Can ChatGPT teach you economics?

What the evidence says

We do not have to speculate about how well these models do on economics exams. People have run the experiments and published the results — and the headline finding is how fast the picture changed between model generations.

The most famous case is the economist Bryan Caplan's. In January 2023 he gave his Fall 2022 labor economics midterm at George Mason to the original ChatGPT, graded it blind, and it earned a D. He was unimpressed enough to bet publicly that no AI would score A's on five of his six exams before 2029. Three months later, GPT-4 retook the same midterm and scored 73/100 — an A, and the fourth-highest score in the class. Caplan reported the result, in his words, with “no small dismay.” That is a D-to-A jump on the same exam, same grader, in one model generation.

The academic studies tell a similar story. Geerling, Mateer, Wooten, and Damodaran gave an early-2023 version of ChatGPT the Test of Understanding in College Economics, the standardized exam used to benchmark principles students in the United States. It answered 19 of 30 microeconomics questions and 26 of 30 macroeconomics questions correctly — the 91st and 99th percentile against students finishing a principles course. Beyond economics, a 2024 scoping review by Newton and Xiromeriti covering 114 multiple-choice question sets across higher education found that GPT-3.5-era models passed about 20 percent of exams while GPT-4 passed about 93 percent.

But the picture is not uniformly rosy, and the cracks appear in a consistent place. A 2025 study in the Bulletin of Economic Research tested ChatGPT on the question bank of a widely used open economics textbook and found it scored a high D on the multiple-choice questions even while earning a low A on short essays — with errors concentrated in higher-order questions, calculations, and graphs. Read all of this together and the honest summary is: results vary sharply by model version and question type. Recent models handle most standard introductory questions well; they remain least reliable exactly where economics exams are hardest — multi-step quantitative problems and anything built on a graph.

Where ChatGPT genuinely helps

Let me give the tool its due, because the value here is real and I would have killed for it as a student. ChatGPT is the office-hours tutor that never sleeps, never sighs, and never makes you feel stupid for asking the same question a fourth time.

• Unlimited re-explanation.A textbook explains crowding out once, in one register. A chatbot will explain it as many times, and as many ways, as you need: “Explain crowding out three different ways — once plainly, once with a metaphor, once with a small numerical example.” If the metaphor lands and the textbook paragraph never did, that is a genuine win. No human tutor on earth offers infinite patience at zero marginal cost.
• Socratic self-quizzing.Flip the roles: “Ask me five questions about price elasticity of demand, one at a time, and grade each of my answers before moving on.” Now you are retrieving instead of reading — producing answers from memory, which is the activity that actually builds recall. This is the single best study use of a chatbot, and almost nobody uses it this way.
• Unsticking yourself on one specific confusion. Every economics student hits a wall at some exact sentence — why the long-run supply curve is vertical, why marginal revenue sits below the demand curve. Paste the exact step that lost you and say so: “I follow everything until this line. Why does marginal revenue fall twice as fast as price?” Targeted questions get strikingly good answers.
• Summarizing and translating readings. Handed a dense journal article or a chapter written in 1980s-textbook prose? Asking for a plain-English summary with the three core claims is fast, low-risk, and a fine way to decide what deserves a careful read.

Where it confidently fails

Now the professor's warning list. Every item below is a failure mode I would warn my own students about, and each one is more dangerous than an honest “I don't know” would be, because the failures arrive in flawless, confident prose.

• Graphs. Large language models are text engines. They will describea monopoly diagram with deadweight loss beautifully — and then, asked to actually draw it, produce curves in the wrong place or a shaded triangle hanging off the wrong intersection. Meanwhile, AP, IB, and university economics exams grade the graph you draw by hand: axes labeled in words, curves shifted the right direction, equilibrium marked. Reading descriptions of graphs trains exactly none of that. The Bulletin of Economic Research study above found errors clustering in graph questions for a reason.
• Arithmetic slips delivered with total confidence. Ask for a midpoint elasticity calculation or a spending multiplier with a marginal propensity to consume of 0.8, and you will usually get the right setup — and, some fraction of the time, a wrong number presented with the same serene certainty as a right one. A student who cannot already do the calculation cannot tell which day they got. Newer models with built-in calculators slip less often, but “less often” is a poor standard for an exam answer.
• Hallucinated facts, figures, and citations.Ask for the unemployment rate in a given year, a real-world example with numbers, or sources for a term paper, and the model will sometimes invent them — plausible-sounding statistics and journal articles that do not exist. Treat every specific figure and every citation as unverified until you have checked it yourself.
• It answers the question you asked, not the misconception underneath it. A student asks, “does a price ceiling shift the demand curve?” ChatGPT answers the literal question — correctly! — while the real problem is that the student has confused a movement along a curve with a shift of the curve, a confusion that will sink a dozen future questions. A human teacher hears the wrong question and treats the underlying disease. A structured course is designed so that exact misconception gets caught and drilled. A chatbot, by default, politely answers the symptom.
• The fluency trap.This is the subtle one. Reading a great explanation produces a feeling of understanding, and that feeling is not learning. Decades of research on retrieval practice show that recognizing an idea when it is in front of you and recalling it from a blank page are different skills — and exams test the second. A chatbot makes the first feel effortless, which makes the illusion stronger, not weaker. I summarize the research behind this on the science page; the one-line version is that explanations you merely read are rented, and only answers you produce are owned.

The workflow that actually works

The fix is not to avoid ChatGPT. The fix is to stop asking one tool to do three jobs. Learning economics has three distinct parts — understanding ideas, building skills, and retaining both — and the chatbot is only the right tool for the first.

• Explanation on demand: the LLM. When a concept will not click, interrogate the chatbot until it does. Re-explain, analogize, work a toy example. This is its home turf.
• Skill: structured practice with feedback.Then close the chat and produce answers yourself — solve problems, draw graphs from a blank page, get told immediately whether you were right. This is the part I care enough about that I built a tool for it: Econ Academy's free practice problem pages need no signup, and the lessons use interactive draggable graphs, so shifting a demand curve is something your hands do, not something you read about. Full disclosure: I built this site, so discount accordingly — but the principle stands whatever practice source you use.
• Retention: spaced repetition.Whatever you learned this week will leak away over the next month unless something resurfaces it at widening intervals. Use any spaced-repetition system — Econ Academy's review engine schedules this automatically — so that October-you still owns what June-you learned.

A concrete weekly loop you can copy: on day one, learn the week's topic from your course or textbook, using ChatGPT to clear up anything confusing the same day. On day two, do practice problems on that topic without the chatbot open; for each miss, ask ChatGPT to explain the error, then ask it for a similar problem and solve that one cold. Midweek, do a blank-page graph drill: redraw every diagram you know from memory, then check. At week's end, take one mixed quiz covering everything so far, and let your spaced-repetition queue run daily for ten minutes. The chatbot appears three times in that loop — and never as the thing producing the answers.

Prompts worth stealing

The common thread: every good study prompt makes the model ask and you answer, never the reverse.

That last one is my favorite, because it forces the model to do what good teachers do by reflex: treat your question as a symptom and go looking for the cause.

Frequently asked questions

Can ChatGPT teach you economics?

Partially. It is excellent at explaining concepts — it can re-explain crowding out or elasticity five different ways until one clicks, instantly, for free. What it cannot do is build the skills economics exams actually grade: drawing correct graphs from a blank page, executing multi-step calculations reliably, and retrieving ideas from memory under time pressure. Use it as an explainer and a quizmaster, and do your practice somewhere that makes you produce answers yourself.

Can ChatGPT solve economics problems?

Often, but unevenly. Published evaluations show recent models handle most standard introductory multiple-choice questions well — GPT-4 earned an A on a real university economics midterm that the previous model failed. The same literature finds errors concentrate in multi-step calculations and graph-based questions, and performance varies sharply by model version and by which question bank is used. Treat any numerical answer it gives you as a draft to verify, not a verdict.

Will ChatGPT give wrong answers in economics?

Yes, and the dangerous part is that the wrong answers arrive in the same confident, fluent tone as the right ones. The typical failure modes are arithmetic slips inside otherwise-correct reasoning, invented statistics and citations, and answering the question you asked instead of the misconception underneath it. None of these come with a warning label. Always check calculations by hand and treat any specific figure or source it cites as unverified until you confirm it.

Can I use ChatGPT to study for AP Economics?

Yes — as one half of a study system. It is genuinely useful for re-explaining concepts and quizzing you on definitions and chains of reasoning. But AP Micro and AP Macro free-response questions grade hand-drawn graphs with labeled axes and explicit step-by-step logic, and reading a chatbot's prose builds neither skill. Pair it with timed practice problems and blank-page graph drills. The free AP review kits on this site are built around exactly that kind of practice. AP Macro review kit

What is the best AI tool for studying economics?

Honest answer: for explanation quality, the differences between the major frontier chatbots are smaller than the difference between using any of them well and using them badly. The binding constraint on your learning is not which model you pick — it is whether you practice retrieving and producing answers instead of just reading explanations. Pick whichever chatbot you already have access to, prompt it to quiz you rather than lecture you, and spend the saved money on nothing, because the good practice tools are free too. The 9 best apps and websites to learn economics

How should I prompt ChatGPT to learn economics?

Make it ask the questions. The default mode — you ask, it explains — feels productive but mostly produces recognition, not recall. Instead: have it quiz you one question at a time and grade your answers, have it generate fresh problems modeled on ones you just missed, paste in your own worked solution and ask it to find your first error without revealing the answer, and ask it to grade your written explanation against a rubric. Every one of those flips you from reader to producer, which is where learning happens.

Related