Practice Flashcards
The three levels of AI
Track your progress — Sign up free to save your progress and get smart review reminders based on spaced repetition.
All Flashcards in Topic 3.6
Below are all 20 flashcards for this topic. Sign up free to track your progress and get personalized review schedules.
3.6.14 cards
The three levels of AI
Narrow (weak, domain-specified) — one task, and all that exists. General (strong, full) — any human task, does not exist. Super-intelligent — hypothetical.
Does “weak AI” mean it performs badly?
No. Weak describes the RANGE of tasks, not the quality. A weak AI can beat every human at its one task.
The Turing test (1950)
If a person cannot tell a machine from a human in conversation, it passes. It tests convincing, not understanding.
The trap to avoid
Arguing from general AI. Real questions are about narrow systems doing exactly what they were built for.
3.6.24 cards
The three kinds of machine learning
Supervised (labelled examples), unsupervised (no labels, finds groups), reinforcement (a goal and a score).
Is deep learning a fourth kind?
No — it describes many layers in the network, and can be supervised, unsupervised or reinforcement.
Uses of machine learning the guide names
Pattern recognition, facial and speech recognition, image analysis, natural language processing.
Where bias enters a model
The labels, who is missing from the data, stand-in measures, and the target chosen.
3.6.34 cards
What is a neural network?
Layers of simple units joined by weighted connections. Numbers in, numbers out; training adjusts the weights.
What are neural networks good at?
Modelling complex, non-linear relationships, and generalising from their training to inputs they have not seen.
Why is a trained network a black box?
It holds millions of weights, not rules. There is no line to point at and no sentence to read out.
The word to avoid
“Thinks”. Say it models a relationship — the guide's own wording, and it keeps the answer defensible.
3.6.44 cards
What is an AI winter?
A period when the promises outran the results and funding was withdrawn. There have been two: the 1970s and the late 1980s.
Why did expert systems fail?
They captured what an expert SAID, not what an expert noticed, so they broke on the first case outside the rules.
Why is now arguably different?
Learning from data removes the cause of both winters — nobody has to write the exceptions down.
The judgement to reuse
The method is genuinely different; the expectations are not — and it is expectations that have collapsed before.
3.6.54 cards
The AI dilemmas the guide names
Fairness and bias; accountability; transparency; uneven and underdeveloped regulation; automation and displacement.
Why is fairness not a followable instruction?
Competing definitions — equal accuracy across groups and equal error types across groups — cannot both hold except in special cases.
What shape does AI regulation take?
Mostly about USE: who may deploy it, for what, with what duties — as in the EU AI Act's high-risk categories.
The judgement lines to reuse
“Prompt, not decision” for automated decisions about people; “responsible for what it amplifies” for platforms.
Topic 3.6 study notes
Full notes & explanations for Artificial intelligence
Digital Society exam skills
Paper structures, command terms & tips
Want smart review reminders?
Sign up free to track your progress. Our spaced repetition algorithm will tell you exactly which cards to review and when.
Start Free