What Are AI Writing Tools and How Do They Work?

Ask an AI writing tool for five books about a niche subject and you will get five. Real-sounding titles, real-sounding authors, the right sort of publisher. Check them and two might not exist — and nothing in the answer looked different from the three that did.

That is not a bug someone forgot to fix. It falls out of how these tools are built, and knowing how they are built lets you predict it — along with most of the other odd things they do.

The one sentence that explains almost everything#

These tools have been fed a very large amount of text and trained to answer one question over and over: given everything so far, what word is likely to come next?

That is most of it. Write one word, add it to what you have, ask again. Repeat a few hundred times and you have a paragraph.

It reads like a person because the patterns came from people. Sentence rhythm, paragraph shape, the way an apology email differs from a product description — all pattern, all learnable from examples.

It invents facts for the same reason. "The capital of France is Paris" and "the capital of France is Lyon" are both fluent English sentences. The model has seen the first far more often, so it usually produces it. But nothing in the machinery checks a claim against the world. It checks what tends to follow what.

Two words you keep seeing: training data and context window#

Training data#

Training data is the pile of text the tool learned patterns from — books, websites, code, forums, a great deal of ordinary internet writing, collected up to a cut-off point. Then the learning stopped.

So the tool knows nothing after that date unless it can search the web. It absorbed the errors and biases in that text, having no way to grade what it read. And it does not store documents you can look up: what remains is a statistical impression, not the originals. That is exactly why a half-remembered citation comes out as a confident, wrong one.

Context window#

The context window is how much text the tool holds in view at once: your instructions, anything you pasted, attached documents, and everything said so far in the conversation.

Everything inside that window shapes the next word. Everything outside it does not exist. That is why long conversations drift — early instructions fall off the back — and why it cannot recall what you said last Tuesday in a different chat unless a separate memory feature is bolted on.

It also explains the most reliable way to get better output. Adding context is not politeness; it is your only lever over what gets predicted next, which is the argument behind learning these tools task-first.

The four kinds of tool you will actually meet#

General assistants. The chat tools — ChatGPT, Claude, Gemini and similar. You describe a situation, they respond. The most flexible category and the one to learn first, because the skill transfers.

Grammar and style checkers. Grammarly is the well-known example. These run over text you already wrote and flag errors, awkward phrasing and tone. Narrower, and safer for it: they correct your content rather than inventing their own, so the made-up facts problem barely applies.

Template-based marketing writers. Forms you fill in — product, audience, tone — to produce ad copy or social posts at volume. Underneath, a general model with a pre-written prompt around it. Useful if you write that copy all day. Close to pointless if you do not.

In-app suggestions. The assistant inside your email client, document editor or notes app. Convenient, because it can already see what you are working on. Usually less capable than the standalone version of the same model.

What the design is good at, and what it cannot be good at#

Split the work by whether it depends on form or on truth.

Why this changes how you use them#

One rule falls out of all this: supply the truth, let the tool supply the form.

Paste your rough notes and ask for a cleaner structure. Give it the real numbers and ask for the paragraph around them. Hand it the report and ask for a summary. The facts come from you; the pattern-matching does the part it is good at.

Run it the other way — "write a paragraph about our industry's growth" — and you have asked a system with no access to your industry to guess what such a paragraph usually contains. It will comply, fluently, and you will spend longer checking it than writing it would have taken.

The same reasoning explains why these tools suit plausible text rather than reliable rule-following. Moving data between systems on a schedule is an automation problem, and the tools built for it are different ones.

Frequently asked questions#

Why do AI writing tools make things up?#

Because they generate text by predicting likely word sequences, not by looking anything up. A false statement that sounds like normal English is as easy to produce as a true one. There is no internal check against reality, so invented details arrive in the same confident tone as correct ones. The behaviour is often called hallucination, and it cannot be fully trained away.

Does the tool remember what I told it last time?#

By default, no. Each conversation starts fresh with only what is in its context window. Some products add a separate memory feature that stores facts about you between sessions, but that is a layer built on top rather than part of how the model works. If a detail matters, include it again.

Are the paid versions actually better?#

Usually yes, but the difference is in capability rather than kind. Paid tiers typically offer stronger models, longer context and higher usage limits. The failure modes are identical: a paid model still invents citations. Free tiers are enough to learn on, and pricing changes often enough that you should check the current page before committing.

Can I tell whether text was written by AI?#

Sometimes, from style rather than any reliable test. Detection software produces both false accusations and misses, so it should not be trusted for anything consequential. What people actually notice is flat rhythm, cautious phrasing, and a recognisable set of overused words — the tells covered in the common mistakes guide.

Is what I type kept private?#

It depends on the provider and your settings. Some services use submitted content to improve their models unless you turn that off; business plans usually do not. Treat anything you paste as leaving your control, and keep client data, information about other people and anything under an NDA out of consumer tools. Our safety and privacy guide covers the settings worth changing.

What to do with this#

You now have a test to apply before you ask for anything: is this a form problem or a truth problem? Form problems get handed over. Truth problems get the facts from you first.

The practical next step is choosing something to use. The beginner tool guide sorts the options by the job you need done rather than by brand.