Concept of the Week: Types of AI
Three ways software does “intelligent” work and how to pick the right one
A brief introductory note
As a layman, the advances in artificial intelligence are overwhelming me, which is why I started this segment of the blog. Too many terms, shifting claims, and not enough proof. I’m not a researcher. I’m a non-specialist who needs reliable answers for real work. So I’ve decided to make available what I learn to others who may be in the same position.
Each Friday I unpack one idea in plain English. I show where it appears in everyday tools, what it changes for cost, quality, speed, and risk, and how to try it safely. There’s a five-minute test you can run and one action you can take that day. I use dated sources. I mark what’s evidence and what’s observation. If something’s wrong, I’ll fix it.
This isn’t a course. It’s a small weekly habit to cut noise and build understanding, one idea at a time. Read in two minutes. Try it in five. Leave with one clearer decision.
What “AI” means here
Software that takes an input and either makes a judgment (classify, rank, route, approve/deny) or produces content (text, image, audio, code).
Rules
What it is: software follows explicit “if X, then Y” instructions written by people.
Everyday examples: a tax calculator; a content management system (CMS) field check that rejects an ID with the wrong length; email rules that file messages.
Tool example: Gmail Filters or Outlook Rules.
Machine learning
What it is: software learns from examples instead of hand-written rules.
EEveryday examples include spam detection in your inbox (where emails are labelled as "spam" or "not spam"), extracting totals from invoices, and grouping articles by theme.
Tool example: Google Vertex AI AutoML or Microsoft Fabric AutoML for a small text classifier without code.
Generative AI
What it is: software that produces new text, images, audio or code by predicting the next piece from patterns in data. Think make, not fetch.
Everyday examples: autocomplete in a phone keyboard; a tool that writes a 150-word note from a brief; an image tool that draws “a red bus at night”.
Tool example: ChatGPT, Microsoft Copilot, or Google Gemini. For documents, prefer modes that quote or cite your source text.
Try it now (3–5 minutes, ChatGPT or similar)
Goal: get a safe first draft from a fixed brief.
Metric: edit time and number of factual fixes.
Paste a ~120-word brief with facts (who/what/where/when).
Prompt: “Draft 150 words using only these facts. Quote exact lines in quotes. UK English. Short sentences.”
Accept text only where quoted lines match the brief. Fix anything that doesn’t.
Good result: publishable in ≤10 minutes with 0 factual fixes.
How to choose next time
Clear, stable steps? Use rules (cheap, auditable).
Lots of examples and a target metric? Use machine learning.
Need a first draft or options fast? Use generative with quotes/citations and a human check.
Common slip
Using a chatbot for narrow, rule-based tasks such as simple routing and redaction is effective. A short rule or tiny classifier is cheaper, faster, and easier to audit.
Takeaway
Rules for fixed steps. Learning for patterns. Generative for drafts – with quotes or citations and a human check.