Skip to content
AI.info

AI literacy basics

AI Literacy: A Map of the Field

Build a durable mental map of AI that connects technical components, human decisions, product workflows, and real-world effects.

By the end you can

Key idea

The first mistake: treating AI as one thing

“AI” can refer to a research field, a mathematical model, a product feature, a business process, or a public policy problem. Those meanings overlap. They are not interchangeable.

A language model may be one component inside a support assistant, and the assistant also includes retrieval, permissions, interface design, escalation rules, logs, and people who remain accountable. Judge only the model and you miss most of the system.

AI literacy begins when you stop asking only “How smart is the model?” and start asking “How does the whole system work here?”

Predict first

Your call, before the verdict

An airline’s website chatbot tells a grieving passenger he can buy his tickets at full fare now and claim the cheaper bereavement fare afterwards. A policy page on the same website says the opposite. He relies on the chatbot, buys, applies, and is refused.

Who do you expect a tribunal to hold responsible: the passenger, who could have read the policy page; the software that wrote the answer; or the airline? And what defence would you expect the airline to try?

Settle on an answer, then read on — the next section is the reveal

Case

A chatbot’s promise, and who answered for it

That question has been tested in a tribunal. On 14 February 2024, British Columbia’s Civil Resolution Tribunal decided Moffatt v. Air Canada (2024 BCCRT 149), after the airline’s website chatbot told a bereaved passenger he could claim a bereavement fare retroactively — advice that a policy page on the same website contradicted. Air Canada submitted that the chatbot was “a separate legal entity that is responsible for its own actions.” Tribunal member Christopher C. Rivers called this “a remarkable submission”: a chatbot “is still just a part of Air Canada’s website,” and the airline was liable for negligent misrepresentation. The model was never a party to the case; the operator was.

Visual

Four levels you must keep separate

The course uses four levels because many arguments become confused when they jump between them without anyone noticing the jump. A claim can be true at one level and false at another. Both speakers can be right.

FigureHierarchy · 4 levels

  • Field

    Artificial intelligence is a broad area of research and engineering with several traditions, methods, and goals.

    • Model

      A model transforms inputs into outputs. It may classify, rank, predict, retrieve, generate, or help choose an action.

      • System

        A system wraps one or more models with data pipelines, software, policies, interfaces, safeguards, and monitoring.

        • Context of use

          People, institutions, incentives, law, culture, and stakes determine whether the system creates benefit or harm.

A strong model can still produce a poor system, and a modest model can create value inside a well-designed workflow.

Three questions that organize AI literacy

First, what output does the system produce: a label, score, ranking, forecast, recommendation, generated artifact, or action? Second, what evidence and assumptions support that output? Third, what happens when the output enters a real workflow?

They work on a spam filter, a medical image model, a recommendation engine, or a generative assistant. They also keep the conversation concrete. A vague argument about whether a machine “understands” something never takes the place of those three answers.

Output, evidence, and workflow form the backbone of this path.

Comparison

Capability, usefulness, and trustworthiness are different tests

A demo often answers only the first test. A serious product decision must answer all three.

Take one example, a speech-to-text tool. Koenecke and colleagues evaluated OpenAI’s Whisper and reported the result in June 2024. In their evaluation, “roughly 1% of audio transcriptions contained entire hallucinated phrases or sentences which did not exist in any form in the underlying audio,” they write. Of those hallucinations, 38% carried explicit harms. Invented violence and implied false authority were among them. Most transcriptions contained no invented passage at all, so the capability claim survives. Whether the same system is trustworthy in a clinical or legal recording is a separate question. It depends on what that 1% does after it leaves the model.

FigureComparison · 3 columns

Capability

Can the system perform the target task under stated conditions?

  • Measured with task-specific evidence
  • May vary sharply across inputs
  • Does not prove product value
  • Example: transcribing clean studio audio

Usefulness

Does the output improve a real human or organizational workflow?

  • Compared with the current baseline
  • Includes time, cost, and friction
  • Depends on user needs and adoption
  • Example: reducing correction time for transcripts

Trustworthiness

Is the system sufficiently reliable, safe, secure, fair, transparent, and accountable for this context?

  • Depends on stakes and affected people
  • Requires failure handling
  • Cannot be reduced to one score
  • Example: protecting sensitive meeting audio

Example

Five ordinary encounters, five different systems

AI becomes easier to understand when you name the exact job it performs. These familiar products do not share one universal kind of intelligence.

  • Email filtering: a classifier estimates whether an incoming message resembles spam, phishing, promotions, or legitimate mail.
  • Navigation: forecasting models estimate travel time, while search and optimization select a route under changing constraints.
  • Photo search: vision and language representations connect images with concepts such as “beach,” “birthday,” or a person’s face.
  • Streaming recommendations: ranking systems choose which items to show first, using behavior signals and product objectives.
  • Writing assistance: a generative model proposes text, while the application may add retrieval, formatting rules, safety checks, and user approval.

Analogy

A map is useful because it leaves things out

A transit map in an unfamiliar city does not show every doorway, tree, or traffic light. It shows the relationships needed to travel. That is exactly why it works.

This course map works the same way, separating the field, model, system, and context so you can navigate first. Printed transit lines, though, stay where they are. AI systems change after deployment.

A beginner map should simplify without pretending that the omitted details do not matter.

Harry Beck, 1933, and the geography he threw away

The transit map that made this trade-off famous is Harry Beck’s London Underground diagram, first issued to the public as a pocket map in January 1933 after a small trial the previous year. Beck discarded true geography: horizontal, vertical and 45-degree lines, roughly even spacing between stations, and a central area drawn far larger than scale. London Underground had first rejected the design as too radical because it did not show real distances. Travellers adopted it immediately, because the deliberately inaccurate picture answered the question they actually had — where to change trains.

Steps

A six-step routine for any AI feature

Use this routine whenever a product is described as intelligent, autonomous, predictive, or generative.

FigureProcess · 6 steps

  1. 1. Name the user goal

    Describe the human problem without mentioning AI. This exposes whether the technology is solving something real.

  2. 2. Name the output

    State exactly what the model returns and what uncertainty or ambiguity remains.

  3. 3. Trace the evidence

    Ask what data, rules, retrieved sources, feedback, or expert knowledge influence the output.

  4. 4. Follow the handoff

    Identify who or what receives the output, what decision follows, and whether review is possible.

  5. 5. Imagine failure

    Choose one common error and one rare high-impact error, then inspect the fallback.

  6. 6. Compare a baseline

    Ask whether simpler software, better information, or process redesign could solve the problem more reliably.

Example

The routine, run once on Air Canada’s chatbot

Here are the six steps applied to the case at the top of this lesson, using only what the tribunal recorded and what was reported from its decision. Watch which steps the record answers and which it leaves open.

  • User goal. A man whose grandmother had just died needed to know whether he could fly at a reduced bereavement fare, and how to claim it. The problem can be stated without mentioning AI, which is the point of the step.
  • Output. A generated answer in a chat window: “If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued by completing our Ticket Refund Application form.” The words “bereavement fares” in the same answer linked to the airline’s own “Bereavement travel” page.
  • Evidence. Here the public record runs out. It shows what the chatbot said, and that the page it linked to said the opposite: the reduced fare could not be claimed once travel was complete. What the chatbot drew its answer from is not in the account of the case. That gap is where an auditor would start, before launch rather than after.
  • Handoff. The answer went straight to the customer, with no person in between, and he acted on it. He bought a next-day ticket and a return, both at full fare, and then applied for the reduced rate as the chatbot had told him to.
  • Failure. The common error is a wrong statement of policy. The rare, expensive one is a wrong statement a customer spends money on. The fallback came afterwards and by hand: an Air Canada representative admitted in writing that the chatbot had used “misleading words” and said the issue had been noted so the chatbot could be updated. The airline offered a $200 coupon. He refused it and went to the tribunal.
  • Baseline. The non-AI answer already existed, on the same website. Air Canada argued that he could have found the correct information there. The tribunal was not persuaded: “It makes no difference whether the information comes from a static page or a chatbot.” It ordered the airline to pay $650.88, the difference between what he paid and the bereavement fare, plus interest and fees: $812.02 in all.

With the facts in front of you, the routine takes minutes. Air Canada met the same facts in front of a tribunal.

What this path promises—and what it deliberately postpones

By the end, you should be able to explain AI without hype, inspect an AI-enabled workflow, recognize major limitations, and evaluate product claims. You should also know which deeper questions belong to later paths. Knowing what you are not yet equipped to judge is part of the skill.

This path will not teach backpropagation, transformer internals, clustering algorithms, or detailed regulation. It will give you the conceptual structure needed to learn those topics without confusing a model with a product or a benchmark with reality.

The goal is informed judgment: enough technical accuracy to ask better questions and enough humility to keep investigating.

Key takeaways