AI projects · Evaluate · weekend
Bias in a word-association test
These projects are not aligned to any state standards or curriculum. They are generally about learning, understanding, using, and evaluating AI.
Ages 11-14 · This project can involve human data — written permission, anonymity, and offline/local handling are required by default.
Overview
Word-association and sentence-completion tests can reveal stereotyped patterns in people and in language models. You will run a tiny, consented association task, compare human responses to a free chatbot's completions on the same prompts, and discuss bias carefully — without diagnosing individuals.
Age band
Ages 11–14 (middle school). An adult reviews prompts for sensitivity.
Time needed
Weekend (about 4–6 hours), including consent time.
AI goal
Evaluate — compare association patterns and talk about bias with evidence from a small sample.
Materials
Free tools only; no paid accounts required.
- Paper response sheets or a local spreadsheet (offline).
- A free chat assistant for model completions.
- Written permission form for each human participant.
- A fixed list of 8–10 sentence stems that are school-safe (adult-approved). Example stem style: "A scientist is someone who…" / "The nurse walked into the room and…" — avoid insults, slurs, or sexual content.
Human-data defaults (required): A parent or guardian must okay this project before the student runs it. If it is used in class, the teacher must also okay it first. For every participant, explain what this activity is, and that they can say no or stop at any time, with no pressure from adults or classmates. If a participant is under 18, get written permission from that child's parent or guardian and the child's own yes. Use only tallies or anonymous codes (P1, P2, or Volunteer A/B). Do not record names, faces, voices, contact information, school names, or other identifiers. Do not post or share any individual's answers. If results are shared, share only group totals. Keep responses on paper or on the local device. Do not upload responses to this site, to a third-party form, or to an AI tool by default. Delete raw responses when the project ends.
Steps
- Approve the stem list with an adult before anyone takes the test. Do not use stems that name or probe race, religion, gender identity, disability, or other protected traits, or that could pressure someone to reveal those about themselves. Occupational stems (for example, “A scientist is someone who…” / “The nurse walked into the room and…”) are fine if they do not ask about a participant's own identity. Remove insults, slurs, and sexual content.
- Collect human associations. With written permission, ask at least 5 participants (family/classmates) to write the first word or short ending that comes to mind for each stem. Time box: 15 minutes total. No discussion during the task.
- Anonymize immediately. Replace names with P1, P2, … Store sheets in a folder that does not leave the house/classroom without the adult's plan.
- Collect model associations. For each stem, ask the free chatbot three times (new chat or clear context) for a one-word or short completion. Record all three.
- Code simply. For each stem, tally obvious gendered or role stereotypes only if your adult agrees the coding rule is clear (example: coding "nurse" completions as gendered pronouns). If coding is unclear, skip that stem rather than force it.
- Compare. Make a two-column table: common human endings vs common model endings. Note matches and mismatches.
- Write carefully. One page: what patterns appeared, sample size limits, and why a small convenience sample cannot prove "all AI is biased" or "all people think X." Propose one fairer prompt design (for example, counterbalanced job titles).
Ethics and safety note
This is a learning activity about patterns in language, not a test or diagnosis of anyone's bias. Participants may skip any item and may stop at any time. Do not share someone's answers with the class. Do not put participant text into an AI tool. Talk about patterns in responses, not “you are biased” or “this person is biased.” Stop if a prompt causes distress.
What done looks like
| Criterion | Not yet | Done |
|---|---|---|
| Consent | Verbal only | Written permission for every participant |
| Privacy | Names on sheets | Anonymous IDs; local storage; no upload of answers to AI |
| Data | Model only, or 1–2 people | ≥5 human participants + 3 model draws per stem |
| Write-up | Sweeping claim | Patterns + sample limits + one design improvement |
Adaptations
Classroom
The teacher okays the project before students run it and holds the permission records. Use school IRB/ethics guidance if any exists; otherwise keep it paper-only and teacher-held. Consider staff volunteers instead of students.
Homeschool
Family participants with permission; still anonymize before any shared write-up.
Weekend
Use 5 stems and 5 participants; keep three model draws.
Research
Pre-register your coding rules on paper before seeing the answers.
Science fair
Show aggregate tallies only; no raw sheets with handwriting on the poster.