AI, machine learning, deep learning, generative AI: what's the difference?

AI, machine learning, deep learning, generative AI. Four terms that often get used as if they mean the same thing. They don’t, and mixing them up is how organizations end up using the wrong tool.

Artificial intelligenceSystems that turn inputs into predictions, content, recommendations or decisions. Includes rules written by people.

Machine learningLearns the rules from past examples instead of being told them.

Deep learningMachine learning with many-layered neural networks. Good with images, audio and text.

Generative AICreates new content: text, images, code. ChatGPT and Copilot live here.

One thing surprises people: a well-designed set of business rules counts as AI. The OECD’s definition explicitly includes rule-based approaches alongside machine learning.

One utility, four kinds of AI

Picture a mid-sized utility. Each layer solves a different problem.

• RulesFlag an odd meter reading

  If a reading is far above the same month last year, send it for review. A person wrote that rule.

• Machine learningPredict transformer failures

  Learns from years of load, age and outage history which units are most at risk.

• Deep learningSpot damage in drone photos

  Finds cracked crossarms and damaged insulators across thousands of images.

• Generative AIDraft the outage notice

  Writes a first draft for customers. A person checks it before it goes out.

Which one do you need?

Start from what the system has to produce, not from the technology.

What does it need to produce?

1. A decision from rules you already knowuse RulesPredictable and easy to audit
2. A number or category from your data tablesuse Machine learningForecasts, scoring, anomalies
3. Something found in images, audio or free textuse Deep learningNeeds lots of labelled examples
4. New text, a summary, or answers from documentsuse Generative AIPlan for human review

The most common mistake runs the other way: asking a chatbot to forecast next winter’s peak load. A language model writes by predicting likely next words. It doesn’t calculate loads. Use a forecasting model, and let the language model explain the forecast if that helps your users.

Each layer brings its own risks

	Good for	Needs	Watch out for
Rules	Known, stable logic	People who know the rules	Cases the rules never covered
Machine learning	Forecasts, scoring, anomalies	Clean historical data	Biased history; drift as conditions change
Deep learning	Images, audio, free text	Lots of data and computing power	Hard to explain a single decision
Generative AI	Drafting, summarizing, search	Good source documents	Confident, wrong answers

That last risk has a name. NIST calls it confabulation (most people say hallucination) and describes it as a natural result of how generative models are designed. It can be reduced, but not switched off.

Governance follows the impact, not the label

Canada’s federal Directive on Automated Decision-Making covers any technology that “either assists or replaces the judgment of human decision makers”, from rules-based systems through to deep learning. It applies to federal institutions, but the lesson travels: a spreadsheet macro that decides who qualifies for a program can need more oversight than a chatbot that answers questions about office hours.

What to do on Monday

• List your AI ideas. For each one, write what goes in, what comes out, and which decision it informs.
• Label each one: rules, machine learning, deep learning or generative AI.
• Ask vendors which layer is under the hood, and what data it learned from.