What's the actual difference between AI and machine learning?

Machine learning is a specific method for building AI. Think of AI as the whole field — any machine that can do something smart. Machine learning is the approach where you don't program rules directly. You feed the computer data and let it figure out the patterns on its own. It's like the difference between teaching someone to cook by giving them a recipe book versus having them taste hundreds of dishes and figure out what works. Both get you a meal, but the path is completely different. I've found this distinction trips up a lot of beginners because the terms get thrown around like they're the same thing. They're not. All machine learning is AI, but not all AI uses machine learning. The older kind of AI — what people now call 'symbolic AI' or 'good old-fashioned AI' — worked on hard-coded logic. If this, then that. A chess program from the 90s that used a database of opening moves and brute-force calculation? That's AI without machine learning. It's not learning from games, it's just following instructions. Modern chess engines like AlphaZero, on the other hand, learned by playing millions of games against itself. That's machine learning. Here's a concrete example. Say you want to build a spam filter. The old AI way: you write rules. If the email contains 'Nigerian prince', mark as spam. If it has too many exclamation marks, add points. You maintain a list of bad words. The machine learning way: you gather 100,000 emails that humans already labeled as spam or not spam. You feed them to an algorithm. It finds the patterns — maybe it notices that spam emails tend to have certain word combinations you'd never think to write a rule for. The algorithm builds its own internal model of what spam looks like. You never told it about Nigerian princes. It figured that out from the data. The big insight here is that machine learning shifts the work from writing rules to collecting and labeling data. That's not always easier. Sometimes writing a few if-statements is way more practical than gathering 100,000 labeled examples. The tool should fit the problem, not the other way around.