To ask a question, one must know enough to know what is not known

Every teacher knows the silence after “any questions?” at the end of a lesson. Thirty students, nobody raises a hand, and it is very tempting to read that as thirty students who understood. Some of them did. Others are so far behind that they couldn’t say what they don’t understand, so they say nothing, and the silence sounds exactly the same from the front of the room.

In 1979 the cognitive scientists Naomi Miyake and Donald Norman put this into a study and gave it the title this post borrows. They had people learn a text editor from a manual and counted the questions they asked along the way. Beginners asked the most questions when the material was easy. People with more background asked the most when it was hard. The number of questions had little to do with curiosity and a lot to do with whether the material was just beyond what the person already knew. Too far beyond it, and the questions stopped. To ask a good question you need a rough map of the subject, enough to see where the hole in it is, and the beginner doesn’t have the map yet. That is what they came for.

Now look at the ChatGPT window. It is an empty box with the words “Ask anything” in it. It is “any questions?” turned into a product, asked to every student at every hour of the day, and it does nothing until the student answers it.

For a student who knows enough to ask, it is extraordinary. They can ask why the electron transport chain needs oxygen at the end, push back when the explanation doesn’t add up, and ask for the same thing a third way at eleven at night without anyone sighing. A student who is lost in the chapter can’t ask any of that, because they don’t know yet that there is a chain, or an end. There is one question they can always ask without knowing anything, though, and it is “what’s the answer?”

That is roughly what happened when researchers from Wharton gave nearly a thousand high school math students in Turkey GPT-4 for practice. With the chatbot, the students solved 48 percent more practice problems correctly. Then it was taken away for the exam, and they scored 17 percent lower than classmates who had never used it. A second group got a version that the researchers had set up to give hints and never the answer. Their practice scores rose even more, and their exam scores did not drop. When the researchers read the logs, the first group had mostly been copying.

The cost of that shortcut isn’t the same for every kid. Twenty years ago, Ryan Baker and his colleagues watched middle school students use a math tutoring program in which the last hint was always the answer, and some students learned to click straight through to it. The ones who did it learned less than students who went off task in any other way. A later study found that the clicking did little harm on steps the students already understood. It hurt on the steps they didn’t. So a tool that waits for questions works best for the kids who already know a fair amount, and worst for the kids who need it most, and the gap between them gets wider.

OpenAI has added a study mode to ChatGPT, which answers with hints and questions instead of the answer. It is a better reply, and the box still waits for the student to start. A lost student types “explain photosynthesis”, gets a well-organised lecture with nowhere to put it, and is back where they started, with a toggle that turns study mode off when the essay is due at midnight.

Good teachers stopped relying on “any questions?” a long time ago, because they know the silence doesn’t tell them anything. They ask the questions themselves. “Why does the plant need light here?” “What would happen if this cell had no mitochondria?” A student can’t tell you where the hole in their understanding is, but they will show you the moment they try to answer. Psychologists have a name for the reason: the illusion of explanatory depth. People are sure they understand how something works until someone asks them to explain it step by step, and then they find the gap. The question finds what the student couldn’t have known to ask about.

There is a second problem, and it comes from what a chat assistant is built for. These models are trained partly on people’s ratings of their answers, and people give high ratings to answers that solve their problem quickly and completely. So the assistant gets very good at making the person in front of it feel done. For most tasks that is exactly right: if you ask for a recipe or an email, feeling done and being done are the same thing.

In learning they come apart. The part that builds memory is the part that feels like effort, pulling something out of your own head, getting it wrong, trying again. Psychologists call these desirable difficulties, and a tool trained to remove difficulty removes them along with everything else. A well-written answer feels clear, and the student stops when it feels clear, which is long before they could explain it themselves. In a chat, the student decides when the lesson is over, and the tool is designed to bring that moment as early as possible.

That is the part I built Lexie to do. Lexie doesn’t wait for a question. You give it your notes or a chapter, and it asks you the questions, one at a time, in an order that builds the map, and it waits for your answer. When the answer shows a gap, Lexie goes after that gap, walking you through it in a step or two instead of handing over the answer. The student doesn’t have to know what they don’t know. The questions find it.

So before you blame the kid, or the AI, look at the empty box they were typing into.