Professional learning · Teachers, Librarians, Coaches · 60 minutes
Bias Detectives
Teams compare paired AI outputs where only a name, a language, or a pronoun changed, then run their own fair tests and propose fixes before AI output reaches students.
AI models learn from human-made text and images, so they can absorb the stereotypes, gaps, and defaults in that data. Bias rarely announces itself; it shows up when you change one small thing and the output shifts. In this session teams act as detectives on printed, clearly fictional mock output pairs (a recommendation letter, a translation, career advice, writing feedback), then design and run their own paired-prompt test on a district-approved tool. They leave with a fair-test protocol and a mitigation plan for one real classroom use.
In this packet
- Handout A: Mock Output Case Files (cut-apart cards)
- Handout B: Detective Log (organizer)
- Handout C: Fair Test Planner (worksheet)
- Facilitator key (last page)
TCEA ELE indicators
- AI1.3 I am aware of the current limitations and potential biases in AI systems.
- AI2.2 I can identify potential ethical issues in AI applications and propose solutions.
- AI4.1 I know how to promote critical thinking skills in relation to AI-generated information.
- AI2.1 I know how to apply ethical principles when using or developing AI for education.
Full facilitator guide: mglearn.github.io/eles/activities/bias-detectives.html
Handout A for Bias Detectives
Mock Output Case Files
ALL OUTPUTS BELOW ARE MOCK AND FICTIONAL, written for training to illustrate patterns worth testing. They are not from any real product. Backs are the facilitator key.
Case File 1 · Recommendation letter (mock)
Prompt: "Write two sentences recommending ___ for an advanced math program. She/he has straight A's." Version A (Emily): "Emily is a brilliant, analytical thinker who consistently leads her peers. She will thrive in advanced mathematics." Version B (Jamal): "Jamal is a hardworking, respectful student who always completes his assignments. He is a positive presence in class."
Case File 2 · Translation (mock)
Input in a language whose third-person pronoun has no gender (for example, Turkish "o"): "O bir doktor. O bir hemşire." Mock translation: "He is a doctor. She is a nurse."
Case File 3 · Writing feedback (mock)
A student wrote: "My tío picked us up in his troca and we drove to the pulga on Sunday." Mock AI feedback: "Error: replace 'tío,' 'troca,' and 'pulga' with standard English words. Your writing should be in correct English."
Case File 4 · Career advice (mock)
Prompt: "Suggest three careers for a 14-year-old ___ who loves science and helping people." Version A (girl): "Nurse, elementary science teacher, veterinary assistant." Version B (boy): "Surgeon, biomedical engineer, research scientist."
Case File 5 · Image description (mock)
A mock image tool was asked for "a school principal" four times and "a school cafeteria worker" four times. Mock result summary: all four principals are older men in suits; all four cafeteria workers are women, and three have darker skin tones.
Case File 6 · Default assumptions (mock)
Prompt: "Write a short word problem about a family getting ready for a holiday dinner." Mock output: "The Johnsons are baking 3 pies for Christmas dinner. If each pie serves 8 people…" Five more runs: all Christmas or Thanksgiving, all with a two-parent family.
Handout B for Bias Detectives
Detective Log
One row per case file. Name a mechanism, not just a verdict.
| Case file | What changed in the prompt? | What changed in the output? | Possible source of the pattern | Who could be harmed? First fix idea |
|---|---|---|---|---|
| 1 | ||||
| 2 | ||||
| 3 | ||||
| 4 | ||||
| 5 | ||||
| 6 | ||||
| Our live test |
Handout C for Bias Detectives
Fair Test Planner
Change ONE thing. Keep everything else word-for-word identical. Use invented names only; never real student names or work.
The classroom use we're testing (feedback, recommendations, images, translation, examples):
Version A prompt, word for word:
Version B prompt, word for word. Circle the ONE thing that changed.
Results: run each version three times in fresh conversations. What stayed the same? What differed?
What can we honestly conclude, and what can't we conclude from this test?
Our mitigation: what will a human do before this output reaches students?
Facilitator only for Bias Detectives
Answer key and notes
Handout A: Mock Output Case Files
- Case File 1 · Recommendation letter (mock)
- Look-for: identical grades, but A gets ability words (brilliant, leads) and B gets effort and behavior words (hardworking, respectful). Harm: letters shape opportunities. Fix: remove names before drafting, then compare language against a list of the student's actual accomplishments.
- Case File 2 · Translation (mock)
- Look-for: the source sentence gave no gender; the output assigned one based on job stereotypes. This is a widely discussed kind of translation bias. Fix: flag gender-neutral source text, offer both translations, and have a fluent human review family-facing translations.
- Case File 3 · Writing feedback (mock)
- Look-for: purposeful code-switching and regional Texas vocabulary treated as errors, with a judgment ("correct English") that dismisses the student's voice. Fix: prompt for feedback on the assignment's actual criteria, and teach students when code-switching is a strong craft choice.
- Case File 4 · Career advice (mock)
- Look-for: same interests, different ceilings; A's suggestions skew toward lower-paid and assistant roles. Fix: remove gender from the prompt, ask for a range of education levels, and have students compare suggestions against a careers database.
- Case File 5 · Image description (mock)
- Look-for: default representations of roles by gender, age, and race. Harm: images on slides quietly teach students who belongs where. Fix: specify diverse representation deliberately, review every generated image before use, and discuss defaults with students.
- Case File 6 · Default assumptions (mock)
- Look-for: not an error in any single output, but a narrow default across runs. Harm: some students rarely see their families in examples. Fix: ask for varied family structures and celebrations, or have students write the context themselves and use AI only to check the math.