AI, ML, deep learning — the three layers explained in plain English.
Published September 21, 2026
Artificial intelligence (AI) is the broad goal of building computer systems that perform tasks we'd normally say require human intelligence: understanding language, recognising images, making decisions, planning, learning from experience. It isn't one technique but a large field, and the technique that dominates today is machine learning.
Artificial Intelligence any technique that makes machines behave "intelligently"
└── Machine Learning systems that LEARN patterns from data instead of following hand-written rules
└── Deep Learning machine learning with many-layered neural networks
└── Generative AI / LLMs deep learning models that generate text, images, code, audio
Everything inside is also everything outside: an LLM is a deep learning model, which is a machine learning system, which is a form of AI. The reverse doesn't hold. A chess engine that searches moves using hand-written evaluation rules is AI, but not machine learning.
Traditional programming: a developer writes the rules, and the computer applies them to data.
data + rules → answers
if (email.contains("free money") && sender.isUnknown()) markAsSpam();
Machine learning: you provide examples with answers, and an algorithm finds the rules, captured as numbers in a model.
data + answers → rules (a trained model)
10 million emails labelled spam / not-spam → a model that predicts spam for new emails
ML wins when the rules are too many, too subtle or too fast-changing to write by hand, like recognising a face or detecting a new fraud pattern, and when there's enough representative data. Hand-written rules still win when the logic is clear, must be exactly explainable, or when data is scarce. Tax calculations shouldn't be "learned".
Three things arrived together:
| Task | Typical approach | Everyday example |
|---|---|---|
| Classify things | Supervised learning | Spam filters, fraud flags, content moderation |
| Predict numbers | Regression | Delivery-time estimates, demand forecasting |
| Recommend | Collaborative filtering, embeddings | "Customers also bought", video feeds |
| Find groups or anomalies | Unsupervised learning | Customer segments, unusual network traffic |
| Understand and generate language | Large language models | Chat assistants, summarization, code assistants |
| See | Deep learning (vision models) | Face unlock, document scanning, defect detection |
| Act and plan | Reinforcement learning | Game playing, robotics, some recommendation tuning |
The learning styles behind these are covered in Types of Machine Learning, and LLMs specifically in How LLMs Work.
Every AI system deployed today is narrow: very capable within the task it was built or trained for, and unreliable outside it. A fraud model knows nothing about images. LLMs are unusually broad, since language covers many tasks, but they still make confident mistakes, have knowledge cut-offs, and don't reliably know what they don't know. Artificial general intelligence (AGI), meaning human-level ability across essentially all tasks, remains a research goal and a subject of active debate, not a deployed product.
Most engineers don't train large models from scratch. They integrate AI:
Q: What's the difference between AI, machine learning and deep learning? A: AI is the overall goal of intelligent behaviour by machines, by any means. Machine learning is the subset where behaviour is learned from data rather than programmed as rules. Deep learning is the subset of ML using many-layered neural networks, which dominates perception and language tasks.
Q: When should you not use machine learning? A: When clear, stable rules exist and exactness matters (billing, permissions); when there isn't enough representative data; when every decision must be fully explainable; or when a simple heuristic performs about as well at far lower cost and complexity.
Q: Why did deep learning suddenly work so well? A: The combination of very large datasets, GPU and cloud compute that made training big networks affordable, and algorithmic advances (better initialization and activation functions, regularization, and later the transformer architecture) that made deep networks trainable and scalable.
Q: Is an LLM "intelligent" or just predicting text? A: Mechanically, it's trained to predict the next token. Doing that extremely well at scale requires internal representations that support useful reasoning, summarization and coding. It's also why LLMs can produce fluent errors: plausibility, not verified truth, is what they're optimized for. Practical systems add retrieval, tools and verification around them.
Q: What skills does a backend engineer need to build AI features? A: API integration and resilience (timeouts, retries, fallbacks), prompt and output design, retrieval over company data, evaluation and monitoring, cost control, and security awareness (data privacy, prompt injection). Deep model training is a specialism most product teams don't need.