Teaching Students to Fact-Check ChatGPT and Claude: A Practical Classroom Lesson

Activity Worksheets
Written By Faysal Hafidi

A polished answer fills the classroom screen. It is clear, confident, and full of scientific vocabulary. “It sounds correct,” Maya says. Her classmate points at one detail and asks where the information came from. Nobody knows. The teacher responds with a better question: What could convince someone who has never used this chatbot that the statement is accurate?

That question transforms the activity. Instead of seeing ChatGPT or Claude as an oracle to trust or an enemy to catch making mistakes, students treat an AI-generated response as a set of claims that can be tested. They learn the difference between fluency and reliability, quotation and evidence, and uncertainty and error. Their final task is not to select a winning chatbot. It is to defend a conclusion using sources outside the chatbot.

This ready-to-use 50-minute lesson is intended for middle and high school educators. It includes a verification routine, a complete lesson sequence, a teacher-prepared claim set, a reusable student worksheet, an answer key, a formative rubric, and adaptations for different subjects. Its narrow goal is to teach students how to fact-check factual claims generated by AI. It is distinct from detecting AI-written homework or testing whether a student understands a correct AI answer.

Why does teaching AI fact-checking matter?

ChatGPT and Claude can explain complicated topics helpfully, yet a convincing response is not necessarily accurate. OpenAI warns that ChatGPT can make errors and invent references; Anthropic explains that Claude may generate convincing but unsupported content. A citation in an AI answer must still be opened. It might support only one sentence, omit an essential qualification, or lead to a source that does not say what the response asserts.

Asking an AI, “Are you sure?” is not independent verification. Nor is asking another chatbot always independent corroboration: different systems may repeat the same myth. Students need a repeatable procedure for checking facts against evidence suited to the subject.

Research on lateral reading supports this direction. Lateral readers leave a webpage to investigate its author, original sources, and outside corroboration instead of evaluating credibility from appearance alone. A 2021 study of college civics students reported stronger fact-checking performance after such instruction than in comparison classes. The Digital Inquiry Group's Civic Online Reasoning work brings related strategies into school teaching. These findings concern particular instruction and populations, not proof that every AI lesson will work equally well.

The UNESCO AI Competency Framework for Students likewise emphasizes critical judgment and responsible human agency. A responsible classroom response is not to train children to distrust every sentence produced by technology; it is to give them the habits and tools to justify their conclusions.

Lesson outcomes, materials, and classroom safeguards

At the end of the period, students should be able to isolate a factual claim, identify an appropriate source, find the relevant evidence, distinguish a supported claim from a false or misleading one, and write an accurate correction. They should also recognize a sentence whose conclusion is true but whose explanation is false.

Time: 50 minutes. Suggested ages: 12–18, adaptable. Preparation: ten minutes to review sources and prepare the display or printed documents. Materials: projector, investigation paragraph, student worksheets, NASA source pages or printed extracts, pens, and optional web-enabled devices.

A teacher-led demonstration is sufficient: students do not need personal ChatGPT or Claude accounts. Comply with local age rules, school policy, accessibility requirements, and privacy protections. Do not submit identifiable student details, grades, personal information, or confidential work to a chatbot. For students without internet access, use a printed source pack; the investigative reasoning can be identical.

The CLAIM routine: Five moves for checking an AI answer

Use CLAIM, a classroom mnemonic proposed in this lesson. It is not a branded chatbot feature or a scientifically validated scale.

Move Question What the student produces
C — Capture the claim What exactly can be checked? One precise statement
L — Locate an authoritative source Who could know this directly? An original record or appropriate expert source
A — Assess the evidence Does the source actually support this wording? Relevant passage plus context
I — Investigate elsewhere What do other dependable sources say? Corroboration, contradiction, or documented disagreement
M — Make a judgment What can I responsibly conclude? Verdict, correction, and remaining uncertainty



Explain that multiple links are not automatically independent sources. Two news articles may repeat one press release. A search-result summary is not the source itself. A government or university domain is not a guarantee that every item on that site is current and correct. The student must assess expertise, context, relevance, and evidence.

A complete 50-minute classroom lesson

Minutes 0–5: Begin with an uncomfortable question

Display this statement: “The AI gave three references and sounds confident, so it is probably right.” Ask students to write whether they agree, disagree, or need more information, followed by a one-sentence justification. Do not reveal the preferred answer yet.

Invite two contrasting explanations. Clarify that the lesson is not about proving ChatGPT or Claude always wrong. The important insight is that professional-sounding wording and a reference list describe how the answer looks, not whether the facts have been established.

Minutes 5–10: Present the investigation paragraph

Project the following teacher-written AI-style paragraph. It deliberately combines accurate and inaccurate claims. It is not represented as a real sampled answer from either platform, so teachers should never cite it as a measured example of model performance.

Venus is the planet closest to the Sun, which explains why it is the hottest planet. The Great Wall of China can easily be seen with the naked eye from the Moon. In 1969, Apollo 11 became the first mission to land humans on the Moon.

Ask each student to underline statements that can be checked. A careful reader may identify more than three: the first sentence contains the claim that Venus is closest, the claim that it is hottest, and a causal explanation connecting the two. These require separate judgments. Have students highlight the assertion they feel most confident about and identify what might make them change their minds.

This step matters because an unexamined paragraph feels like one coherent story. Fact-checking begins by breaking it into propositions that can be investigated independently.

Minutes 10–17: Think aloud with a first source

Begin with “Venus is closest to the Sun.” Ask what kind of evidence would settle the question. Rather than ask ChatGPT or Claude to grade themselves, search for a space agency's planetary facts. Open NASA's Venus Facts: Venus is the second planet from the Sun. Mercury is first.

Then open NASA's Solar System temperature explanation. Venus is indeed the hottest planet, but its dense atmosphere traps heat; its temperature cannot be explained by calling it the nearest planet to the Sun.

Write a corrected model sentence: “Mercury is closest to the Sun, while Venus is hottest because its dense atmosphere traps heat.”

Ask the class why replacing “Venus” with “Mercury” in the first phrase is not enough. The causal explanation must also change. This is how a fact-checking lesson becomes a science lesson rather than a word-swapping exercise.

Minutes 17–32: Investigate in pairs

Pairs take the remaining claims. One student is the investigator, responsible for locating evidence, and the other is the skeptic, responsible for challenging whether the evidence matches the exact wording. Switch roles after seven minutes. Both students record the final judgment individually.

For the Great Wall statement, provide NASA's explanation. Students should discover that the Great Wall is not visible from the Moon with the naked eye. It is also challenging to distinguish from Earth orbit without magnification. Help students distinguish observation from the Moon, images from Earth orbit, and unaided vision from photography with lenses. These are not interchangeable conditions.

For Apollo 11, provide NASA's Moon Missions timeline. Students can verify that Apollo 11 achieved the first human landing on the Moon in July 1969. Ask whether an earlier spacecraft orbiting the Moon would contradict the claim. It would not: the claim is specifically about a crewed landing.

Emphasize that one accurate claim is intentionally included. Students should not assume that every sentence that sounds machine-written is wrong. Good verification recognizes what is supported as well as what must be corrected.

For advanced students, introduce one complication: a source without a visible date, an outdated reference, or two apparently credible sources that disagree. The correct judgment may be “not yet verified.” A lack of evidence is not automatically proof that an event never occurred.

Minutes 32–40: Put the evidence on trial

Invite volunteers to present one checked claim. Give their classmates four questions: What precisely was claimed? Who published the evidence? Which passage directly supports or contradicts it? What did you revise after checking?

Model a respectful challenge: “Your source shows that Venus is hot. Does it show that the Sun's distance is the reason?” This is more educational than saying “wrong answer.” Another productive prompt is “Your search snippet says one thing. What does the full source say?”

Tell students that “Google said so” or “the AI agreed” does not count as a source evaluation. They must open the material, inspect provenance, consider the context, and connect a specific part of it to the exact claim.

Minutes 40–47: Rewrite without AI assistance

Ask students to close the chatbot and rewrite the passage from their verified notes:

Mercury is the planet closest to the Sun. Venus is the hottest planet in our Solar System because its dense atmosphere traps heat. The Great Wall of China is not visible with the naked eye from the Moon. Apollo 11 landed the first humans on the Moon in July 1969.

Students should be able to explain at least one correction aloud. A polished final paragraph alone would not show whether they understand why a statement was changed. Encourage them to say, “The original sentence was misleading because the temperature was correct but the cause was not.”

Minutes 47–50: The exit ticket

Each student answers three short questions on paper: Which claim initially sounded convincing but failed your evidence test? What precise source or passage changed your conclusion? What will you check first next time an AI gives you a factual answer?

Keep these tickets. They provide useful formative evidence about the student's checking strategy rather than a simple memory test on planetary science.

Copy-ready fact-checking worksheet for students

Task: Investigate each statement in the AI-style response. Your goal is not to find mistakes at any cost; your goal is to justify a judgment using evidence beyond the chatbot.

Claim I am checking: __________________________________

Exact words or assumption I need to test: _________________________

My prediction before searching: Supported / Contradicted / Unsure

Most appropriate original source (organization, document title, URL, date if available): _________________________

Evidence from that source, in my own words: _________________________

Does the passage support the entire claim? Yes / Partly / No

A second source or an explanation of the first source's expertise: _________________________

Final decision: Supported / Contradicted / Misleading or incomplete / Not yet verified

My corrected explanation: __________________________________

What evidence would strengthen my answer further? _________________________

Students complete one copy for each selected claim. With limited time, have pairs investigate separate claims and then exchange their evidence notes before writing their independent conclusions. Teachers may collect the sheets as formative assessment or print them without any digital tool.

Teacher answer key and source verification

Claim Judgment Explanation
Venus is closest to the Sun and is hottest for that reason. Misleading and partly false Mercury is closest. Venus is hottest because of its dense heat-trapping atmosphere, not because it is the nearest planet.
The Great Wall is easily visible from the Moon with the naked eye. Contradicted NASA rejects this myth. Magnified photographs from Earth orbit do not establish unaided-eye visibility from the Moon.
Apollo 11 first landed humans on the Moon in 1969. Supported NASA confirms the July 1969 first human lunar landing.

The teacher should not award full marks merely because a student guesses the correct label. A sound answer connects a precise claim with a specific passage and explains the correction. Changing one's initial judgment in response to evidence is a sign of learning.

An eight-point formative assessment rubric

Score four dimensions from zero to two: 0 means absent or incorrect, 1 means partial or teacher-supported, and 2 means independent and well justified.

  • Claim selection (0–2): Does the student isolate a testable assertion and split combined claims when necessary?

  • Source selection (0–2): Does the student use an appropriate expert or original record and identify who published it?

  • Evidence reasoning (0–2): Does the student explain how the source supports the specific wording, including limitations?

  • Correction and uncertainty (0–2): Does the student produce an accurate revision and state what remains unresolved?

An eight-point score describes performance on this task, not a permanent AI-literacy certification. Repeat the assessment using unfamiliar topics. Look for improving independence and fewer cases where students paste unsupported links.

Can teachers run the activity with real ChatGPT or Claude responses?

Yes, if school policy and student access rules permit it. A teacher can ask either system to explain a science topic, display the result, and freeze that answer for the class to investigate. Select the material before the lesson when possible; model outputs change with the prompt, configuration, tool access, and date.

For example, the teacher can ask: “Explain why Venus is hotter than Mercury for ninth graders in 120 words. List claims that a reader should verify and do not invent citations.” Such a prompt may produce a useful starting passage, but it does not guarantee accuracy.

You may ask a model to identify candidate claims to check, but do not let it act as the final referee. An AI system agreeing that it made an error is not independent evidence. Neither is agreement between ChatGPT and Claude. The class must verify the result by consulting appropriate original documents or by carrying out a direct disciplinary check.

Avoid announcing that one model is more truthful based on a single demonstration. Such a comparison would require systematic sampling, defined tasks, and consistent evaluation conditions. The learning goal here is to distinguish unverified claims from evidence-based conclusions, not to rank products.

Subject-specific adaptations

Science: Separate the fact from the explanation

Use a model-generated account of an experiment. Students identify measured observations, claims about mechanisms, and predictions. Ask which part is a measurement and which part is an inference. An authoritative research or science-education resource may confirm the observation while qualifying the proposed mechanism. The Venus example demonstrates exactly this distinction.

History: Trace quotations back to records

Give an AI summary with an attributed historical quotation. Students locate a primary document or reputable archival transcription. They ask whether the quotation exists, whether the person is correctly identified, and whether the context changes the meaning. Secondary sites repeating the same unsourced line are not independent corroboration.

English language arts: Check quotations and interpretations

Use a class text students already know. Provide an AI-written interpretation containing a quoted passage and a claim about a character. Students verify that the quote appears in the text and determine whether the interpretation has textual support. Unlike an invented quotation, an interpretation may be debatable rather than strictly true or false.

Mathematics: Verify directly rather than searching

Provide a worked solution containing one invalid step. Students recompute independently, substitute the answer into the original problem, and explain where the logic breaks. For mathematical claims, a calculation or proof can be a stronger check than a website. Verification is not synonymous with browsing the internet.

Social studies: Check numbers and missing context

Choose a teacher-reviewed statistic and ask students to identify its source, population, time period, and denominator. A number may be technically accurate yet misleading if applied to the wrong year or population. This is an advanced but important step beyond labeling everything simply “true” or “false.”

Seven common mistakes teachers can anticipate

1. Asking a second chatbot for confirmation. Guide students toward sources independent of the chat, not another fluent response.

2. Copying links without opening them. Require a cited passage and an explanation in the student's own words.

3. Treating every statement as false. Keep one clearly supported claim in the exercise to reward proportional judgment.

4. Trusting the first search result. Teach students to inspect the publishing organization, the original document, and the context before reaching a verdict.

5. Relying on impressive-looking domains alone. Expertise is topic-specific. NASA is appropriate for Solar System facts; it is not the automatic authority for every field.

6. Treating uncertainty as failure. Sometimes the responsible answer is “I have not verified this yet.” Reward transparent limits rather than invented certainty.

7. Confusing a corrected paragraph with understanding. Require an independent explanation of what changed and why. Correct copying still does not demonstrate reasoning.

Younger learners, accessibility, and classrooms with no internet

For students aged approximately nine to eleven, replace the three-claim paragraph with two short claims and use preselected, readable evidence sheets. Ask “Which line of the source helps us know?” rather than requiring open-web searching. Provide vocabulary support, accessible text or oral reading, and additional time when appropriate.

The entire lesson can run offline. Prepare three evidence cards, one for each subject claim, and ask students to annotate the original statement in one color and its supporting or contradicting evidence in another. This can make the reasoning more visible and reduce the advantage of students who simply search faster.

In multilingual classrooms, students may first explain their reasoning in a familiar language before rewriting the conclusion in the teaching language. Keep the cognitive challenge constant even when you adapt the reading load or digital setup.

Turn a one-off activity into a classroom habit

A single lesson can introduce the procedure; regular use helps students apply it beyond astronomy. Once a week or when AI-supported research appears in an assignment, ask: “Which claim would cause the most misunderstanding if it were wrong, and what independent check would you perform?” Rotate the subjects and evidence types so students do not associate verification with one topic.

A teacher-as-coach approach strengthens the routine. Instead of instantly labeling a claim correct or incorrect, the teacher asks diagnostic questions, observes source choices, gives specific feedback, and gradually removes scaffolding. Students become responsible for explaining why a claim is warranted. For educators interested in strengthening this questioning and learner-autonomy practice, the Learning Coach training pathway offers a place to explore the available training information without implying that every specialization or credential is immediately open.

The most valuable outcome is not that a student can recite the CLAIM acronym. It is that the student pauses when an answer looks persuasive, asks what evidence would count, follows the source trail, and corrects a statement without embarrassment. AI can accelerate the production of answers; teaching must still cultivate responsibility for knowing why those answers deserve to be believed.

Frequently asked questions

Do students need access to ChatGPT or Claude for the lesson?

No. A prepared AI-style passage and printed source documents are enough. The core work is performed by the students, not the chatbot.

Is asking both ChatGPT and Claude a valid fact-check?

It may expose inconsistencies but does not independently establish truth. A verifiable original source, expert record, or mathematical proof is still needed.

What if a source supports only part of a claim?

Label it misleading or incomplete, describe precisely which part is supported, and rewrite the sentence so it no longer implies more than the evidence shows.

What is the difference between fact-checking and testing student understanding?

Fact-checking evaluates whether a statement matches evidence. Testing understanding evaluates whether the student can explain, apply, and reason independently. A lesson can target both, but success on one does not guarantee the other.

Should every factual sentence in every AI answer be checked?

Checking effort should be proportional to the importance and risk of the claim. Teach students to prioritize surprising assertions, quotations, statistics, disputed topics, and facts that would materially change the assignment's conclusion.

Resources and references

Faysal Hafidi
Previous

Education News in England, October 2026: Six Developments Teachers Need to Watch
Next

When AI Gives the Right Answer: How Can Teachers Know Students Really Understand?