Chapter 2 · Cheat Sheet
Cheat Sheet: What is Artificial Intelligence?
Chapter 2 Cheat Sheet
Operational Definition
An AI system receives inputs and infers how to produce outputs such as predictions, content, recommendations, or decisions for objectives in a specific context.
Do not infer consciousness, intention, or human-like understanding from the label.
Six-Axis Classification
| Axis | Ask | Example value |
|---|---|---|
| Field | Where does it belong? | AI, NLP, computer vision |
| Method | How is behavior produced? | rules, search, optimization, ML |
| Model family | What representation is used? | decision tree, neural network, LLM |
| Capability | What output is produced? | predict, retrieve, generate, plan |
| Application | Which workflow uses it? | policy assistant, fraud triage |
| Authority | What may it affect? | inform, recommend, draft, decide, execute |
Terms That Are Not Synonyms
- AI: broad field and family of systems.
- Machine Learning: methods that fit behavior from data.
- Deep Learning: ML based on multi-layer neural networks.
- Foundation Model: broadly pretrained model adapted across tasks.
- LLM: large-scale language-centered model; many are Foundation Models.
- Generative AI: capability to construct a new artifact.
- Application: complete product and workflow around one or more models.
Optional Capabilities
An AI system may perceive, represent, infer, learn, plan, generate, act, or adapt. It does not need every capability.
Remember:
- symbolic AI can infer without learning,
- a learned model can remain frozen after deployment,
- a model can recommend without acting,
- application code grants authority.
System Contract
For which users, the system receives which inputs and authorized context, uses which inference method to produce which output, may influence which decision or action, and falls back to which safe behavior when evidence is insufficient.
Model Versus Application
| Model responsibility | Application responsibility |
|---|---|
| infer candidate output | authenticate users |
| represent learned patterns | authorize context and actions |
| expose model-specific signals | validate claims and citations |
| return output or error | handle fallback and escalation |
| identify model/version | evaluate, monitor, audit, and roll back |
Authority Ladder
Increasing authority generally requires increasing control:
- Inform: present retrieved or inferred information.
- Recommend: rank or suggest an option.
- Draft: prepare an artifact for review.
- Decide: select an outcome under policy.
- Execute: create an external side effect.
Do not let the model assign itself authority.
Failure Contract
For each material failure, define:
- detection signal,
- immediate bounded behavior,
- user-visible explanation,
- human or service owner,
- recovery or rollback evidence.
Useful behaviors include reject, abstain, ask for more evidence, use a deterministic fallback, route to a person, or reverse an action.
Architecture Review Questions
- What observable output does the system produce?
- Which evidence supports it?
- Which behavior is deterministic, searched, optimized, or learned?
- Where does the model boundary end?
- Which data and tools may the system access?
- What authority does the application have?
- What happens when evidence is absent, conflicting, stale, or unauthorized?
- Which test blocks release?
- Who owns quality, security, cost, and incidents?
- How is each component versioned and rolled back?
When Not to Use AI
Prefer deterministic software when a formula, policy rule, or query produces the authoritative answer; no reliable evaluation target exists; or the organization cannot own the data, outcome, and failure path.
Interview Answer Frame
- Define the system operationally.
- Classify it on the relevant axes.
- Separate model from application.
- State authority and failure behavior.
- Name one measurable trade-off and its owner.