A 30-Minute AI Source-Checking Lesson: From Answer to Evidence

Featured image: AI-generated conceptual illustration, not a photograph of a real class.

An AI answer can become the beginning of a lesson rather than the end of one. Instead of asking learners only whether they can produce a clear response, ask a more demanding question: can they show which parts deserve to be believed?

This 30-minute activity turns that question into a practical classroom exercise. It is designed for adult learners, staff workshops and appropriately supervised educational settings. The goal is not to catch a chatbot making a mistake. It is to practise moving from an attractive answer to an accountable explanation.

Why teach more than prompting?

In Generative AI Literacy: Twelve Defining Competencies, Ravinithesh Annapureddy, Alessandro Fornaroli and Daniel Gatica-Perez propose a broad competency framework encompassing foundational knowledge, prompting, programming, and ethical and legal considerations. The paper was submitted in November 2024 and has a March 2025 journal reference. Its breadth is a useful reminder that using generative AI responsibly involves more than phrasing instructions well.

The lesson below is an original teaching proposal, not an intervention tested or endorsed by those researchers. Its central design choice is simple: assess the learner’s evidence trail, not merely the polish of the final paragraph.

Before the lesson: prepare a small evidence pack

Choose a low-stakes topic with two or three short, accessible primary sources. A museum’s collection entry, a public scientific explainer or a university’s description of a research project can work well. Avoid medical decisions, live legal disputes or other topics where a beginner’s mistake could cause harm.

Prepare a short AI-generated answer on that topic and label it clearly. Check it yourself before class. It does not need to contain a false statement: a correct answer still needs supporting evidence. If you introduce an altered claim for teaching purposes, label the material as a deliberately constructed exercise and disclose the alteration during the debrief.

Provide printed copies as an alternative to online access. Use public information, not learners’ personal records. A shared sample answer also means participants do not need individual chatbot accounts to take part.

Minutes 0–5: separate the claims

Give learners a paragraph of approximately 120 words. Ask them to underline three factual claims and circle any language that expresses interpretation, prediction or opinion. Then invite them to split sentences containing several claims into separate statements.

For example, the fictional sentence “The project launched in 2022, attracted 5,000 participants and transformed local learning” contains two potentially checkable statements and one broad evaluative claim. A source confirming the launch date would not, by itself, establish the participant figure or the claimed transformation.

Minutes 5–15: follow the evidence

Working individually or in pairs, learners match each claim to a source passage. For each match, record the source title, organisation or author, publication date where available, and the exact location of the supporting information.

Ask four questions: Does the source actually exist? Does it discuss this specific claim? Does it support the wording used? What does it leave unresolved? A relevant link is a starting point, not a completed check.

Use four possible judgements: supported, partly supported, contradicted, or not established by the available evidence. Keep the last category distinct from false. Failure to find support in a short exercise does not prove that a statement is wrong.

Minutes 15–23: repair the answer

Learners rewrite the paragraph using only claims they can support. They should remove unsupported detail, narrow overgeneralised language and mark genuine uncertainty. If the evidence reports an association, the revised wording should not silently turn it into a causal conclusion.

A fictional example makes the distinction clear. “Everyone preferred the new tool” might become “In the small group described by the project report, most respondents preferred the new tool.” The revised sentence is less dramatic, but its scope is easier to defend. Learners must use the actual numbers and limitations in their chosen source rather than copying this illustrative wording.

Minutes 23–30: explain the judgement

Finish with a short explanation from each learner: one claim retained, one claim changed, and one question still unanswered. Ask what additional evidence would change their conclusion.

The facilitator can use a simple four-part checklist: the learner identified a specific claim; located relevant evidence; matched the strength of the wording to the evidence; and explained a remaining limitation. This is a classroom feedback tool, not a validated psychological or educational assessment.

What this activity can—and cannot—show

The exercise makes a learner’s reasoning visible. It does not establish that one half-hour session produces lasting AI literacy, nor does it measure every skill involved in responsible AI use. Learners may need subject knowledge, reading support or additional time to interpret a source accurately.

A prepared evidence pack also simplifies the real world. Later sessions can introduce conflicting sources, outdated material, inaccessible documents and unsupported citations. Increase that complexity gradually rather than treating speed as proof of competence.

For accessibility, allow spoken explanations, paired work, screen-reader-friendly documents or enlarged print. Keep the learning objective consistent while varying how participants demonstrate it.

The takeaway: confidence should come with a reason

The most useful closing question is not “Did the AI get it right?” It is “What evidence justifies the answer we are willing to use?” That question leaves room for AI to help with drafting while keeping responsibility for the final judgement visible.

Try the activity with one short answer this week. Keep the original response, the evidence notes and the corrected version together. The difference between those three items is the lesson.