What AI actually is: a guide for drug and alcohol workers

Start with the task, not the label

Artificial intelligence can describe systems that classify images, predict an outcome, recommend content or generate text. Those systems have different purposes and risks. Calling all of them ‘AI’ can hide the practical question: what is this particular system doing, with which information, for whom and with what consequence?

Generative AI produces new text, images, audio or other material from a user’s instructions. A large language model generates language by learning patterns from vast collections of text. It can produce remarkably coherent answers without knowing whether those answers are true.

Capability is not judgement

A tool can summarise a public report, improve a document structure or draft fictional training material. It does not understand a service user, hold a professional relationship or accept accountability. Fluency can make an answer feel more informed than it is.

The safest everyday uses have a defined purpose, non-identifiable information, clear source boundaries and an output that a person can independently check. The higher the consequence of an error, the stronger the controls and human expertise need to be.

A practical mental model

Treat generative AI as a fast, fallible assistant. Give it the task, audience, boundaries, source material and required output. Then examine omissions, bias, certainty and evidence. If the work affects care, safety, rights or access to services, do not treat a general-purpose AI response as the decision.