AI vs ML vs Deep Learning

Walk into any product review in 2026 and someone will say their feature is 'AI-powered' before they've described what it does. The phrase is doing a lot of work — and most of it is rhetorical. Artificial intelligence is the umbrella term for any software system that does something we'd call intelligent if a human did it, which is so broad that it captures everything from a thermostat with three temperature rules to GPT-5. What that means in practice is that 'AI' describes the behaviour, not the mechanism. A chess engine that brute-forces every move with hand-coded heuristics is AI. A model trained on every game ever played is also AI. The label tells you nothing about how the thing was built.

Picture two ways to build a spam filter. The first is a senior engineer writing 200 lines of if-statements: block this sender, flag that keyword, score these patterns. The second is showing a system a hundred million emails labelled spam or not-spam and letting it work out the pattern itself. Machine learning is the second approach: instead of programming behaviour directly, you give the system examples and let an algorithm derive the rules from them. The first approach is engineering. The second is closer to teaching, with all the messiness that implies.

For decades, getting an ML system to do anything useful with images, speech or text required a small army of domain experts to hand-design the inputs. You couldn't just hand a model raw pixels — you'd hand it carefully engineered features like edge counts, colour histograms, or speech spectrograms, and the model learned from those. Deep learning collapsed all of that. A deep neural network takes raw inputs — pixels, audio waveforms, characters — and learns the useful features for itself, layer by layer.

The cleanest way to hold these three terms in your head is as three nested circles. AI is the outer circle: any system that behaves intelligently. Machine learning sits inside it: AI systems that learn from data. Deep learning sits inside ML: the subset that uses large neural networks. That asymmetry matters. The terms are not synonyms with different vibes — they describe progressively smaller and more specific commitments about how a system is built.

A rule-based system is software where every behaviour was written down by a human. If income > $50k AND credit_score > 700 → approve loan. If subject_line contains 'lottery' → flag email. If user hasn't logged in for 14 days → send re-engagement email. A learned system has no explicit rules. It has a trained model — a pile of weights — that maps inputs to outputs. The 'rules' are implicit and statistical, and they shift every time the model is retrained.

Question one: is there a trained model, or are these hand-written rules? This is the rules-versus-learning diagnostic from the previous section, applied as the opening move. If the answer is 'no model, just rules', the conversation is no longer about AI — it's about software, with all the normal evaluation criteria. A useful follow-up: 'what kind of model, and what does it predict?' One sentence of specificity here separates a serious team from a marketing one. You don't need to understand the technique; you need to hear that they do.