Workshop kit · AI in practice
Checking AI-generated quotes
A 90-minute session for communication teams. Pairs generate a mock expert interview, hold it to the same standard as any quoted material, and leave with a written plan for verifying every claim.
01
Why this session
Generative tools are very good at producing a quote that sounds like someone said it. That makes them useful for drafting and dangerous at publication, and the gap between the two is a habit, not a tool.
August 2024
A reporter at a Wyoming weekly resigned after a competing paper found AI-generated quotes attributed to seven people, the state's governor among them.
May 2025
A syndicated summer reading list that ran in two major daily papers recommended books, credited to real authors, that did not exist. A reader had to reach the 11th of 15 entries to find a real one.
Both reached print under a news organization's name. This session practices the check that would have stopped them.
If fluency does not tell us who said something, what does?
02
The outcome
By the end of the session, each pair evaluates an AI-generated expert interview against a four-point editorial standard and writes a verification plan covering at least five attributed claims. In the plan, every flagged claim names who or what could confirm it, and what happens if it cannot be confirmed before deadline.
A plan meets the bar when all four points of the standard appear across those claims and no flagged claim is left without a route or a fallback. "The output generally needs checking" does not meet it, however long it is.
03
Before the day
- Send a short survey the day before. Ask how often each person uses generative AI and how they feel about it, then show three brief AI-generated claims and ask which would need verification before publication, with a confidence rating. Use the first two answers to build pairs. Keep the claim items, because closure repeats them.
- Prepare the opening quotes. Pick one short quote from a real published interview on a topic your team covers, then prompt an AI tool for a quote from a comparable expert on the same topic. Display both without labels.
- Set up an anonymous poll with one question for each round of voting, so results can go on screen live.
- Print one copy per person of the four-point standard and the verification plan, with digital copies for screen reader users.
- Generate and save a backup transcript with at least five attributed claims, for anyone who would rather not generate their own, and one AI-generated image for the visual version.
- Print the prompt scaffold for anyone new to AI tools.
- Check the room: captions on for any video, and your team's approved AI tool working on participant laptops.
04
Agenda
The stakes
Open with the two real cases above.
The quote vote
Show the two unlabeled quotes. Everyone votes anonymously on which one a real person said. Without showing results, partners explain their choices to each other, then vote again. Show both rounds side by side and ask what changed minds. The reasons people give are usually about tone and confidence, which is exactly what a model is good at producing. Reveal the answer.
Where this leads
Explain that the plans written today will become the starting material for the team's own verification checklist.
How the output is made
Generative models predict likely text, which is why they can produce a confident quote no one ever said.
The standard, modeled
Walk through the four points with one concrete example each. Then think aloud through one claim from the backup transcript, mark it cleared, flagged or unverifiable, and fill in its row of the verification plan on screen. Leave that row up during practice.
Where the human sits
Ask the room where the human check currently sits in your own publishing workflow, if it sits anywhere.
Generate
Pairs pick a current or recent project and prompt the AI tool to simulate an interview with a relevant expert.
Annotate
Pairs mark at least five claims as cleared (C), flagged (F) or unverifiable (U) against the standard.
Plan
Pairs complete the verification plan: each flagged claim, who could confirm it, and the fallback if they cannot by deadline. Early finishers take an extension.
Share
Each pair shares one flagged claim and its reasoning. Record recurring problem types where everyone can see them.
The disclosure question
Would we ever publish material drawn from a simulated interview, and if so, with what disclosure? The PRSA Code of Ethics provision on disclosure gives the room a reference point.
Retake and commit
Everyone retakes the three survey claims and the confidence rating, then completes: "One verification practice I will adopt in my work is…"
Close the loop
Name the point in your workflow where this check would live and who would own it. Collect the plans.
05 · Handout
The four-point editorial standard
Apply each question to every attributed statement in the transcript.
01
Factual accuracy
Can each factual claim be confirmed in a source outside the transcript? Are numbers, dates, names and titles correct? Is anything stated with more certainty than the evidence supports?
02
Attribution integrity
Is anything presented as a quote from a real, identifiable person? Did that person actually say it, and could we show where? Would they recognize it as theirs?
03
Voice authenticity
Does this sound like a specific person or like a general expert? Does it have the hedges, specifics and occasional disagreement real experts tend to offer?
04
Publication risk
If this ran under our name and turned out to be wrong or invented, who would be harmed: the named person, our audience, the institution? What would a correction cost?
06 · Handout
Verification plan
One per pair. The completed plan is what the facilitator collects.
| Claim from the transcript | Standard point(s) | Who or what could confirm it | Fallback if not confirmed by deadline |
|---|---|---|---|
- In our current workflow, this check should happen at: ________
- The person responsible for it would be: ________
- Each partner: "One verification practice I will adopt in my work is…"
07
Prompts
Describe a type of expert rather than naming a real person. The point is to see how convincingly the tool invents an expert voice.
Starter prompt
Act as [a type of expert, for example a structural engineer, a public health researcher, or a sports nutritionist]. A writer from a university communications office is interviewing you for a short feature about [topic or project]. Answer the four questions below in that expert's voice, using two to four sentences each. 1. [Question] 2. [Question] 3. [Question] 4. [Question]
Follow-up, once the interview is generated
Which of your answers include specific facts, figures, names, or dates that a writer would need to verify before publishing, and where would someone check each one?
Compare the tool's list with the claims you flagged yourself. Anything you flagged that the tool missed belongs on your plan, and is worth noticing.
08
A version for photographers and videographers
The scenario: a striking photo arrives through a community account, captioned as taken on campus last night, and the social team wants to repost it within the hour. The facilitator supplies an AI-generated image prepared in advance. Pairs apply the same four points to the image.
Factual accuracy
Do the buildings, signage, weather and light match the place and time in the caption?
Attribution integrity
Who made the image, can they be reached, and can they provide the original file or other frames from the same moment?
Voice authenticity
Are shadows, reflections, hands and text consistent? Does it match the account's other photos? Can metadata or a reverse image search support where it came from?
Publication risk
If it turned out to be fabricated, what would reposting it cost the institution, and who would be misrepresented?
Pairs fill in the same verification plan, listing each visual detail they could not confirm as a claim.
09
Running the room
People differ in three ways that matter here: job role, comfort with AI tools, and how they feel about the technology. The goal stays the same for everyone.
By comfort with AI
| Never used it | Paired with an experienced user and given the printed prompt scaffold. The partner runs the tool while the newer person leads the evaluation, so an unfamiliar interface does not get in the way of the skill. |
| Uses it daily | Extension: coax output that passes all four points, then explain what made it pass and whether passing should be reassuring. |
| In between | The standard activity, with the scaffold available on request. |
By attitude toward AI
Enthusiasts and skeptics both tend to evaluate output poorly at first: enthusiasts because a fluent answer is easy to accept, skeptics because it is easy to dismiss before reading closely. Introduce the standard as a checklist that works the same wherever someone starts. Where the pairing allows, put an enthusiast with a skeptic so both urges get argued out over the same claim. Anyone who declines to use an AI tool evaluates the backup transcript instead. The goal is evaluation, so skipping generation costs them nothing.
Access and participation
- Handouts in print, large print and digital form.
- Captions on any video, with key points restated aloud.
- Pair talk for verbal processors, the written plan for people who think in writing.
- Sharing is by pair, so no one is put on the spot, and anonymous voting keeps the loudest voices from settling the answer.
10
Measuring it
During practice
Ask each pair to name one claim they flagged and one they cleared, and to explain why the cleared one is safe. A pair that cannot explain a clearance is applying the standard mechanically. Redirect by asking what evidence outside the transcript supports it. The two rounds of the opening vote are an early read: if answers shift after partners talk, the discussion is working.
The plan itself
| Criterion | Meets | Approaching |
|---|---|---|
| Applies all four points | Each point is addressed in the plan. | One or more points skipped, or treated as one generic check. |
| Identifies specific claims | Names at least five specific claims. | Fewer than five, or "needs checking" without naming what. |
| Assigns verification routes | Each flagged claim has a named source or person. | Verification described as general fact-checking. |
| Addresses the deadline case | States a fallback: cut, rephrase or hold. | No fallback, or the fallback is to publish anyway. |
Before and after
The three claim items and confidence rating from the day-before survey are repeated at closure and compared for the group as a whole. The claim items measure skill. Confidence is reported alongside but treated as secondary, since confidence can rise after a short session without any gain in judgment.
11
After the session
Within two weeks, combine the collected plans, the recurring problem types and the commitment statements into a draft verification checklist for AI-generated content. Take it to the next staff meeting for revision, along with the bigger questions about attribution and intellectual property the session only opens. Participants will see their own flagged claims in the draft, which ties 90 minutes to a decision the office makes about its own publishing process.
Resources
- Public Relations Society of America, Code of Ethics, including the Disclosure of Information provision.
- KUNC / Wyoming Public Media, "Wyoming journalist resigns after using AI-generated quotes", August 16, 2024.
- Axios Philadelphia, "Philadelphia Inquirer prints summer reading list full of AI-generated fake titles", May 20, 2025.
- IBM Technology, "What is Human In The Loop with AI?" (video, March 2026).
- Mazur, E. (1997). Peer Instruction: A User's Manual. Prentice Hall. The source of the vote, discuss, revote method.
Designed in fall 2026 as part of ENGR 61095, AI and Automation, at Kent State University, and revised after three peer reviews. Pairs with Prompt to Publish, the field guide.