Facilitator guide: Bias in AI

This guide helps a teacher, parent, or ARK ambassador lead the Bias in AI module with students in grades 6 to 8. Like the module, it is a draft under expert review.

Learning objectives

Students will be able to:

• Explain what bias means and how it can get into an AI tool
• Name three places bias can enter: the data, the goal people choose, and how a tool is tested and used
• Recognize who can be harmed when an AI tool works better for some people than others
• Suggest practical ways to catch and reduce bias before a tool is used

Timing (60 minutes)

Time	Activity	How to run it
5 min	Warm-up	Ask: “Has an app ever guessed wrong about you?” Take two or three answers.
10 min	Step 1: What is bias?	Read it together. Pause on the dog example before moving to people.
15 min	Step 2 and the Sort	Walk through the diagram, then have pairs sort the cards and compare.
15 min	Step 3: the scenario	Read the story, vote on A to D by show of hands, then read each choice’s feedback.
5 min	Step 4: quick check	Individually or as a class.
10 min	Steps 5 and 6	Quiet reflection on paper, a few volunteers share, then close with the three questions.

Before you start

• Bias can touch on race, gender, disability, age, and where people live. Keep the focus on how systems work, not on individual students, and let anyone pass.
• Don’t ask students to share personal information, or examples that single out a classmate.
• The reflection in step 5 stays on each student’s own device. If students write on paper instead, they keep it.

Discussion prompts

Talk about itFor a class, a club, or around the dinner tableShowHide

• Should an AI tool ever make the final choice about who gets a job? Why or why not?
• Who should get a say in how a program picks people?
• Is a random lottery fairer than a well-tested AI tool? When might each be better?
• If you could ask the makers of an app one question about fairness, what would it be?

Answer notes

Sort (step 2)

• The data it learned from: the face recognition tool trained on adults, the tool trained on old neighborhood records, and the music app trained on one country’s listeners.
• Choices people made: the “most like our best past employees” goal, testing the voice assistant only with its builders, and using an adult writing tool on middle schoolers.

Some cards could spark debate (people also choose which data to collect). That’s a good discussion, not a wrong answer.

Scenario (step 3)

There’s no single right answer. Strong responses mention checking the data and testing with different groups (choice B). Many classes combine B with D, openness and a way to appeal.

Quick check (step 4)

1. B: an unfair pattern, not a single mistake.
2. False: the bias comes from data and human choices.
3. A: a feedback loop.
4. A: ask what it learned from and who it was tested on.

Sources

Each step lists its sources at the bottom. Reviewers: please confirm each source supports the sentence it’s attached to.

• Step 1, face recognition accuracy differs across groups: Ketan Kotwal and Sébastien Marcel, “Review of Demographic Fairness in Face Recognition,” IEEE Transactions on Biometrics, Behavior, and Identity Science, 2025, arxiv.org/abs/2502.02309; and Vítor Albiero and others, “Gendered Differences in Face Recognition Accuracy Explained by Hairstyles, Makeup, and Facial Morphology,” IEEE Transactions on Information Forensics and Security, 2021, arxiv.org/abs/2112.14656.
• Step 2, the résumé-screening tool that favored men: Jeffrey Dastin, “Amazon scraps secret AI recruiting tool that showed bias against women,” Reuters, October 10, 2018, reuters.com.
• Step 2, predictive policing feedback loops: Danielle Ensign, Sorelle A. Friedler, Scott Neville, Carlos Scheidegger, and Suresh Venkatasubramanian, “Runaway Feedback Loops in Predictive Policing,” Conference on Fairness, Accountability and Transparency, 2018, arxiv.org/abs/1706.09847.