implemented well. We have observed that AI curriculum design worldwide faces three universal problems:

Problem One: Update Speed Can't Keep Up with Technology Iteration ​

AI technology iterates too fast — large models have a major version every half year, and new tools emerge constantly. But university curriculum updates slowly; reports indicate some schools' AI courses lag industry iteration by 15+ months.

Solutions:

• Keep the core foundation stable (basic concepts and principles change slowly)
• Use "modular tool" design for the practice methods layer — each tool is a module; when a tool updates, just swap the module without changing the entire course
• Build a "rapid curriculum update mechanism," such as refreshing practice cases and tool versions every semester
• Introduce real enterprise projects so students encounter the latest industry practices

Problem Two: Disconnect Between Theory and Practice ​

This is the most frequently mentioned problem — too much concept teaching, too little hands-on practice. Students can recite Transformer's attention mechanism formula but don't know how to run a local model with Ollama.

Solutions:

• Adopt "learning by doing" instructional design — let students run a project first, then return to principles
• Pair every theory course with corresponding practice; practice hours should be no less than 50%
• Front-load practice environments — in the first week, get every student's computer running their first AI application
• Use projects to connect knowledge points, not knowledge points to piece together projects

Problem Three: Insufficient Faculty Capacity ​

88.12% of schools report insufficient AI faculty, and 54.13% of teachers lack confidence in AI teaching. This is a global problem — UNESCO data shows that as of 2022, only 7 countries had developed teacher AI frameworks or plans.

Solutions:

• Don't wait until all teachers become AI experts before offering courses; "teach while learning"
• Build teacher AI capacity development mechanisms (specialized training, teaching workshops, enterprise secondments)
• Adopt a "teacher + technical assistant" model — teachers handle instructional design and domain knowledge; technical assistants handle environment setup and tool support
• Share curriculum resources — good practice cases, project templates, and environment configurations can be shared across schools

Curriculum Design Is Just the First Step; Environment Implementation Is Key ​

We want to emphasize an easily overlooked fact: no matter how perfect the curriculum design, without runnable practice environments, students still can't learn.

Layer 1 (Core Foundation) is manageable — online AI services can teach it, no local environment needed.

But Layers 2 and 3 don't work that way. RAG systems need local LLMs + vector databases; image generation needs ComfyUI + tens of GB of model files; video generation needs even higher hardware... none of these are achievable by "opening a webpage."

The reality is:

• Many schools spend heavily on GPUs but can't set up or maintain software environments
• Teachers spend enormous time helping students troubleshoot environment issues rather than teaching domain knowledge
• Students give up before even starting because they can't install tools or run models

The core of these problems is not "whether there's hardware" but "how to rapidly deploy, uniformly manage, and continuously update software environments." This is precisely the capability many schools lack most.

Some solutions are already visible in the market — some tools simplify local running of large models, some platforms turn image generation into node-based interfaces, and some all-in-one AI application deployment tools support rapid deployment of common model versions (original, quantized, accelerated). The value of these tools: schools don't need to research how to quantize each model or configure dependencies; they just choose appropriate versions based on curriculum needs and can quickly set up a usable teaching environment.

AI curriculum design is a big topic; there's no one-size-fits-all standard answer. But we believe one criterion is universal:

A good AI course is not about how many concepts students remember, but whether they can actually build things with AI after learning.

From this criterion, the three-layer framework (Core Foundation → Practice & Methods → Cross-Disciplinary Application) is just a skeleton. The real substance is that each layer has runnable practice environments, operable project cases, and verifiable competency standards. With the framework in place and environments keeping up, AI education can truly move from "concept popularization" to "capability building."