Verify the Bot

Overview

AI chatbots explain science fluently, and they can blend accurate explanations with classic misconceptions, invented citations, and numbers that sound right. The fix isn't "don't trust AI" or "trust AI"; it's a routine: split the answer into claims, check each claim against two independent sources, then keep, fix, or cut. Participants practice the routine on printed mock AI answers about seasons and plant growth, run it live on a district-approved chatbot, and then plan a 15-minute version they can lead for their department, because teaching colleagues to verify is how the habit spreads to students.

Objectives

• Participants will split an AI-generated science explanation into individual, checkable claims (facts, numbers, citations, explanations).
• Participants will verify claims with at least two independent, credible sources and label each claim keep, fix, or cut.
• Participants will revise a prompt or follow-up question to make an AI output easier to verify, and explain why that doesn't replace verification.
• Participants will plan a short workshop to teach the routine to colleagues or students.

Step by step

1. 10–5 min

   Hook

   Sounds right to me

   Project Mock Answer 1 (seasons). Ask the room to rate it 1–5 for "I'd let students use this." Most rate it high. Then ask: "How many separate facts did it just assert?" Nobody knows. Say: "That's the problem. We judged the whole thing by its tone. Today we check it one claim at a time."

   Facilitator noteMock Answer 1 includes a common misconception: that Earth is closer to the sun in summer. In fact, Earth is closest to the sun in early January, during Northern Hemisphere winter.

2. 25–13 min

   Model

   The routine: Split, Source, Compare, Decide

   Think aloud through the first three sentences of Mock Answer 1. Split: underline each separate claim and number it on Handout B. Source: for each, find a credible, independent source (a textbook, a science agency, a university education page), not another AI answer. Compare: find a second source that doesn't copy the first; do they agree? Decide: keep, fix (write the corrected claim), or cut. Narrate one dead end honestly.

   Facilitator noteTwo sources that both copied the same blog post aren't two sources. Model checking where the second source got its information.

3. 313–27 min

   Practice

   Ledger race

   Pairs finish Mock Answer 1 and complete Mock Answer 2 (where plant mass comes from) on Handout B, using devices for sources. Pay special attention to the citation in Mock Answer 2: trace it. Can you find the study, the researcher, the university? Pairs compare ledgers with the other pair at their table.

   Facilitator noteKey for Answer 2: plants gain most of their mass from carbon dioxide in the air, not from soil. The "Dr. Helena Marsh, Maple Hollow University" citation is fictional and untraceable. That's the lesson: AI can generate citations that look real. Absence of the source is a finding.

4. 427–39 min

   Apply

   Live round: check a real answer

   Each pair asks the approved chatbot a science question from their current unit (or the facilitator's prepared question), then runs the routine on the answer. Next, they try one prompt move to make verification easier, such as: "List each factual claim separately," "Say which claims you are less certain about," or "Name the type of source a student should use to check each claim." Discuss: did it help? Does it replace checking?

   Facilitator noteBe accurate: asking a chatbot to check itself or to rate its own confidence can help surface weak spots, but the model's self-assessment can also be wrong, and any sources it names still need to be found and read. The routine still applies.

5. 539–52 min

   Create

   Plan your 15-minute workshop

   Each person uses Handout C to plan a 15-minute version of this session for their department, PLC, or students: the mock or real AI answer they'll use (from their own subject), the claims they expect people to catch, the two sources they'll point to, and the one habit they want colleagues to leave with. Pairs rehearse the opening two minutes for each other.

   Facilitator notePush for subject-specific examples. A biology teacher's misconception-laden AI answer will land better with the science department than any generic example.

6. 652–60 min

   Debrief

   What will students turn in?

   Ask: "If students use AI in your class, what evidence will show they verified it?" Collect answers: a claim ledger, a keep/fix/cut annotation, a corrected paragraph, two cited sources per claim. Agree as a group on one common verification artifact the department could use across courses.

   Facilitator noteThis connects to the 2026 ELE question: technology use by itself isn't evidence of learning. A completed claim ledger is.

Adaptations

Grades 6–8 science teachers

Use Mock Answer 1 only and simplify to three moves: split it, check it twice, fix it. Let students verify with the textbook plus one teacher-selected website.

Librarians

Lead the Source and Compare moves: teach how to tell whether two sources are independent, and how to search library databases for the original study behind a claim.

Other departments

Swap the science answers for a history summary checked against primary sources, or a math explanation checked by working the problem. The routine carries over unchanged.

Pairs well with

• Digital Inquiry Group: Civic Online ReasoningLateral reading lessons that strengthen the Source and Compare moves.
• AI Essentials Prompt LibraryTry prompts that make outputs easier to verify, then verify anyway.