Chapter 2 · Cheat Sheet
Cheat Sheet: What is Artificial Intelligence?
Cheat Sheet - What Is Artificial Intelligence?
Three definitions, three purposes
| Who | Definition | Minimax chess engine | Hand-coded rules |
|---|---|---|---|
| Research field (Dartmouth 1955; Russell and Norvig) | any machine simulation of learning or intelligence: search, logic, planning, learning, perception | AI | usually not |
| OECD (Nov 2023) and EU AI Act Art. 3(1) | a machine-based system that infers, from input, how to generate outputs that can influence environments; autonomy varies, adaptiveness optional | likely outside (Commission guidelines, 6 Feb 2025: classical heuristics) | outside (Recital 12) |
| Marketing | whatever sells | "AI-powered" | often "AI-powered" |
A requirement or policy should name which definition it uses. The SEC fined Delphia $225,000 and Global Predictions $175,000 (18 March 2024) for AI claims they could not substantiate.
EU AI Act dates (as of September 2026)
In force 1 Aug 2024 · prohibitions 2 Feb 2025 · general-purpose AI models 2 Aug 2025 · transparency (Art. 50) 2 Aug 2026 · Annex III high-risk 2 Dec 2027 (moved by the Digital Omnibus on AI, Reg. (EU) 2026/1744, in force 27 Jul 2026) · Annex I product-embedded high-risk 2 Aug 2028.
Six axes
| Axis | Question | Examples |
|---|---|---|
| Field | Which area of computing? | NLP, vision, robotics |
| Method | How is behaviour produced? | rules, search, optimisation, machine learning |
| Model family | Which model is inside? | decision tree, neural network, foundation model, LLM |
| Capability | What kind of output? | classify, predict, retrieve, generate, plan, act |
| Application | Which product surrounds it? | spam filter, code editor, ride-hailing |
| Authority | What may an output do? | inform, recommend, draft, decide, execute |
Not synonyms: machine learning is a method; an LLM is a model family; generative AI is a capability. Two meanings of inference: deriving outputs from inputs, and running a trained model (as opposed to training it).
Authority levels
| Authority | Sheridan–Verplank | Example |
|---|---|---|
| Inform | 2–3 | search results |
| Recommend | 4 | credit score shown to a loan officer |
| Draft | 5 | Copilot ghost text, accepted with Tab |
| Decide | 6 | paid automatically unless vetoed in time |
| Execute | 7–10 | Gmail spam filter (reversible); Waymo Driver (SAE Level 4) |
Raise authority with reversibility, speed of detection and low cost per mistake, never with capability alone.
Four real systems
| System | Model family | Authority |
|---|---|---|
| Gmail spam filtering (RETVec: +38% detection, −19.4% false positives, Nov 2023) | neural classifier | execute, reversible |
| GitHub Copilot inline suggestions | LLM | draft |
| Waymo Driver (SAE Level 4; fleet response advises, never drives) | neural networks + planning | execute, within an operating domain |
| Air Canada chatbot, 2022 | not disclosed | designed to inform; treated as the airline's promise (C$812.02, Feb 2024) |
Review decision
C_none = N e a L
C_full = N r_c + C_none (1 - k)
C_routed = N s r_c + C_none (1 - k q)
full beats none when e > r_c / (a L k)
routed beats none when e > s r_c / (a L k q)
full beats routed when e > (1 - s) r_c / (a L k (1 - q))
With a = 10%, L = C$812.02, r_c = C$1.00, k = 90%, s = 10%, q = 80% (illustrative except L): none below 0.17%, routed from 0.17% to 6.16%, full above 6.16%; full beats none at 1.37%. Volume N cancels. Showing e < 0.17% at 95% confidence takes about 1,750 clean answers.
Klarna, for calibration
Feb 2024: 2.3 million conversations in month one, two-thirds of chats, ~700 agents' work, under 2 minutes instead of 11, ~US$40M profit improvement estimated. May 2025: CEO said cost had been "a too predominant evaluation factor" and quality suffered; humans brought back.