you teach adults to catch AI hallucinations by having them catch one. In a 75-minute lab, each learner asks an AI tool for a specific, checkable fact from their own field, such as a rate, a code or a cited statistic. They circle every claim in the answer, check the one that matters most against the official source, and write down what they fixed. The habit sticks because they saw the tool get something wrong, or saw how close it came, with their own eyes.
Version 1. ed.cafe has not run this lab with a cohort yet. Circle #1 is still forming. The timings and prompts below are a starting design, and we will revise them after the first run.
The question you are actually asking
You have probably told a class that AI "can make things up." Heads nod. Then, a week later, someone pastes an AI-written paragraph into a work email with a statistic nobody checked.
A warning is not a skill. Your learners know AI can be wrong in the same way most drivers know brakes can fail. Knowing it doesn't change what they do at the moment it matters. What changes behavior is a practiced routine, run once with someone watching, on a task that looks like their real work.
The U.S. Department of Labor's voluntary AI Literacy Framework (TEN 07-25) names this skill "Evaluate AI Outputs." Its wording is plain: "Workers must cross-check AI-generated outputs against trusted sources or known information to identify false claims, outdated references, or fabricated content." The framework says what workers should do. It does not give you a lesson for teaching it. This lab is one.
Why a lab beats a lecture: one documented example
The best warm-up is a real case from the course your learners are taking. Scott Kirsner, writing for MassLive (syndicated via AOL), worked through the Google AI Professional Certificate. During one exercise he asked Gemini for its sources. Gemini replied: "I have to confess… I generated illustrative data and realistic-sounding citation titles to demonstrate how the infographic layout would look." Kirsner's verdict: "the infographic was pure imagination."
Read that quote aloud at the start of the lab. Notice what it shows. The output looked finished. The citation titles sounded real. The only thing that exposed the problem was a person asking, "Where did this come from?"
That question is the whole lab. Everything else is practice in asking it quickly.
The 75-minute lab script
Run this as Evening 2 of a four-evening cohort, or as a stand-alone session. You need one device per pair, a projector, and printed check cards (below).
- Minutes 0 to 10: Warm-up. Read the Kirsner quote aloud. Ask: "What would have happened if he had sent that infographic to his boss?" Take two answers. Don't lecture.
- Minutes 10 to 20: Live demo. Run one prompt from the list below on the projector. Circle every number, name, date and citation on screen. Check one live, out loud, including the moment you open the official site.
- Minutes 20 to 45: Pairs run a field prompt. Each pair picks the prompt closest to their job. One person types; the other circles claims on paper. Then they check one claim using the two-minute routine.
- Minutes 45 to 60: Fix and write it up. Each pair writes a corrected version of the sentence and a one-line note: what was wrong, which source they used, and the date they checked.
- Minutes 60 to 70: Debrief. Use the questions in the debrief section below.
- Minutes 70 to 75: Close. Show how this links to "checked and fixed" on the work-sample rubric. Assign: repeat the routine once on a real (non-private) work task this week.
Field prompts that invite checkable claims
Each prompt asks for a specific fact plus its source. Specific facts can be checked against one official page in a few minutes. Vague answers ("customer service matters") cannot.
Bookkeeping: "What is the IRS standard mileage rate for business use in 2026? Give the IRS notice number and a link." Check against: IRS.gov, searching the notice number yourself.
Medical billing: "What does claim adjustment reason code 50 mean? Quote the official description and give the source." Check against: the official code list published by the code maintainer, not a billing blog.
Retail or operations: "What is the Massachusetts minimum wage in 2026? Cite the law or official page." Check against: Mass.gov.
Customer support, or any field: "Give me three statistics about customer satisfaction with chatbots, each with the study name, publisher, year and a working link." Check against: open each link. Does the page exist? Does it say that number?
The catch: no prompt fabricates on command. Tools change, and sometimes the answer will be right. That is fine. The skill is the check, not the gotcha. A learner who confirms a correct fact has still practiced the routine. In the debrief, ask how they would have known if it had been wrong. The citation prompt is the one most likely to produce something that doesn't hold up, because it asks for several sources at once.
The two-minute check card
Print this and hand one to every learner. It fits on an index card.
THE TWO-MINUTE CHECK
1. CIRCLE every number, name, date, code and citation.
2. PICK the one claim that would cause harm if wrong.
3. ASK "Who would publish this officially?"
(IRS, Mass.gov, the code maintainer, the study's author)
4. GO to that source yourself. Type the address.
Do not click the link the AI gave you.
5. COMPARE word for word. Same number? Same year?
6. MARK Match / Wrong / Can't find.
"Can't find" counts as wrong until proven otherwise.
7. FIX Write the corrected sentence + source + date checked.Step 4 matters most. An AI-supplied link can point to a real website that doesn't contain the claim, or to a page that doesn't exist. Going to the publisher directly separates "this site exists" from "this site says this."
A worked example with the bookkeeping prompt
Illustrative example: a learner asks for the 2026 mileage rate. The tool replies with a rate in cents per mile, a notice number, and a link.
- Circle: the rate, the year, the notice number, the link. Four claims.
- Pick: the rate. If a client bills mileage with the wrong rate, the books are wrong.
- Ask: who publishes this? The IRS.
- Go: the learner types irs.gov and searches "standard mileage rate 2026." They don't click the AI's link.
- Compare: suppose the IRS page shows a different rate, and the notice number the tool gave belongs to an earlier year.
- Mark: Wrong. The tool gave last year's figure with confidence.
- Fix: "The 2026 business rate is [rate from IRS page], per IRS Notice [number from IRS page], checked 15 Oct 2026."
That is a passing "checked and fixed" note. It names the source, shows the error and shows the correction.
Running it with mixed-skill groups
Mixed groups are normal in adult classes. Three adjustments help:
- Pair a fast typist with a careful reader. The typist drives the tool. The reader holds the pen and circles claims.
- Let ESOL learners prompt in their strongest language if the tool allows, then check against the English-language official page together. The comparison step is the same in any language.
- Give learners who finish early a second job: check a claim the AI did not attach a source to. That is harder and closer to real work.
Debrief questions
Keep the debrief to ten minutes. Ask, in this order:
- "What did you find?" (Match, wrong, or can't find. Tally on the board.)
- "How did the wrong answer look? Did anything about it seem off before you checked?"
- "How long did the check actually take?"
- "Which claims in your answer would you never have checked if you were in a hurry?"
- "Where at work would one wrong number cause the most trouble?"
Question 2 usually produces the key insight. Fabricated answers rarely look fabricated. They look finished.
Where this lab stops working
The lab teaches one routine well. It has limits you should name in class.
- It checks facts, not judgment. A correct fact used in the wrong context still produces bad work. Bias and missing perspectives need a different exercise.
- Some sources are paywalled or hard to find. Teach "can't find" as a valid result, and say who to ask instead.
- Tool behavior changes. Re-run your demo prompt before each class. If it stops producing anything interesting, swap it.
- Never use real private data. Every prompt above uses public facts. Do not let learners paste a real client's details into the tool "to make it realistic." Use invented names and amounts, labeled as invented.
If your organization's staff want practice first, the City of Boston runs free AI and data literacy workshops for community organization staff, covering "privacy and your data, hallucinations, bias, and misinformation," with sessions through 21 December for hosts with 8 or more participants.
Decision checklist
- Did every learner circle claims before checking any?
- Did they go to the source directly instead of clicking the AI's link?
- Did each pair write a fix with a source and a date?
- Did you name "can't find" as a result, not a failure?
- Did you link the lab to the "checked and fixed" criterion learners will be scored on?
The written fix from this lab is exactly what criterion 2 of the 12-point work-sample rubric asks for, so start there when you plan how to score it.
Sources
- TEN 07-25: Artificial Intelligence Literacy Framework, U.S. Department of Labor ETA, 13 February 2026
- Scott Kirsner on completing Massachusetts' free Google AI course, MassLive, syndicated by AOL, 2026
- Analyze Boston and AI Literacy Workshops, City of Boston Department of Innovation and Technology, 2026 cycle
