L08A | English F3 Student Learning Log

Persona: I'm a Form 3 (Grade 9) student and I only read English. I know basic Python (variables, for loops, lists) and school algebra up to quadratics. I learn from the official pages, the videos with captions, and the book. Kaggle notebook steps are written as a practice log that follows the official notebooks.

What I understood (in my own words)

AI is powerful, and used in the wrong way it can hurt people, so model builders have to think about consequences first.
A feedback loop is 'the more it shows you, the more extreme it gets': you watch what the model shows, so the model thinks you want even more.
If some groups are missing from the data, the model will be less accurate for them. That's representation bias.
When an algorithm gets it wrong, someone must be able to appeal and someone must be responsible for fixing it.
Every dataset should come with a 'datasheet' saying where it came from and what it is and isn't suitable for.

Where I got stuck, and how I got unstuck

Where I got stuck	How I solved it
Some US legal cases had background I didn't know	I looked up the linked article on pre-trial detention and rewrote each case as a version that could happen at my school.
Recourse vs accountability	Recourse is the affected person's way to appeal. Accountability is who is responsible for the decision.

My answers to the exercises

Q1. Explain a 'feedback loop' in your own words and give an example from your own life.

A feedback loop is when a model's output changes the data collected next, and that new data pushes the model further the same way. Example: I watched a few funny cat videos on a short-video app; then my feed was all cats, I watched more, and it showed even more, until other topics almost disappeared.

Q2. If a face recognition system was trained mostly on photos of light-skinned men, what problems could happen? What kind of bias is this?

The system would make many more mistakes on darker-skinned or female faces, e.g. failing to recognise them or matching the wrong person. This is representation bias: some groups are under-represented in the training data.

Q3. Why must there be a way to appeal when an algorithm makes a decision about someone?

Algorithms always make some mistakes, and their decisions can affect someone's schooling, job or benefits. Without a way to appeal, a wrongly judged person can't get it fixed, the mistake continues, and nobody takes responsibility.

Q4. If you wrote a 'datasheet' for the image dataset you collected in L02, what are at least four things it should include?

Source (DuckDuckGo search, which keywords). 2) Collection date and size (images per class). 3) How labels were made (folder names, who checked them). 4) Known limits (e.g. few crushed cans, mostly product photos rather than real-life shots). 5) Suitable and unsuitable uses.

Q5. Find two ethical risks in the app you deployed in L02, and suggest one fix for each.

Risk 1: my recycling classifier was trained on clean product photos, so dirty or crushed items get misclassified and go in the wrong bin. Fix: add real canteen photos and show the confidence score. Risk 2: uploaded photos might include people's faces. Fix: state that photos aren't stored and ask users to photograph only the item.

My Zettelkasten card

Card｜Model builders must decide in advance who gets hurt when it fails

AI will always make mistakes. The real questions are who is affected and whether they can appeal. This lesson made me look at AI from the side of the person being judged, not just ask how to raise accuracy.

Our library tried an automatic system that flags overdue books and blocks borrowing. A friend returned a book on time, but the scanner missed it, and they were blocked for two weeks. Getting it fixed meant asking the librarian, who had to email IT. Without that manual process, the record would have stayed wrong. That's recourse and accountability in real life. I don't think 'it can fail' means we should never use automation, though. The point is to plan the fix before it fails.

My next step: write a half-page datasheet for my recycling classifier listing the data source, size and known limits.

🔗 Related notes to grow next: → Tech ethics: '95% accuracy means 5% of people are judged wrongly' → Product design: 'Every automated system needs a human override' → Media literacy: 'If a photo looks too perfect, ask whether it was generated'

Back to lesson note：L08A_資料倫理_202611041526