AI ethics
AdvancedFour gates of AI ethics: bias (skewed data breeds discrimination), privacy (faces and voices are personal data under the PDPO), deepfakes (synthetic voices and faces used in scams — verify by other channels), responsible use (AI may assist learning, never ghost-write, and must be declared).
- Biased data, biased AI: audit the data spread, top up every group, monitor outputs.
- Faces, voices, whereabouts are personal data under the PDPO; get consent before using anyone's.
- Deepfakes fake faces and voices: call back to verify, agree a code word, count fingers and match lips.
文體小知識：時事議題課結合新聞案例和實用建議，學完可以即刻應用在日常生活，保護自己。
§Text Passage
1AI learns from data, so bias in the data becomes bias in the AI. A real case: a multinational's recruitment system trained on a decade of CVs learned, because past hires were mostly men, the pattern 'male CVs score better' and automatically downgraded women applicants. Algorithmic bias is not AI malice but unrepresentative training data — remedies are auditing the data spread, adding examples from every group, and monitoring outputs continuously.
2AI feeds on mass data, and faces, voices, whereabouts and shopping habits are all personal data. Hong Kong's Personal Data (Privacy) Ordinance (PDPO) requires: state the purpose when collecting, collect no more than needed, never stray from the stated use, and give people the right to access and correct. When an app asks for contacts or location, ask: does this feature truly need it? Uploading classmates' photos to an AI tool can breach their privacy too.
3Deepfake technology generates convincing fake video and voices: swapped faces, imitated voices. Scammers already call elders in a synthesised 'child's voice' to steal money. Defences: on any distress call, ring the person back through another channel; agree on a family code word only insiders know; watch for artefacts — finger count, spectacle glare, lips not matching the sound.
4For students, responsible use has a clear line: AI may explain concepts, suggest improvements and spark ideas, but handing in AI-generated work as your own teaches nothing and breaks academic honesty. The right way: think first yourself, use AI to check and inspire, verify everything, and declare AI assistance honestly. AI is the study partner, never the ghostwriter.
1Summary 重點摘要
Four gates of AI ethics: bias (skewed data breeds discrimination), privacy (faces and voices are personal data under the PDPO), deepfakes (synthetic voices and faces used in scams — verify by other channels), responsible use (AI may assist learning, never ghost-write, and must be declared).
2Vocabulary cards Vocabulary
Systematic unfairness in an AI caused by unbalanced training data.
Data identifying an individual, e.g. name, facial image, voice, address, phone number.
Hong Kong's law on collecting and using personal data: state the purpose, collect minimally, access and correct.
AI-generated or altered media that passes as real — swapped faces, imitated voices.
Content that is untrue; AI hallucinations and deepfakes both produce it.
The principle of declaring sources honestly and never passing off others' (or AI's) work as your own.
Voluntary permission after the person knows the use; uploading someone's photo without it breaches their rights.
AI systems should disclose their data sources, workings and limits to users.
When AI causes harm, a clear person or body must answer for it — never 'the AI did it'.
3Step by step Steps
- How bias gets learnedAlgorithmic bias means an AI systematically treats some groups unfairly because its training data was unbalanced. Example: in a hiring system's training data, 800 of 900 successful CVs were male, so the model learned the pattern 'male markers correlate with hiring' and systematically scored women down. The key insight: nobody wrote the malice in — the AI induced it from skewed historical data, which is why 'AI has no feelings' guarantees nothing about fairness.
- Audit and repair the training dataThe answer to bias is deliberate human auditing before and after training. Before: inspect the dataset spread — gender, age bands, ethnicities — and top up any under-represented group until each has fair coverage. After: test accuracy group by group, checking for instance that male and female CVs receive similar scores; if a clear gap remains, adjust the data and retrain.
- Three privacy questionsPersonal data is information identifying an individual — name, address, phone number, facial image, voice — governed in Hong Kong by the Personal Data (Privacy) Ordinance. To judge whether an app deserves trust, ask three questions: what data does it collect; is the stated purpose reasonably related to its function; can I refuse to provide it? A torch app demanding your contacts fails all three, so switch to another app.
- Four deepfake checksDeepfake technology synthesises convincing faces and voices with AI. Four checks for a suspicious video: count the fingers, since AI portraits often err there; inspect reflections, whether glasses and metal surfaces mirror a plausible scene; match lip movement to the sound; and verify the event through an official channel. For a video announcing 'the principal closes school', confirm with the school's website or a teacher before forwarding.
- Drilling for voice scamsWith a few seconds of sampled audio, scammers can clone a relative's voice. Drill the scenario: a call from your 'son' in tears, begging for an urgent transfer. The correct routine: stay calm and hang up, never redial the number shown on the call; ring the number stored in your contacts to reach the real person; and use a family code word agreed in advance — a caller who cannot answer it is exposed. Rule: every transfer request gets confirmation through a second channel.
- The line for AI in homeworkIn homework, AI assists rather than ghost-writes. Allowed: asking it to 'explain Newton's third law at S3 level', 'check this paragraph's grammar' or 'comment on my report outline' — these aid understanding and revision while the learning stays yours. Not allowed: having AI 'write the whole book report' and submitting it as your own, which is ghost-writing and breaks academic honesty. The test: the core thinking and analysis must be yours, with AI only checking and suggesting.
- Declare use and verify factsResponsible use carries two duties. First, declaration: if AI helped with an outline or polishing, state it honestly at the end, such as 'this outline was prepared with AI assistance', so the teacher understands the creative process. Second, verification: AI hallucinates, inventing professional-looking figures, book titles and events, so every fact and number it supplies must be checked item by item against textbooks, library catalogues or official websites before entering your report.
4Diagrams Diagrams
5Worked example Example
Example: four students use AI in a group report — judge which way is responsible and why. A prints the AI-written report as their own; B drafts first, then uses AI to spot errors and suggest improvements, verifying and revising personally; C inserts an AI-fabricated 'field photo' as real evidence; D has AI write the whole report and submits without reading it.
- A: AI ghost-writes and passes as their own — breaks academic honesty, no learning happened → irresponsible.
- B: thinks and drafts personally, uses AI to check and suggest, verifies and revises personally — AI assists while the learning stays theirs → responsible.
- C: passes off a fabricated image as real evidence — manufacturing fake data, worse than copying → irresponsible.
- D: submits unread — besides ghost-writing, no one can answer for the truth of content the AI may have hallucinated → irresponsible.
- Answer: Only B is responsible. The principle: AI may assist thinking and checking, but thinking, verifying and answering must stay human; AI content is never your own work, and fabrication is worse still.
6Common mistakes Mistakes
7Quick check Quick check
Q1What is the root cause of a hiring AI systematically downgrading women applicants?
Q2Which Hong Kong law governs collecting and using personal data? Name two of its requirements.
Q3A 'relative's' voice calls asking for a bank transfer — what is the right response?
Q4Which use of AI in homework is responsible? A. having AI write the report and submitting it B. drafting yourself, then using AI to check and suggest, verifying and revising
Q5Name two ways to spot a deepfake video.
8Recap Recap
- Biased data, biased AI: audit the data spread, top up every group, monitor outputs.
- Faces, voices, whereabouts are personal data under the PDPO; get consent before using anyone's.
- Deepfakes fake faces and voices: call back to verify, agree a code word, count fingers and match lips.
- The homework line: AI may explain, suggest and brainstorm; ghost-writing breaks academic honesty.
- Declare AI use, verify its facts; the final responsibility is always human.