Why compare instead of trusting one answer
Students who use AI tools usually see one answer at a time. That answer arrives fluent and well organised, which makes it easy to accept. A simple way to build responsible habits is to ask the same question twice, in a fresh session or with slightly different wording, and compare the two answers. Differences between them are a direct, visible lesson about how language models produce text: they generate plausible wording, and plausible is not the same as checked.
This method works as a short classroom module or as a homework routine. It needs only access to an AI tool the school already permits, a sheet of paper or a shared document, and one reliable reference such as a textbook. It does not require technical knowledge of how models work, and it does not ask students to decide whether AI is good or bad in general.
Step one: ask the same question twice
Choose a question with a checkable answer, for example a factual question from the current unit in science or history. Ask it once and save the answer. Then open a fresh session, or rephrase the question slightly while keeping its meaning, and ask again. Save the second answer. Note the exact wording of both questions so that the comparison is fair.
Avoid questions about private people, personal advice, or topics the school has asked students not to put into AI tools. The goal is to examine the answers, not to collect more information than the lesson needs. Teachers may prefer to generate the two answers themselves and hand them out, which keeps the activity within school policy and saves time.
Step two: line the answers up claim by claim
Break each answer into separate claims, one sentence or idea per line. Put the two lists side by side. Students are often surprised by how much of a long answer is framing, repetition, or general statements that cannot be checked. Those lines can be marked as background and set aside. What remains is a short list of specific claims that the answers actually make.
Number the claims. Matching claims go on the same row. Claims that appear in only one answer get their own row with a blank on the other side. This table is the core artefact of the method. It turns two smooth paragraphs into something a student can reason about, one row at a time.
Step three: mark agreement, conflict, and silence
Label each row with one of three marks. Agree means both answers say the same thing. Conflict means they say different things about the same point, such as different dates, causes, or quantities. Silence means only one answer mentions the point. Encourage students to be strict: two answers that use different wording but mean the same thing count as agreement; two answers that sound similar but give different specifics count as conflict.
Discuss what the marks mean. Agreement does not prove a claim is correct, because a model can repeat the same mistake twice. Conflict shows at least one answer is wrong or incomplete. Silence shows that the answer depends on how the question was asked. Each of these is a useful observation about the tool, and students can express it in a single sentence.
Step four: check one disputed claim
Pick one conflict row and check it against the reference the class trusts, such as the textbook, a teacher-approved website, or a primary document. Write down what the source says and which answer, if either, it supports. One careful check teaches more than a hurried attempt to verify everything, and it models the habit of going to a source before repeating a claim.
Then check one agreement row in the same way. Sometimes both answers are right; sometimes both are wrong in the same direction. Finding either outcome teaches students that agreement between AI answers is not a substitute for a source. Record the result honestly, including when the source is unclear.
Writing the human sentence at the end
Finish by asking each student to write one sentence in their own words that they are now prepared to say, with the source named. For example: according to our textbook, this happened in this year, and one of the two AI answers had the date wrong. This sentence is the outcome of the module. It is short, attributable, and can be marked.
Teachers can collect the comparison tables and sentences as evidence of the skill. They show whether students can separate background from claims, label differences accurately, and use a source. None of this depends on which AI tool was used, so the method stays useful as tools change.
Running it as a single lesson
In a single lesson, spend the first few minutes on the question and the two answers, the middle on lining up and marking claims in pairs, and the last part on the source check and the human sentence. Pairs work well because one student can read claims aloud while the other writes them into the table. Collect one table per pair rather than per student so that the discussion stays focused on reasoning, not on copying neatly.
Repeat the routine with a different subject a few weeks later. Students usually get faster at separating background from claims and more careful with the labels. That improvement is a fair, observable measure of AI literacy that does not depend on any claim about the tool itself.
Common mistakes and how to avoid them
A common mistake is comparing answers to an opinion question, where differences are expected and nothing can be checked. Start with factual questions. Another is treating the longer or more confident answer as the better one. Length and tone are not evidence. A third is trying to check every row in one lesson, which turns a thinking exercise into a race.
Families can use a lighter version at home: ask the question twice, read both answers aloud, and talk about one difference. The habit is the same in both settings. Before repeating what an AI tool says, see whether it would say the same thing again, and look up the part that changed.