Professional learning · Leaders, Librarians, Coaches, Teachers · 60 minutes
The Glass Box Checklist
Teams interrogate a fictional AI vendor's glossy fact sheet, sort the questions that actually matter, and build a transparency and accountability checklist they can use on real tools.
AI tools arrive in schools with confident marketing: "safe," "unbiased," "research-based," "personalized." Transparency means being able to see behind those words: what the tool does, what data it uses, where it fails, and who answers for it when it's wrong. In this session teams read a fictional vendor fact sheet, sort a bank of evaluation questions, and build a short checklist that turns vague reassurance into answerable questions. Then they test the checklist on the published documentation of a tool their district already uses or is considering.
In this packet
- Handout A: Mock Vendor Fact Sheet: BrightPath Writing Coach (reading)
- Handout B: Question Sort (card sort)
- Handout C: Glass Box Starter Checklist (checklist)
- Facilitator key (last page)
TCEA ELE indicators
- AI2.3 I am aware of the importance of transparency and accountability in AI systems.
- AI2.1 I know how to apply ethical principles when using or developing AI for education.
- AI1.3 I am aware of the current limitations and potential biases in AI systems.
Full facilitator guide: mglearn.github.io/eles/activities/glass-box-checklist.html
Handout A for The Glass Box Checklist
Mock Vendor Fact Sheet: BrightPath Writing Coach
FICTIONAL PRODUCT AND COMPANY, written for training. Highlight checkable claims in one color and vague claims in another.
1BrightPath Writing Coach: Every student, a confident writer. BrightPath uses advanced artificial intelligence to give every student instant, personalized feedback on their writing, 24 hours a day. Teachers save hours every week.
2How it works. Students paste or type their writing. Our proprietary AI engine analyzes it and returns feedback on organization, evidence, grammar, and style, plus a predicted rubric score. The feedback is generated by a large language model licensed from a third-party AI provider and customized with BrightPath's writing expertise.
3Research-based. BrightPath is grounded in the science of writing instruction and has been proven to accelerate student growth.
4Safe and fair. BrightPath is bias-free and appropriate for all learners, including English learners and students with disabilities. Our AI never makes mistakes that a teacher would not catch.
5Your data. Student writing is stored securely in the cloud. We may use de-identified student writing to improve our services and may share data with trusted partners. Accounts can be deleted by contacting support.
6Teacher control. Teachers can view all student feedback in the dashboard. Predicted scores can be synced automatically to your gradebook.
7Ages. Designed for grades 3–12. Setup takes under 5 minutes; students sign in with their school account.
Handout B for The Glass Box Checklist
Question Sort
Cut apart the strips. Sort each question into the category where it fits best, then star your five non-negotiables.
How it works and its limits
Data and privacy
Accountability and recourse
Evidence of learning
Handout B for The Glass Box Checklist
Question Sort: cards to sort
Cut apart the strips. Sort each one onto the mat and be ready to explain why.
Sheet 1 of 2
What kind of AI produces the output, and does the vendor say which parts are generated versus rule-based?
What known limitations or failure cases does the vendor publish?
Has the tool been tested for how well it works for English learners, students with disabilities, and different dialects?
Does the tool tell students when content is AI-generated?
Is student work used to train or improve AI models, and can the district opt out?
Who are the "trusted partners," and what do they receive?
How long is student data kept, and is it deleted automatically when the contract ends?
Does the tool meet the district's data privacy agreement requirements for student records?
Who is responsible when the feedback or predicted score is wrong: the teacher, the district, or the vendor?
Can a teacher review and override every score before it reaches the gradebook?
How can a student or family question or appeal an AI-influenced decision?
How does the vendor notify schools of changes to the AI model or data practices?
Handout B for The Glass Box Checklist
Question Sort: cards to sort
Cut apart the strips. Sort each one onto the mat and be ready to explain why.
Sheet 2 of 2
What independent evidence shows students' writing improves, not just that they used the tool?
What evidence will our own teachers collect to decide whether it's helping?
Handout C for The Glass Box Checklist
Glass Box Starter Checklist
Cross out, add, and rewrite items so every one can be answered from documentation or a vendor reply. Note where you found each answer.
| Yes | Partly | No / Not found | |
|---|---|---|---|
| The vendor states plainly what the AI does and does not do. | |||
| The vendor publishes known limitations and failure cases. | |||
| Performance has been checked across student groups (language, disability, dialect), with results shared. | |||
| Students can tell when content is AI-generated. | |||
| Student data is not used to train AI models, or the district can opt out. | |||
| All third parties that receive student data are named. | |||
| Data retention and deletion timelines are stated. | |||
| The tool is covered by a district data privacy agreement. | |||
| A teacher reviews and can override AI outputs before they affect grades or placement. | |||
| Responsibility for errors is stated, and students or families have a way to question outputs. | |||
| The vendor notifies schools of model or data-practice changes. | |||
| There is evidence, beyond usage data, that the tool improves learning, and we have a plan to collect our own. |
Notes
Facilitator only for The Glass Box Checklist
Answer key and notes
Handout B: Question Sort
| Item | Sort | Why |
|---|---|---|
| What kind of AI produces the output, and does the vendor say which parts are generated versus rule-based? | How it works and its limits | Knowing it's a large language model tells you to expect fluent but sometimes wrong feedback. |
| What known limitations or failure cases does the vendor publish? | How it works and its limits | A vendor that names limits is more trustworthy than one claiming none. |
| Has the tool been tested for how well it works for English learners, students with disabilities, and different dialects? | How it works and its limits | "Bias-free" is not a testable claim; results by student group are. |
| Does the tool tell students when content is AI-generated? | How it works and its limits | Users can't apply judgment to AI output they don't know is AI output. |
| Is student work used to train or improve AI models, and can the district opt out? | Data and privacy | "Improve our services" can include model training. |
| Who are the "trusted partners," and what do they receive? | Data and privacy | Unnamed sharing is a transparency gap. |
| How long is student data kept, and is it deleted automatically when the contract ends? | Data and privacy | "Contact support" is not a retention policy. |
| Does the tool meet the district's data privacy agreement requirements for student records? | Data and privacy | District agreements set enforceable rules consumer terms don't. |
| Who is responsible when the feedback or predicted score is wrong: the teacher, the district, or the vendor? | Accountability and recourse | Accountability must be named before something goes wrong. |
| Can a teacher review and override every score before it reaches the gradebook? | Accountability and recourse | Automatic sync removes the human decision point. |
| How can a student or family question or appeal an AI-influenced decision? | Accountability and recourse | Recourse is what makes transparency meaningful to the person affected. |
| How does the vendor notify schools of changes to the AI model or data practices? | Accountability and recourse | Models get updated; a tool approved last year may behave differently now. |
| What independent evidence shows students' writing improves, not just that they used the tool? | Evidence of learning | Usage minutes are not evidence of learning. |
| What evidence will our own teachers collect to decide whether it's helping? | Evidence of learning | Local evidence matters more than marketing claims. |