Whose Story Is Missing?

Overview

Teachers increasingly use AI to generate classroom images, word problems, stories, and translations. Those outputs can quietly repeat whatever was most common in the data a model learned from: one kind of family, one kind of scientist, one version of a town's history. In this session, teams audit eight mock AI outputs for representation, name the likely cause of each pattern, and decide whether a better prompt or a human fix is needed. Then they revise one output so every student could see themselves in it and access it, using UDL and basic accessibility checks.

Objectives

• Participants will identify who is represented, missing, or stereotyped in AI-generated classroom materials and explain why such patterns can occur.
• Participants will distinguish problems a better prompt can fix from problems that require human knowledge, community voices, or a different approach.
• Participants will revise one AI-generated material using a UDL and accessibility checklist so more students can see themselves in it and use it.

Step by step

1. 10–5 min

   Hook

   Picture a scientist

   Ask everyone to close their eyes: "Picture a scientist at work. Where are they? What are they wearing? What are they doing?" Eyes open; three people describe their picture. Then say: "Our minds fill in defaults from everything we've seen. AI tools do something similar with everything they were trained on. The question for teachers is what those defaults teach our students about who belongs." Share your own example of a material that missed your students.

   Facilitator noteKeep the tone curious, not accusatory. Everyone has defaults; the skill is noticing them.

2. 25–20 min

   Explore

   Audit the outputs

   Teams take the eight cards and complete Handout B for at least six: who or what shows up, who is missing or stereotyped, why the pattern might exist, and whether the fix is a better prompt, a human fix (knowledge, community voices, review by someone fluent), or both. Each team marks the one card they think would do the most harm in their own classroom if used as-is.

   Facilitator noteCard 7 is also a hallucination: the AI invented a founder for a fictional town. Watch for teams that catch representation but miss accuracy, or the reverse. Both matter.

3. 320–28 min

   Model

   Why it happens, and why "add diversity" isn't enough

   Explain in plain terms: "A model generates what is likely based on patterns in its training data. If some people, places, or family types appear less often in that data, or mostly in stereotyped ways, outputs can reflect that." Then model a fix on Card 1 out loud. First try a better prompt ("scientists of different ages, genders, and backgrounds, working in the field and in labs, including a scientist who uses a wheelchair"). Then critique your own fix: "Better, but notice I'm still choosing who counts. Prompting for 'diversity' can produce token or stereotyped images too. The strongest fix may be real photos of real scientists from our community, or letting students draw scientists they know."

   Facilitator noteStay general and accurate about causes: training data, defaults, and design choices. Avoid claims about how any specific product was built.

4. 428–43 min

   Create

   Fix one for real

   Each team picks the card they marked as most harmful and revises the material on chart paper (or in a document) so it could go to their students tomorrow. They must use Handout C: at least one engagement choice (relevance, student choice), one representation choice (more than one way to take in the content, such as text, image, audio, or objects), one action and expression choice (more than one way to show learning), plus the accessibility items (alt text, readable layout, meaning not carried by color alone). Teams note which fixes used AI and which needed a human.

   Facilitator noteUDL's three principles (engagement, representation, action and expression) come from CAST. Push teams past "add pictures" to real options students can choose.

5. 543–52 min

   Debrief

   Gallery and hard questions

   Teams post their revisions and do a quick 3-minute walk. Then discuss: "Which fixes could a prompt handle? Which needed a person, such as a family member, a colleague fluent in a language, or students themselves? What would you tell students about AI images and stories so they can notice missing voices on their own?"

   Facilitator noteListen for teachers moving from "AI is biased, don't use it" to "AI outputs need a human review for representation, just like any textbook."

6. 652–60 min

   Transfer

   A student version

   Each person plans how students will audit an AI output in their class. Suggested frame: students examine one AI-generated image or text and answer "Who is here? Who is missing? How could we find out what's true?" Then they add a missing perspective from their own knowledge, family, or community. Partners share plans and complete this sentence for each other: "The evidence that students noticed missing voices will be…"

   Facilitator noteThis builds a welcoming classroom (Teacher ELE 3.1) when students' own experiences become the correction.

Adaptations

K–2 teachers

Use Cards 2, 5, and 6 and a simple student routine: "Who is in the picture? Who is not? Draw someone who should be." The teacher drives any AI tool on the projector.

Secondary teachers

Have students write a better prompt, rerun it with the teacher, and then argue whether the new output is actually more accurate or just more varied.

Librarians

Pair the audit with a search for real, authored sources (local archives, community interviews, diverse book lists) that fix gaps no prompt can fill, such as Card 7's missing history.

Emergent bilinguals

Use Card 8 with students as experts: they evaluate a machine translation of a class message and teach the class what it got wrong.

Pairs well with

• Digital Accessibility Readiness AssessmentGo deeper on the accessibility side of the checklist.
• Gen AI Literacy Breakouts (K–8)Student activities on AI limits and bias.