This has become the question I get more than any other about learning: can I learn marketing from AI instead of taking courses?
The answer is yes, with one condition. The condition is not about which model you use. It is about how you use it. Without it you are not learning, you are reading answers.
There is real research that shows the difference, so let us start there rather than with advice.
What the research actually found
In 2025 a team at Harvard published a study in Scientific Reports on an AI tutor in an introductory physics course. The paper is published in full, and the numbers below are taken from it directly.
The design was a randomised controlled crossover trial. 194 eligible students each learned one topic in class through active learning led by experienced instructors, and another topic at home with an AI tutor called PS2 Pal. The conditions were swapped the following week, so every student experienced both, which is what makes the comparison clean.
The result: median post-test score was 4.5 for the AI group against 3.5 for the in-class group, with a pre-test median of 2.75. The effect size from linear regression was 0.63, which the authors call an underestimate because of ceiling effects, and quantile regression put it "in the range of 0.73 to 1.3 standard deviations".
It also took less time. The AI group spent a median of 49 minutes against an assumed 60 minutes of learning in class. Self-reported engagement was 4.1 out of 5 against 3.6.
Stop reading there and the conclusion is that AI beats the lecture. Wait, because the important part is in the detail.
The condition the study itself names
The tutor in that study was not an open chatbot. It was constrained in three ways, and those constraints are what made it work:
- A prompt written by the instructor for each specific question, not a general one.
- Detailed step-by-step correct answers placed inside the prompt to reduce errors.
- A platform that forced sequential progression through each part of a problem, because the prompt alone did not reliably scaffold multi-part problems.
Read that third point again. The researchers found that instructions alone were not enough, so they built a platform that made students move one step at a time. The limitations section notes that building it took several months.
So what succeeded was not "ask AI to teach you". What succeeded was a human instructor preparing the questions and the answers, then putting AI inside a structure that released one step at a time.
How to build that condition for yourself
You are not going to build a platform. But you can copy the idea, which is that the AI should be asking you, not answering you.
The method I use and recommend: do not say "explain targeting in Facebook ads to me". That produces an article you read and forget. Say something closer to this: "I run a clothes shop in Shubra with a budget of 3,000 EGP a month. Ask me one question at a time about targeting, and after each answer tell me what is wrong with my reasoning. Do not give me the full answer."
The difference between those two prompts is large. In the first you are a reader. In the second you are solving. And the study found that time on task did not correlate with post-test scores, which means what mattered was not how long you sat there but what you did while you sat there.
A second exercise that works: make it play the client. "You own a restaurant, I am going to pitch you a marketing plan, object the way a real restaurant owner would." That trains something courses do not train.
What you have to verify yourself
This is the most practical part. AI states wrong things with confidence, and marketing is full of numbers and definitions that change.
The rule: every metric definition, every platform limit, every ad policy gets checked against the platform's official documentation. Not against the chat.
I see this mistake repeatedly. Someone asks for a metric definition, gets an answer that looks right, and builds a decision on it. Official definitions change and the model was trained on older text. So make it a rule: the number, the definition and the policy get confirmed on the platform's own page, and the chat helps you understand them rather than establishing them.
In data analysis the problem is sharper still, because the definition is the whole thing, and I went through that in how to learn data analysis for marketing.
The limits of the study, which matter
I have to report what the authors said about their own result, because it changes how you should use it.
The study measured first exposure to new content, at the understanding, applying and analysing levels. Not mastery, and not originality. The authors wrote plainly that they "do not presume that structured AI tutoring will always outperform in-class active learning in all contexts".
It was one institution and one course, and the effect may depend on the instructor-crafted prompts and the quality of the videos he made. Nobody measured long-term retention, so there is no number anywhere in that paper telling you students still remembered it a month later.
One last thing stated in the limitations: the feedback was reactive, meaning the tutor responded only when a student asked. That is the gap there is no way around. Somebody looking at your work and telling you it is wrong before you thought to ask is the thing AI does not do on its own.
So use it for what, and not for what
The ordering I recommend after all of that is simple.
Use it for: the first explanation of a new topic, generating practice questions, playing the objecting client, simplifying official documentation you have already read and found confusing, and turning your own work into a draft you then edit.
Do not use it for: verifying numbers and definitions, judging your own work, deciding who your target market is, or decisions that need current knowledge of your local market.
The general order for learning with no budget, which AI speeds up rather than replaces, is in how to learn marketing without courses. And if you do not yet know where to start at all, what to learn first in marketing is the order.
My wider view on AI in marketing, away from learning, is in what is wrong with AI.
And if you conclude that what you actually want is structure, deadlines and somebody to correct you, that is a different thing and it is bought rather than improvised. The honest comparison between free and paid is one our friends at UCourses wrote in free versus paid marketing courses.
The whole thing in one sentence
AI is an excellent tutor when it is constrained to ask you one step at a time, and a poor source when you ask it for a number. The study people cite to argue it beats the lecture is the same study that says what worked was the constraint, not the model.