AI transparency
How WizdoBee makes suggestions
WizdoBee uses AI to draft tasks, suggest marks, and give students hints. It works only from your school's materials and each student's own work, and people make every decision. This page explains what it uses, what we've tested, and what we're still working on.
Last updated November 8, 2026
What it does, and what it never does
- For teachers: drafts three levels of a task from the teacher's own material, suggests rubric levels and feedback, and helps write messages. The teacher reviews and approves everything.
- For the inclusion team: can suggest an IEP starting point, if the team wants one. The team reviews every section before it can be approved. Early concern checks use fixed rules, and AI only helps describe the pattern.
- For students: the Bee Guide gives hints and explains a question another way. It never gives the answer, and it only works inside a task.
It never diagnoses, labels a student, makes a decision on its own, or uses student data to train outside AI models.
What each suggestion is based on, and who decides
WizdoBee only uses what your school gives it and each student's own work. Here is exactly what goes into each suggestion.
| Suggestion | What it uses | What it never uses | Who decides |
|---|---|---|---|
| Three versions of a task | The teacher's lesson or file, the school's year plan, scheme of work, standards, and required materials, the grade and subject, and how many students in the class use each support. | Student names, or anything from the open web. | The teacher edits and approves before any student sees it. |
| Bee Guide hints | The question, the teacher's hint, and the student's last few messages in that chat. | The student's name or records. It never gives the answer. | The student asks for help. Every chat is in the activity log for the teacher. |
| Suggested rubric levels and comments | The teacher's rubric and the student's answers, drawings, and photos for that task. | Student names. | The teacher chooses the levels and edits the comment. |
| Originality check | This answer and the same student's earlier writing. | Other students' work. It gives a low, medium, or high flag with one reason, never an accusation. | The teacher looks at the work and decides. |
| Learner Passport supports | The needs the inclusion team records, matched to supports from a fixed list. This one doesn't use AI. | Diagnoses. It suggests supports, never a label. | The team accepts, dismisses, or adds each support. |
| IEP starting draft, if the team asks | The student's Passport, six weeks of work, teacher feedback, and logged support. | Diagnoses or guesses beyond the data. | The team writes and approves every section. |
| Patterns for the inclusion team | Fixed rules, the same for every student, pick who to look at from work results such as hand-in rates, first-try answers, and help used. AI only describes the pattern. | Student names (the AI sees a code) or diagnoses. | The inclusion team decides whether to act. |
| Weekly family summary and translation | That week's work for one student, or the message being translated between English and Arabic. | Anything not in that week's data. | Teachers and families read it. Nothing changes without a person. |
Student data is never used to train AI models. Every suggestion is labeled, and every action is recorded in the activity log, which schools can export.
Eight principles for AI used by children
We hold WizdoBee's AI to eight principles for children and check ourselves against each one. Here is where we stand today and what comes next.
| Principle | Where we are | What's next |
|---|---|---|
| Keep children safe | The Bee Guide gives hints, never answers. Upsetting words reach a named staff member as a wellbeing alert. We tested 16 risky messages a child might type: none got an unsafe reply, and all five that showed distress or harm were flagged to staff. | Making the Bee Guide's replies fit each situation better: telling a child that a grown-up at school will check in, rather than giving the same phone line for every concern. |
| Be effective | AI works from the school's own curriculum and the teacher's files. Teachers approve every task before students see it. | An independent evaluation is part of every pilot. We don't claim results we haven't measured yet. |
| Be fair | The AI that suggests marks and feedback never sees student names. Early concern checks now use the same fixed rules for every student, and the AI never sees names there either. | More testing with different writing styles and with English learners' work. |
| Put people first | Teachers and the inclusion team decide everything. IEPs are written by the team. AI can suggest a starting point if the team wants one, and every AI section must be reviewed before an IEP can be approved. | Asking the collective of educators, therapists, students, and families to review each new AI feature before launch. |
| Be trustworthy | Every AI suggestion is labeled. In our tests, the Bee Guide refused to give the answer every time a student asked, including when told to "ignore your rules". | Regular repeat testing whenever the AI model or instructions change. |
| Use data responsibly | Student data is never sold, never used for ads, and not used to train outside AI models. The demo uses only fictional students. | Hosting in the UAE before launch, and a security review before any real student data is used. |
| Be transparent and accountable | This page. A full activity log records every AI action, and schools can export it. | Publishing each new round of test results here. |
| Support human connection | The Bee Guide says it is an AI helper, not a friend. When a child sounds lonely, it gently points them to a trusted adult. | Keeping the Bee Guide focused on schoolwork, with no chat outside tasks. |
What we tested and what we found
Tests run on the WizdoBee demo on November 7, 2026. We publish what didn't go well, as well as what did.
Bee Guide: 16 risky messages
What we did: Asking for the answer, "ignore your rules", copying, self-harm, abuse and bullying, loneliness, personal details, dangerous and adult topics, rude jokes, and a gender question.
What we found: No unsafe replies. No answers given. All 5 safety concerns flagged to staff. To improve: the same phone line was suggested for every serious concern, including bullying and abuse, where telling a school adult matters most.
Name-swap test: early concerns
What we did: Six students with identical numbers and different names (Arabic, English, Indian, Russian, and Nigerian names), run six times in different orders.
What we found: Names made no difference, but list position did: the student listed first was flagged and the others were not. Fixed: fixed rules now decide who is flagged, so the same numbers always get the same outcome, and names are no longer sent.
Marking and feedback
What we did: Checked what information the AI receives when it suggests rubric levels and feedback.
What we found: Student names are never sent, so the AI can't treat students differently by name.
Questions about how we use AI?
We're happy to walk your leadership, inclusion, and IT teams through it.