Reading thirty answers at once

The pile of scripts as data

After a test you have thirty answers to question 7, and somewhere in them is the reason a third of the class got it wrong. Reading thirty answers for a pattern is slow, and by script twenty you have stopped noticing. It is the kind of reading a model is good at: it does not tire and it can hold thirty short answers in one view.

The output (three or four groups, each an error with the students’ own words as examples) is a reteaching plan. This lesson is how to get it and how not to be misled by it.

Getting the answers in

Typed answers from an online quiz paste straight in. Handwritten answers need transcription; the next lesson covers the phone-camera route. For a first attempt, choose a question with short answers (a line or two) and type them yourself. Thirty lines is ten minutes and it is worth doing once by hand to see what the analysis gives you.

Replace names with numbers. The model does not need to know who wrote what, and the rules about student data apply.

Here are 30 answers from Class 10 to: "Explain why a ball thrown upward slows down." The correct idea is that gravity acts downward throughout, decelerating the ball.
Group the answers by the misunderstanding they show. For each group: name the misunderstanding in one sentence, list the answer numbers, and quote one answer as the clearest example. Then say which group is largest.
Do not correct the answers. Do not invent a group that has fewer than three answers.

What comes back, and what it is worth

You will get something like: nine answers say the ball “runs out of force”; seven say gravity “gets stronger” as it rises; five are right but incomplete; and so on. The first group is the classic impetus misconception and it is worth a whole starter. The second is a different error that needs a different fix. Without the grouping you would have taught one lesson to both.

That is the value, and it is real. Now the caution.

It will find patterns in anything

A model asked for groups produces groups. Give it six answers and it will return three categories with percentages, and the categories will look as neat as the ones from sixty. Give it thirty answers that are all the same and it will find subtle distinctions. It has no threshold below which it says “there is not enough here to say”. The “no group under three answers” line in the prompt helps; the real defence is the check question: read the answers it put in each group and see whether they belong.

That reading takes five minutes and it is where you catch the group that is two misconceptions merged, or the “correct” group that contains a wrong answer with the right vocabulary. Do it before you plan a lesson around the groups.

Across the whole test

For a whole paper, the spreadsheet is the tool. Marks per question per student in a grid (Google Sheets, Excel, LibreOffice Calc, all free or already on the laptop) and a row at the bottom for the class average per question. The two questions with the lowest averages are your reteaching priorities before any AI is involved. Then paste those two questions’ answers for the grouping above.

The model can also read the grid. “Here is the marks grid. Which questions did students who scored well overall still get wrong?” That finds the question that was harder than intended, or badly worded, which is a question-bank correction as much as a teaching one.

Over a term

Record the groups. After three tests you have a list of the misconceptions this class actually holds, in their words, which is the most useful planning document you own. It feeds the “misconception” line of every future prompt, and it is the thing the model could never have told you, because it came from your students.

The honest limit

Grouping short answers works well. Grouping essays does not: the model summarises rather than analyses, and the categories become vague (“weak structure”, “unclear argument”) which is no better than a tired marker’s impression. For long work, choose one question, one criterion, and ask about that: “Which of these fifteen introductions state a clear position, and quote the sentence that does it?” Narrow questions get honest answers. Broad ones get fluent ones.