Stretching the student who is ahead

The student the lesson was not built for

Every class has one or two who finish the work in the time it takes others to read the question. The earlier lessons gave you extension tasks (the harder question about the same shirt) and those keep her in the conversation. But an extension task is thirty minutes of work, and she has a term of them to get through, and the honest truth is that a strong student in a large class spends most of her time waiting.

The chatbot removes the ceiling. She can ask it anything, follow any thread, go as far as she likes. For the curious student this is a genuine gift, and teachers should say so.

The trap

She is also the student most likely to absorb its errors. Three reasons. She is confident, so she does not check. She is going beyond what you teach, into topics where you are less able to catch a mistake and where the model is more likely to make one: the further from the syllabus, the further down the reliability gradient. And you are busy with the twenty-eight students who need you more, so nobody is looking at what she is learning.

The check question is the result: a strong student with fluent, detailed, wrong ideas, delivered with the model’s confidence and her own.

Make her the checker

The stretch that fits this student is not more content. It is the skill she is missing and the model cannot give her: telling right from wrong in confident material.

Give her the task from the check question. “You said three things about quantum mechanics that I am not sure about. Find out, from sources that are not a chatbot, which are right, which are wrong, and which are unsettled. Present it to the class in five minutes next week.”

This is harder than anything on the syllabus. It requires finding real sources, reading them, comparing them with what the model said, and (the difficult part) holding an unsettled question as unsettled. It is exactly the skill the whole course has been about, aimed at the student most able to learn it and most at risk of not.

Other stretches that use the tool well

• ”Find the error.” Give her a generated explanation of a syllabus topic that contains one deliberate mistake, and the model’s confident tone. She finds it. Then she writes one for the class.
• The Socratic prompt. A card: “Do not explain. Ask me questions until I can explain it myself.” Used on a topic just beyond the syllabus, this makes the model into a tutor that withholds answers: the design the evidence lesson found does not harm and can help.
• The evaluation task. She writes twenty questions on the topic, asks the model, marks it, and reports the accuracy and the pattern of errors. She learns the topic by writing the questions and learns the tool by marking it.
• Teaching it back. She uses the model to prepare a five-minute explanation of a topic for the class, and is told the class will ask questions, so she had better understand it rather than recite it.
• The project lesson. The bigram model from the teaching-AI block is a two-week stretch for a strong student in any subject, and it produces the understanding that makes every other use of the tool safer.

Keep her in the room

A student who is always ahead can drift into a private world with the machine. The extension tasks that keep her in the same conversation as the class (the harder question about the shirt) matter more, not less, now that she has an unbounded alternative. Two-thirds of her stretch should be about what the class is doing; one-third can be her own thread.

The honest limit

The checker task works because there are sources beyond the model to check against: a textbook, a library, a teacher who knows the field, the internet read carefully. For a student in a school with no library and one textbook, “check it against real sources” may mean checking it against the model’s own web search, which is better than nothing and not independent. Be honest with her about that: the skill is the same, the sources are thinner, and knowing that the sources are thin is itself part of the skill.