What is semantics – and why was the philosopher arguing with his quill pen about it?

Last time we left our cartoon philosopher mid-scribble, ink flying, with one word still hanging unresolved in the air: semantics.

It sounds like something you should already remember from school, but that most of us just nod along to in practice. Yet semantics has become highly useful in a world where we increasingly use ordinary language to control technical systems. So what does it actually mean?

The short answer: Semantics is about meaning – what words and sentences mean, not just how they're constructed.

Syntax, semantics, and pragmatics walk into a bar

To understand semantics, we need to separate three things that tend to get blurred together:

• Syntax is about how words are put together.
• Semantics is about what the words and the sentence mean.
• Pragmatics is about what someone intends by an utterance in a given context.

Take the question: "Could you pass the salt?"

Syntactically, it's a correctly formed question. Semantically, it asks whether the person has the ability to pass the salt. Pragmatically, most of us understand it actually means "Please pass the salt." That's why almost no one answers "yes" and keeps eating.

We humans use situation, relationship, tone, and prior experience to grasp the intent behind words – so fast we barely notice. And that understanding is exactly what an AI neither truly has nor entirely lacks – it guesses its way there, based on patterns. It often works astonishingly well. But "astonishingly well" isn't the same as "understands exactly what I mean."

Semantics' little trap

Imagine asking an AI: "Write a short and funny summary."

That sounds clear. But almost every important word leaves room for interpretation.

What does short mean – fewer than a hundred words, a minute's reading time, half the length of the original? What does funny mean – warm and light, ironic, absurd, funny to a developer or to a board of directors? And what does summary mean – a neutral account, the key takeaways, or a persuasive pitch where nuance gets sacrificed?

The problem isn't necessarily that the AI is bad at following instructions. The problem is that the instruction contains gaps – and the model politely fills them with its own likely assumptions.

Aha: The less of your intent you express, the more of the result you hand over to the model's assumptions.

A borrowing from Kahneman

In Thinking, Fast and Slow, Daniel Kahneman describes two ways of thinking: System 1, fast, intuitive, automatic – and System 2, slower, effortful, analytical.

A language model obviously doesn't have these two psychological systems. But the analogy is useful: when an AI quickly generates a seemingly obvious answer, it's easy for us to treat that answer as though someone had already thought the question through carefully. In reality, the model may have simply followed the most probable path through an ambiguous instruction – it has no System 2 to pause and ask whether you really meant what you wrote.

Which, paradoxically, makes you the machine's System 2: the one who has to pause, think it over, and rephrase before the answer is already locked in.

Kahneman's research also shows that how information is framed shapes how we judge it. "90% survival" feels different from "10% mortality," even though it's the same figure. The same happens when we phrase questions to an AI: ask it to describe the "risks" of adopting AI in a business, then ask about the "opportunities" – you'll likely get two quite different stories. The problem arises when we forget the frame was built into the question and start treating the answer as a neutral description of reality.

A better instruction might be: "Analyze both the opportunities and risks of adopting AI in the business. Use the same evaluation criteria for both, surface key assumptions, and flag what requires more information." No guarantee of truth – but a considerably better start.

From theory to practice: what do people actually think?

Enough philosophy and thought experiments – what is semantics actually used for?

Here's a good example: you launch an ad campaign and want to know how it landed on X and Instagram. The easy way is to count comments, likes, shares, positive and negative words. The smart way is semantic analysis – letting the AI read how people are actually expressing themselves, not just that they are.

The difference is enormous. "Wow, what a hit" could be genuine praise – or the most toxic line in the whole comment section, depending on what was said right before it, the account's usual tone, and whether it's followed by three crying-laughing emoji. Plain keyword counting only sees positive words. Worth distinguishing a few related terms here: sentiment analysis judges whether a text is positive, negative, or neutral. Semantic analysis identifies meanings, topics, and relationships. Pragmatic interpretation tries to understand intent in the actual situation – which, incidentally, is exactly what our quill-wielding philosopher has been arguing is the whole point all along.

The same principle works well for product reviews. Instead of just sorting by star rating, you can categorize by how people write: a sober, detailed rundown of features (a buyer who did their homework)? An emotional outburst after a bad delivery experience (not really about the product at all)? A comparison against a competitor (a buyer still on the fence)? Same star rating, three completely different insights.

One small but important detour: irony has long been AI's Achilles' heel. "Great service, only waited 45 minutes" could historically be read as positive by a model stuck on the words rather than the intent – like an overly literal exchange student. The latest generations of language models have gotten considerably better at catching sarcasm and understatement, precisely by weighing in more context: what was said before, the account's general tone, which words tend to follow which emotions. Still not foolproof – irony works precisely because we say one thing and expect the listener to understand another, which makes it a fairly brutal test of whether the context truly came along for the ride.

Same words, different worlds

Even a crystal-clear phrasing can go wrong if it lacks its context. "Book the table" means one thing at a restaurant, something else in a carpentry shop, and perhaps something else again at an auction house. The word is the same. The syntax is the same. But the meaning is shaped by where it's spoken, by whom, and by the larger system of assumptions it belongs to – what things exist, what they're called, how they relate, what one's allowed to do with them.

And that system of concepts and relationships has a name of its own: ontology.

The philosopher gets the last word

Semantics, then, isn't just an academic word for people with too many books and strong opinions about commas. It's a practical tool for spotting the gap between what we say, what the words mean, what we intend – and what the listener actually perceives.

When the listener is an AI, that gap matters even more. Expressing yourself precisely, then, isn't about writing perfect "magic prompts." It's about thinking clearly enough to say what you mean – and checking whether the answer actually matches the intent.

The philosopher, in other words, had a point. Even if he still has ink all over the desk.

Want to go deeper: George Lakoff & Mark Johnson's Metaphors We Live By is an accessible entry point into how linguistic imagery shapes thought. Steven Pinker's The Stuff of Thought and S.I. Hayakawa's Language in Thought and Action go further into the relationship between language, thought, and action. For those who want to go deeper still, there's Alfred Korzybski's "the map is not the territory" in Science and Sanity, Wittgenstein's idea that meaning is created through use in Philosophical Investigations, and John Searle's Speech Acts.

Next part

Words get their meaning from context. But how do we describe the context itself – the things that exist within it and the relationships between them? Next time we tackle ontology: a word that sounds harder than it needs to be.