What is machine learning, and how do four sources define it?
Machine learning is software that gets better at a task by studying examples, not by following rules a programmer typed in. Show a spam filter ten thousand emails marked spam or fine, and it works out what spam looks like. Nobody wrote that list of rules. The table sets four published definitions side by side.

| Source | Definition as written | What it stresses |
|---|---|---|
| Tom Mitchell, Machine Learning (McGraw Hill, 1997) | "Machine Learning is the study of computer algorithms that improve automatically through experience." | Improvement over time |
| Stanford HAI | "a branch of artificial intelligence that enables computers to learn patterns and make decisions from data" | Patterns and decisions from data |
| AWS | "a type of artificial intelligence that performs data analysis tasks without explicit instructions" | Data analysis tasks |
| IBM | "a subset of AI that allows for optimization" | Optimization |
The wording differs and the core doesn't. In all four, the program's behavior comes from data. Mitchell's version is the oldest and the one textbooks still cite. The idea is older still: Arthur Samuel's paper in the IBM Journal of Research and Development, July 1959, opens with "Two machine-learning procedures have been investigated in some detail using the game of checkers."
How does machine learning work?
A machine learning system goes through the same four stages whatever it is learning. First, people collect examples, such as past transactions with a fraud label. Second, a model starts with random guesses and is scored on each example. Third, training nudges the model's internal numbers, called parameters, to cut the score's error. That loop runs over the data many times.
The fourth stage is the one beginners skip. The finished model is tested on examples it has never seen. If it scores well on its training data and badly on new data, it memorized instead of learning, which practitioners call overfitting. A model is only useful if it handles tomorrow's data, not yesterday's.
This is also why the data matters more than the algorithm. A model trained on narrow or skewed examples gives narrow or skewed answers, and no amount of tuning fixes that. Why AI gets things wrong covers what that looks like in a chatbot.
AI vs machine learning vs deep learning: where each one stops
Artificial intelligence is the whole field, machine learning is one method inside it, and deep learning is one kind of machine learning. They nest like boxes. Not all AI learns from data: a chess program built from hand written rules is AI and is not machine learning. IBM states the nesting directly in its comparison page.
| Term | Scope | What IBM says | Example |
|---|---|---|---|
| Artificial intelligence | The broadest term | "used to classify machines that mimic human intelligence" | A rule based chess engine, a chatbot, a recommender |
| Machine learning | A subset of AI | "a subset of AI that allows for optimization" | A spam filter trained on labeled emails |
| Deep learning | A subset of machine learning | Automates much of the feature extraction piece of the process | Image recognition, speech to text, ChatGPT |
The practical split between classic machine learning and deep learning is who picks the clues. In classic machine learning, a person decides which features matter, such as the number of links in an email. IBM says deep learning "automates much of the feature extraction piece of the process", so the network finds its own clues in raw pixels, audio or text. The price is data and computing power: deep networks need far more of both.
So the answer to "machine learning vs AI" is that machine learning is a part of AI. The answer to "deep learning vs machine learning" is that deep learning is a part of machine learning. What is AI and types of AI cover the outer box, and what is generative AI covers the newest branch.
What is deep learning?
Deep learning is machine learning done with neural networks that have many layers. IBM says a network with more than three layers, counting the input and the output, can be called a deep learning algorithm. Each layer turns the last layer's output into something a bit more abstract: pixels become edges, edges become shapes, shapes become a face.
Deep learning took over image recognition around 2012. The NeurIPS paper by Krizhevsky, Sutskever and Hinton trained a deep convolutional network on 1.3 million images from the LSVRC-2010 ImageNet set. On its test data it reached top 1 and top 5 error rates of 39.7% and 18.9%, which the authors call a clear gain over the best earlier results. That was a big jump. Large language models such as the ones behind ChatGPT, Claude and Gemini are deep networks too. What is a neural network explains the layers, and what is an LLM covers the text case.
What are the three types of machine learning?
Machine learning is usually split by what the model gets as feedback. Supervised learning trains on examples that carry the right answer, such as emails marked spam or fine. Unsupervised learning gets no answers and looks for structure on its own, such as grouping customers by buying habits. Reinforcement learning gets a reward or a penalty after each action, and learns a strategy, as in a game.
AWS counts four learning styles, because it adds semi supervised learning, which mixes a few labeled examples with many unlabeled ones. Most explainers teach three. Supervised vs unsupervised learning takes the first two apart with examples.
Three real cases of machine learning, from 1959 to 2012
Checkers, 1959. Arthur Samuel's program learned to play checkers at IBM, and his paper in the IBM Journal, volume 3, July 1959, dates the idea to the 1950s. It's a reinforcement style problem: play, score the result, adjust. You don't need a labeled answer for every move, only a way to score how a game ended.
Fraud, today. AWS gives fraud as its business example: the system finds patterns in known data to predict whether a new transaction is genuine. That's supervised learning when past transactions carry a fraud label, and the model learns what labeled fraud looks like. Banks can't write a rule for every scheme, so they let the data set the rule.
Photos, 2012. The NeurIPS image classifier above, published in 2012, is the case where deep learning beat the hand built approaches by a wide margin. It needed a large labeled set of photos and GPU computing, and it set the pattern for the decade after. Today's image tools, voice assistants and chatbots all descend from it.
Stanford HAI lists email spam filters, recommendation systems, image recognition, voice assistants and fraud detection as common uses. Every one of them follows the same loop: examples in, a model out, a test on new cases.
What are the limits of machine learning?
Machine learning finds patterns and doesn't understand them. A fraud model flags transactions that look like past fraud, and it will miss a new scheme that looks like nothing it has seen. It also copies whatever is in its data: gaps, errors and bias included.
Generative models add a second problem. They produce what fits the pattern of their training data, which is why a chatbot can state a false fact in a confident voice. What is an AI hallucination covers that case. Whenever a decision affects someone's money, health or job, a person has to check the model's output. AI ethics covers bias and accountability in more depth.
What questions should you ask an AI about machine learning?
An AI chat tool is a good tutor for this topic if you give it your level and ask it to check you. These prompts work in Claude, ChatGPT, Gemini or Whizi. Fill the brackets.
- "Explain machine learning to me. I work as [your job] and know [what you know, such as basic statistics or no math]. Use one example from my field, then ask me a question to check I understood."
- "Take this task: [your task, such as sorting support tickets]. Say whether it needs machine learning or whether simple rules would do, and give the three questions that decide it."
- "I have [describe your data, such as 2,000 labeled emails]. Which type of machine learning fits, supervised, unsupervised or reinforcement, and why?"
- "Explain overfitting using an example from [a sport or hobby you like]. Then give me a quick quiz of three questions."
- "Compare deep learning and classic machine learning for [your task]. Make a table with data needed, cost and one risk each, and mark anything you aren't sure of."
- "Here is a claim from an article: [paste]. List what a machine learning claim like this must show to be believable, such as the test data and the error rate, and say which of those the claim is missing."
Where does Whizi fit when you are learning machine learning?
Whizi is one app instead of a pile of single purpose tools. Pro carries Claude, GPT and Gemini together, and its first week is $1.99. On Pro you can put prompt 1 or prompt 5 to two models side by side and read only the lines where they disagree. Web search is on by default in each chat, so a model can look up a current fact.
The free plan gives you 10 messages a day on Whizi AI, Auto and Claude Haiku, with no card. That's enough to try the prompts above. A free explainer page or a free course is still the better place to start if you want a full syllabus, and Whizi doesn't replace one. See Whizi pricing for what each plan holds.
- Ask what data a system learned from before you call it machine learning
- Remember the nesting: deep learning sits inside machine learning, which sits inside AI
- Ask how a model was tested on examples it never saw in training
- Match the learning type to your data: labeled examples mean supervised learning
- Check any machine learning claim for its test data and its error rate
Frequently asked questions
What is machine learning in simple words?
Machine learning is teaching a computer by showing it examples instead of writing rules. A spam filter shown thousands of emails marked spam or fine learns what spam looks like. The more good examples it sees, the better it gets, and it is tested on emails it has never seen.
What is the difference between AI and machine learning?
AI is the whole field of machines that mimic human intelligence, and machine learning is one method inside it. A chess engine built from hand written rules is AI without being machine learning. IBM describes machine learning as a subset of AI. Most modern AI, including ChatGPT, uses it.
What is the difference between deep learning and machine learning?
Deep learning is a type of machine learning that uses neural networks with many layers. IBM says classic machine learning depends on people to pick the features, while deep learning automates much of that step. Deep learning needs more data and computing power, and it runs image recognition and chatbots.
Is ChatGPT machine learning?
Yes. ChatGPT is a large language model, which is a deep learning model, which is a kind of machine learning. It learned patterns from a very large body of text and generates answers one piece at a time. What is generative AI covers how that branch differs from older AI.