AI & The Teaching of Economics

Departmental Informal Discussion

Sérgio O. Parreiras

My approach:

I build together with AI, exactly the way I want it.
I don't want to wait for a developer nor settle for off-the-shelf tools.
I don't let AI create my content.
No way I am giving up what I like the most about my job.
I treat AI like a consultant.
I ask for its options with pros and cons—then I decide and AI implements.

Overview

Four areas where AI helped

Course Utility Apps—leveraging Canvas API
Scraping Tools—NPR podcasts, FRED, etc.
Overhaul of Lecture Notes—PDF → HTML
Practice Apps—ECON 101H and ECON 510

Course Utility Apps

Participation Tracker

Mac desktop app (PySide6/Qt): live session tracking with click-to-mark participation, student photos, year, major
Color-coded warnings: magenta = low participation, red = 3 consecutive absences
Canvas API upload: batches 3 sessions into one assignment, auto-creates assignment, posts grades
iPad web version: GitHub Pages, reads/writes same Dropbox JSON. Tap-to-mark for walking around classroom
TA recitation participation tracker: simplified version
Desktop/Web versions share the same JSON data layer via Dropbox sync

Scraping Tools

The Librarian

Scraping pipeline: NPR Planet Money and This American Life. All stored as JSON.
JEL classification: every episode batch-classified with JEL codes using OpenAI API
Three search modes:
- Text search: boolean query parser (AND, OR, "exact phrase")
- JEL search: topic → JEL codes → matching episodes
- RAG search: Voyage AI embeddings + ChromaDB + Claude for Q&A over transcripts
One-click Canvas push: send episode links straight into a Canvas module

Overhaul of Lecture Notes

PDF → HTML (click to swap example)

Before: Static PDF

KaTeX-rendered math—as clean as $\LaTeX$/Beamer
GSAP/ScrollTrigger animations—scroll-driven progressive disclosure
Interactive components—sliders, live updates
Real data visualizations—scraped FRED data animated in the browser

After: Interactive HTML

Untapped Returns

Practice Apps — ECON 101H & ECON 510

Why are textbooks so expensive? We choose them—but we don’t pay for them. What we’re really buying is the ancillaries: test banks, homework platforms, auto-graded exercises. That’s the moat that locks in $300 textbooks.

AI collapses the cost of building ancillaries. With AI, a single instructor can create practice apps that offer students unlimited, algorithmically generated problems with instant feedback—something no static textbook can match.

Cost Functions (101H)

Randomly generated cost table puzzles
3 difficulty levels, Firebase leaderboard

Backward Induction (510)

Randomly generated extensive-form game trees
Animated resolution with SPNE path highlighted

Not All AI Are Created Equal

Chat AI (ChatGPT, Gemini, Claude.ai)

Copy-paste loop: describe → paste → breaks → repeat
AI only sees what you feed it
You are the debugger and the file system
Hard to customize or extend

Agentic AI (Claude Code, Cursor, Copilot)

Lives in your project—reads, writes, searches, and runs code directly
Breaks → sees the error → fixes it—without you lifting a finger
Teach it how you work—reusable skills and agents that learn your workflow

Same intelligence. The difference is agency.