AI Literacy: A Map of the Field

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.

Context of use is also where the law operates. The EU Artificial Intelligence Act has been in force since 1 August 2024. It assigns obligations by intended purpose rather than by technique. One and the same ranking model carries no high-risk duties while it sorts a product catalogue. It becomes high-risk under Annex III, point 4(a) once it is used “for the recruitment or selection of natural persons … to analyse and filter job applications.” Nothing about the model changed. The level of the map did.

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.

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

Take the transcription example. 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. Word-level accuracy was high, 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.

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

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.

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

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.