What AI is, and is not

You have spent your career making computers do exactly what you tell them. AI flips part of that: instead of writing the rules, you show the machine examples and let it find the rules. This lesson gives you the map and the vocabulary, so the rest of the course (and every news headline) makes sense.

Watch

AI, Machine Learning, Deep Learning and Generative AI ExplainedIBM Technology · 10 min

The words, in order of size

These terms are nested. Each one is a smaller, more specific slice of the one before it.

Artificial intelligence (AI)         any technique that makes machines act intelligently
└─ Machine learning (ML)             systems that learn patterns from data
   └─ Deep learning                  ML using large neural networks
      └─ Generative AI               models that create text, images, audio, code
         └─ Large language models    generative AI for text (ChatGPT, Claude, Gemini)

• AI is the umbrella. A chess program using hand-written search rules counts, and so does a chatbot.
• Machine learning is the part that learns from data instead of following only hand-written rules.
• Deep learning uses neural networks with many layers. Almost every recent breakthrough comes from here.
• Generative AI creates new content rather than only labelling or predicting.
• Large language models (LLMs) are generative models trained on vast amounts of text. They power today's chat assistants.

Rules in, answers out versus data in, rules out

Traditional software: a developer writes the rules, the program applies them to data, and you get answers.

Machine learning: you provide data and the right answers, and training produces the rules (called a model).

An IT analogy: a firewall with hand-written rules versus a system that studies months of traffic and learns what "normal" looks like. The first is predictable and explainable. The second can catch things you never thought to write a rule for, and can also surprise you.

Narrow and general

Every AI system you can use today is narrow: very good at what it was built and trained for, with no general understanding behind it. An image classifier cannot write an email, and a chatbot is not a database. "General" AI that matches human flexibility is a research goal, not a product, and experts disagree about when or whether it will arrive.

LLMs feel more general because language is a flexible interface. They can attempt thousands of tasks. That breadth is real, but so are the gaps, which later lessons cover.

The AI effect

People joke that "AI is whatever computers cannot do yet." Once something works reliably, such as spell check, route planning or spam filtering, we stop calling it AI and call it software. Expect today's impressive demos to become tomorrow's ordinary features, and plan your skills accordingly.

Why now

Three things came together:

1. Data: the internet and digital records produced huge training sets.
2. Compute: GPUs made it practical to train very large models.
3. Better methods: the transformer design (lesson 4) scaled far better than earlier approaches.

On top of that, chat made it usable. Anyone can now try a powerful model by typing a sentence.

Myths to drop early

• "AI understands like a person." It predicts patterns. Whether that counts as understanding is a debate, but it behaves differently from a human mind.
• "AI is always right, or always wrong." It is neither. Quality varies by task, and it can be wrong while sounding certain.
• "You need a maths degree to work with AI." Building models from scratch needs deep expertise. Using and integrating them well mostly needs good engineering judgement, which you already have.
• "It will replace IT." Tasks change. People who can direct AI, check its work and fit it into real systems become more valuable, not less.

Resources

Check yourself

1. Explain to a colleague, in three sentences, the difference between AI, machine learning and generative AI.
2. List ten software features you use daily. Mark each as hand-written rules, machine learning, or generative AI.
3. Pick a rules-based system you maintain. What would it mean for it to "learn" instead? Name one benefit and one risk.