Key Idea: Stop asking whether a system understands. Ask what examples it was trained on, what it was asked to predict, and who is in the room when it is wrong — three answerable questions, each reaching a different failure.
Paper 1
- Part a: distinguish narrow from general AI, or outline machine learning
- Part b: explain why training data determines performance by group
- Part c: evaluate AI in decisions about people
Paper 2
- Q2: analyse a source's claim about an automated system
- Q3: compare sources on accuracy and explanation
- Q4: synthesise sources on hiring, scoring or diagnosis
Everywhere else
- Algorithms: a learned rule is one nobody wrote
- Identity: inferred attributes are model outputs
- HL 5.2: the four entry points structure every recommendation
Three questions, three different failures
| Question | What it reaches | What fixes it |
|---|---|---|
| What examples was it trained on? | Who it handles worst | Collect more, deliberately |
| What was it asked to predict? | Past patterns in the objective | Predict a different thing |
| Who is present when it is wrong? | How the output is used | Time and a real option to disagree |
The second failure survives perfect data. A system asked to predict who was hired learns the pattern of who was hired before, however representative the dataset is.
What changed, and what did not
Recent progress in one paragraph
- The methods are decades old. What changed is the availability of very large datasets and the computing power to train on them.
- Capability therefore concentrates where data and compute are held, rather than where the best ideas are.
- Nobody wrote the rule. A learned system derives its own, which is why responsibility became hard to locate.
- All deployed systems are narrow. They cannot transfer an ability to a task they were not built for.
Important: A system that formally proposes can decide in practice: if a person must justify every disagreement and has thirty seconds per case, the suggestion is being followed rather than considered.
Exam-style questions
Contrast a system built from written rules with one that learns its rule from examples.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
Examine the view that decisions affecting people should never be fully automated.
🔒 Model answer plan
See the mark-by-mark plan — for / against / judgement, with marking guidance — in study mode.
Quick check
The three questions What examples? What was it asked to predict? Who is present when it is wrong?
Why does more data not fix everything? It reaches representation only. The objective still encodes the past decisions the system learned from.
Why is capability concentrated? Progress came from data and compute rather than from new methods, and few organisations hold either at scale.
When does advice become a decision? When the reader has no time and must justify disagreeing.
Exam tips
- Never argue about whether a machine understands. Ask the three questions instead.
- 'Nobody wrote the rule' is why responsibility is hard to locate. Say it explicitly.
- Locate the propose-or-decide line before evaluating any system.