Privacy, bias and copyright — practical AI ethics for schools

On Thursday after school, the English panel head stared at her screen in the staff room. "This reading passage reads smoothly, but every nurse is a woman and every engineer is a man — twelve girls in my Secondary 2 class want to study engineering, and this material cannot go out." The student teacher beside her murmured, "I only asked the AI for three supplementary passages at the same level…"

Privacy, bias and copyright often sit in the same paragraph of a usage policy, yet each triggers different risks, different owners and different remedies. Lumping them together leaves teachers unsure where to look first. This lesson follows jailbreaks and prompt injection and turns ethics from slogans into classroom steps.

---

Three strands, three questions

A five-minute staff-meeting wrap-up — "watch privacy, avoid bias, respect copyright" — does not help when a teacher is standing at the desk wondering whether an SEN support summary may be pasted in, whether an AI circular raises copyright issues, or whether a student polishing sentences with AI counts as cheating.

Each strand asks something different. Privacy asks whether identifiable student data is being sent somewhere you cannot control. Bias asks whether the output quietly assumes some pupils matter less or fit a stereotype. Copyright and integrity asks whether input and output stay within existing licences, and whether pupils have declared how AI was involved.

A policy that cannot answer "may I paste this piece of work, where, and what do I delete first" is wall decoration.

Privacy: data minimisation

What to delete before send

Before pressing send, ask one question: is there a single word in this text that would let someone identify which student it is? If yes, and that word does nothing for the task, delete it. This is data minimisation — data never sent needs no rescue later.

In practice, five categories must never go into any AI tool: identity card numbers, home addresses and parent contact details, full SEN diagnostic reports and individual education plans, counselling case notes, and complete student files carrying name, class and student number. Compositions, answer scripts and presentation notes may be analysed on a school-compliant platform after removing name, class and student number. Curriculum outlines, past public exam papers and lesson plans without pupil data are generally lower risk, though textbook reproduction rights still apply.

One teacher pasted "Secondary 3 Class B, Chan XX, Mathematics Extended Part" together with a full answer script into a free chat tool, hoping for feedback. The problem was not laziness; the guidance had never listed five items that must never be pasted.

Three questions for any vendor

When choosing a platform, the IT team should get written answers: where data is stored, whether it is used for model training, and who decides how long it is kept. Vendors who cannot answer clearly should not handle student work. The Privacy Commissioner's model framework on artificial intelligence and personal data protection is a useful procurement checklist for Hong Kong schools.

Bias: visible in the classroom

Bias is not abstract. Four shapes appear often: default gender pairings for occupations, background assumptions tied to names from different ethnic groups, geographic and cultural defaults skewed towards English and Western content, and undervaluing unconventional but thoughtful writing during assisted marking. Lesson 9 explains where bias comes from; here we focus on catching it at the output end.

When generating materials

Specify explicitly in the prompt: "Hong Kong setting", "balance gender across characters", "use local festivals and food names". Fixing three passages afterwards costs more than writing the constraints once. After each batch, spot-check three items at random. If all three show the same tilt, that is systematic skew, not bad luck — revise the prompt or pause that use.

When assisting with marking

AI works best giving "three specific suggestions for improvement"; scores always come from the teacher. That design removes the step where bias risk is highest. In subjects that reward originality, a distinctive but coherent piece should not be marked down because the AI prefers familiar phrasing.

Copyright and academic integrity

Copyright in an AI context has three separate strands; discussing them together guarantees no conclusion.

First, copyright in training data. Models were trained on vast amounts of online content, and litigation and legislation continue to evolve. A school cannot resolve this strand and should not be the one making the judgement — choose vendors with clear terms and keep records.

Second, material the school puts in. Scanning a whole textbook and pasting it in is an ordinary reproduction question unrelated to AI. Textbooks, purchased item banks and licensed materials remain bound by their original licences.

Third, use of generated output. Internal teaching use is generally less controversial. External publication, posting on the school website or commercial use requires checking vendor terms and human review for protected song lyrics, poetry or exam passages embedded in the output. The Hong Kong Copyright Ordinance (Cap. 528) provisions on educational use are worth a read by panel heads and the librarian.

On academic integrity, policy should describe honest use, not only bans. Grade tasks: Level A — no AI at all; Level B — AI for ideas with declaration; Level C — AI required with process log submitted. A hundred-word note with the formal version — which suggestions were accepted, rejected and why — can itself be part of the assessment, turning "did you cheat" into "can you explain your thinking".

Three layers inside Edor.ai

The platform's position on all three strands is specific enough to verify. Privacy: input gating masks phone numbers, identity card numbers and addresses; student personal-data fields are encrypted at rest and only learning content is transmitted; conversations and AI logs purge on the school's retention schedule; provider calls run under no-training terms; data exports free at any time. Bias: marking feedback requires teacher review before anything is final; the scaffolded tutor gives hints not answers, with hint levels set by the teacher; every interaction writes to ai_audit_logs. Copyright: knowledge-base answers must carry citations; uploaded school material remains bound by its original licence. Further reading: security and privacy and the student AI safety net.

Where people go wrong

Treating encryption as the first line of defence. Once data is pasted into a consumer tool, you cannot control backend encryption or audit trails. Self-check: have all identifying fields been removed before send? If not, encryption cannot undo what was already transmitted.

Replacing spot-checks with "I'll be careful". Promises do not reveal systematic skew. Self-check: in the past month, have you randomly reviewed at least three AI-generated items? If not, bias may already be circulating in class unnoticed.

An integrity policy that only says "no AI". Banning without a compliant alternative pushes teachers and pupils into channels the school cannot see. Self-check: does guidance list a compliant platform, level definitions and a declaration format? If a colleague asks "so what may I do" and nobody can answer, the policy is not yet operational.

Try it once

What to do: Pick one AI task you will run next week — generating a reading passage, drafting feedback on a composition, or writing a parent circular — and spend ten to fifteen minutes on three steps.

Step 1 (3 minutes): Write the text you plan to paste in. Highlight every field that could identify a student and delete or replace each with "Student A".

Step 2 (5 minutes): Run the task. Check the output against four bias items — occupational gender pairing, name-based assumptions, geographic defaults, tone towards unconventional writing — and mark each "pass / revise / stop using".

Step 3 (2 minutes): If pupils will receive this work, write one sentence stating the AI use level (A/B/C) and what they must declare.

What to hand in: One A4 sheet: left column shows input before and after deletion; right column shows bias check results plus the AI use level.

How you know it worked: Input carries no identifiable student data; all four bias checks have a clear verdict; the AI use level matches your school or panel policy. If three or more checks say "revise", the prompt needs rewriting, not line-by-line manual fixing.

Next lesson tackles another topic often left vague — how to judge whether a model is genuinely good: automated evaluation and leaderboards.