Most students have a favorite teacher. In the future, could that teacher be … a chatbot? The answer depends in part on whether AI succeeds in becoming the next great educational technology innovation, or joins the long list of ed-tech disappointments.

Two kinds of AI-driven education are likely to take off, and they will have very different effects. Both approaches have real promise, but neither will make everyone happy.

The first category will resemble learning platforms such as Khan Academy, Duolingo, GPT-4, and many other services. Over time, these sources will become more multimedia, quicker in response, deeper in their answers, and better at creating quizzes, exercises and other feedback. For those with a highly individualized learning style — preferring videos to text, say, or wanting lessons slower or faster — the AIs will oblige. The price will be relatively low; Khan Academy currently is free and GPT-4 costs $20 a month, and those markets will become more competitive.

For those who want it, they will be able to access a kind of universal tutor as envisioned by Neal Stephenson in his novel “The Diamond Age.” But how many people will really want to go this route? My guess is that it will be a clear minority of the population, well below 50%, whether at younger or older age groups.

To see why, consider other current options, such as watching YouTube to master scientific or mathematical knowledge (usually free), reading old editions of textbooks (cheap on Amazon), or hiring an expert tutor (more expensive). None of those approaches is currently very popular. People may resort to these services if their education requires it, but otherwise they are not taking advantage of these opportunities even when they are free. They will go to their graves not knowing about entanglement problems in quantum mechanics.

Chatbots will probably make education more fun, but for most people there is a limit to just how fun instruction can be. These kinds of AIs will appeal to the infovores, but most people still will prefer interacting with friends and family — a known and usually cheap source of fun, not unlike listening to music or daydreaming. Educational chatbots won’t be more fun and fulfilling than what you would choose to do with that time on your own.

Experience with YouTube, which most people use, is the best reason to be optimistic about the future of AI tutoring bots. Yet a lot of people use YouTube for fairly specific purposes: to answer how-to questions (reboot their iPad, fix a flat tire, etc.), to watch and listen to music, or just to hear people talking for fun (PewDiePie).

All of which is to say: These kinds of AI chatbots might replace teachers to some degree, and give more individualized feedback. But they won’t significantly increase the total time spent learning.

There is, however, another way AI education could go — and it may end up far more widespread, even if it makes some people uneasy. Imagine a chatbot programmed to be your child’s friend. It would be exactly the kind of friend your kid wants, even (you hope) the kind of friend your kid needs. Your child might talk with this chatbot for hours each day.

Over time, these chatbots would indeed teach children valuable things, including about math and science. But it would happen slowly, subtly. When I was in high school, I had two close (human) friends with whom I often talked economics. We learned a lot from each other, but we were friends first and foremost, and the conversations grew out of that. As it turns out, all three of us ended up becoming professional economists.

This could be the path the most popular and effective AI chatbots follow: the “friendship first” model. Under that scenario, an AI chatbot doesn’t have to be more fun than spending time with friends, because it is itself a kind of friend. Through a kind of osmosis, the child could grow interested in some topics raised by the AI chatbot, and the chatbot could feed the child more information and inspiration in those areas. But friendship would still come first.

Many parents will undoubtedly find it weird for their children to be so close to chatbots. Some may even forbid any relationship at all, as they do now with a human they deem a bad influence. Nevertheless, the friendship-first model is the most promising formula for AI-driven education. And the larger question may be not whether it will work, but how comfortable parents are with it.

Tyler Cowen is a Bloomberg Opinion columnist, a professor of economics at George Mason University and host of the Marginal Revolution blog.