How AI writing tools actually work.

If you've used a few AI writing tools, you've probably noticed that two tools using the "same AI" can give you wildly different results. One summary is clear and to the point; the other reads like a tired LinkedIn post. The model is identical. What's different is everything around it.

This piece is a short, non-technical walkthrough of what's actually happening when you hit "generate" on a tool like the ones on this site. It won't make you a machine-learning researcher, but it'll make you a more effective user — someone who gets useful output the first time instead of the third.

The model is a prediction machine

Strip away all the marketing language and a large language model — the thing under the hood of ChatGPT, Gemini, Claude, and so on — is a very sophisticated next-word predictor. You give it some text. It looks at the patterns it has learned from training on enormous amounts of human writing, and it predicts what word should come next. Then it predicts the next word after that. And the next. Tens or hundreds of words later, it stops.

That's it. That's the trick. The reason the output sounds intelligent is that to predict the next word well, the model has to have absorbed a great deal of how language works, what facts go with what topics, and how arguments are usually structured. It hasn't "understood" anything in the way a human does, but in terms of producing fluent, on-topic text, the prediction approach turns out to be remarkably effective.

What a "prompt" really is

Every AI writing tool is, underneath, just a structured prompt being sent to a model. A prompt is the text you give the model to start from. Whatever you typed into the tool is part of the prompt, but it's usually not the only part. The tool's author has wrapped your input in additional instructions — telling the model what role to take, what format to output in, what mistakes to avoid.

That wrapping is most of what separates a good tool from a bad one. A poorly built summarizer might pass your text to the model with the instruction "summarize this." A well-built summarizer says something more like: "You are summarizing for a busy reader. Use bullet points. Keep each bullet under 15 words. Capture the key argument, the most important evidence, and any concrete examples. Do not editorialize. Do not include filler like 'in conclusion'." The model is the same. The output is dramatically different.

The model is the same. The output is dramatically different.

This is why "AI is the same everywhere" isn't quite true. The model might be, but the prompt almost never is.

Why output varies between runs

If you generate the same thing twice, you'll usually get two different outputs. That's not a bug — it's a setting called temperature. Temperature controls how predictable the next-word predictions are. At a low temperature, the model picks the most likely next word almost every time, and outputs are repetitive and safe. At a high temperature, it samples from a wider pool of possibilities, and outputs are more creative but less reliable.

Most tools sit somewhere in the middle. A good caption generator uses higher temperature on purpose — variety is the point. A good translator uses lower temperature — you want consistency, not creative interpretation.

Why the model sometimes makes things up

The polite word is hallucination. The model is built to produce plausible-sounding text, not to retrieve true information. If you ask it to summarize an article you've pasted in, it has the article in front of it, and the summary will mostly stick to what's there. If you ask it to write about a person or a paper without giving it the source material, it will sometimes confidently invent details — a quote that was never said, a paper that doesn't exist, a date that's wrong.

This is the single most important thing to know about using AI writing tools well: treat any factual claim as a hypothesis until you've verified it elsewhere. Especially numbers, names, dates, and quotes.

Why some tools cost money and others are free

Running an AI model is not free. Every request you send costs the tool's operator some fraction of a cent — sometimes more, depending on the model and the length of the response. That fraction adds up. A free tool has to pay for its requests somehow: with ads, with sponsored content, with a paid tier hidden behind the free one, or by being run by someone with deep pockets who's hoping to recover the cost later.

Free with ads is, in our view, the most honest model. You see what you're paying with (your attention), you can leave whenever you want, and there's no signup wall pretending you'll get more value once you give up your email.

Things you can do to get better output

None of these require learning prompt engineering. They just require being a little more deliberate.

1. Give context up front. "Rewrite this email" is okay. "Rewrite this email to my boss, who I haven't spoken to in two weeks, asking for an extension on a deadline" is much better.
2. Be specific about format. If you want bullets, ask for bullets. If you want them to be one sentence each, say so.
3. Say what to avoid. "Don't use the word 'leverage'." "No exclamation points." "Don't add a closing greeting." These work surprisingly well.
4. Give an example if you can. Even a one-line example of the style you want shifts the output far more than a paragraph of description.
5. Iterate, don't restart. If the first output is 80% right, ask for changes to the 20% rather than starting over. "Same thing, but shorter and less formal."

What AI writing tools are not

They are not a search engine, even though they sometimes sound like one. They are not a calculator, even when they confidently produce numbers. They are not a fact-checker. They are not a replacement for the part of writing where you decide what you actually think — they're best at the part where you've already decided and just need help saying it well.

Used that way, they save time and often make your writing clearer than the version you'd have produced alone. Used the other way — as a substitute for thinking — they produce the bland, fluent, slightly-off prose that has people complaining about "AI slop."

The model is a prediction machine. You're the one with something to say. That order matters.

---

More reads on this site: The truth about AI detectors · Six prompts that make a big difference.