We are three years past the moment AI writing tools became unavoidable and the conversation about them is still curiously bad. The reviews compare them on prose quality. The benchmarks compare them on hallucination rate. The marketing compares them on the underlying model. None of these are the question that actually matters when you sit down to use one.
The question that matters is which part of writing the tool is trying to do for you. There are at least four distinct jobs that get bundled under 'AI writing,' and the tools are good at very different ones. Choosing the wrong category is more expensive than choosing the wrong tool within a category.
What follows is the taxonomy I use to sort them, with a recommendation in each category and an honest note on where each one breaks.
Category one: the drafter
Drafters produce a first draft from a brief. You type 'write a 600-word blog post explaining X to a non-technical audience,' and you get 600 words back. The best of these have gotten genuinely good. The worst of them are still embarrassing.
The honest use case here is the one nobody likes to talk about: the email you have been avoiding, the project update nobody will read closely, the LinkedIn post you are writing only because the algorithm rewards consistency. For genuine writing — the kind a real reader will engage with — drafters produce something that needs to be substantially rewritten. The economics tend to be unfavorable. You spent twenty minutes on the brief and the edit. You could have spent twenty-five and written something better, in your own voice.
Where drafters are unambiguously good: localization, format conversion (turn this transcript into a recap), and the long tail of writing that does not need to be excellent, just present.
Category two: the polisher
Polishers take writing you have already done and clean it up. Grammar, flow, tone. These are the most successful of the AI writing tools — the technology has been good enough at this for a decade, and the modern LLM-based versions are just the polished surface on top of work that started with Grammarly in 2009.
The trap with polishers is that they have a strong, consistent aesthetic. They like short sentences. They dislike commas. They smooth over the rhythm choices that distinguish a writer with a voice from a writer without one. Used heavily, every polished piece starts to sound the same, and that sameness is the polisher's voice, not yours.
The right way to use a polisher is as a copy editor, not as a co-author. Accept the corrections that are obviously corrections — typos, missed words, broken parallel structures. Reject the suggestions that are stylistic preferences dressed up as improvements. The discipline takes some practice and is worth developing.
Category three: the summarizer
Summarizers compress long input into shorter output. Meeting transcripts into recaps. PDFs into bullet points. A long Slack thread into a one-paragraph TL;DR. These are also genuinely useful, and have been since before the current AI wave — the difference is that the modern versions are good enough to trust on first pass for most low-stakes documents.
What summarizers do well: they extract the explicit claims. What they do badly: they miss the implicit ones. A meeting summary will give you the agenda items. It will not tell you that the head of marketing and the head of engineering disagreed about whether the new feature is ready, even if everyone in the meeting could feel it.
This is fine when you are summarizing a document. It is dangerous when you are summarizing human conversation, because you will get a summary that reads like consensus when the meeting was actually a small political defeat for one of the participants. Use accordingly.
Category four: the thinker
Thinkers are conversational tools you use to develop an idea. You bring a vague intuition, you talk it through, you leave with a sharper version of the same intuition. This is the most interesting category and the hardest to get right.
Done well, a thinker is a sparring partner. You say 'I think X is true,' and it pushes back, offers counterexamples, asks the second-order question you would have asked yourself eventually. The output is not the conversation. The output is what you understand after the conversation.
Done badly, a thinker is a yes-machine. It agrees with everything. It compliments your premise. It restates your idea back to you in slightly more elegant language and you mistake the elegance for confirmation. Every major model has this failure mode and you have to actively prompt around it.
The single best prompt I have found for surfacing the thinker mode: 'Argue against this. Find the strongest objection a serious critic would raise.' Almost any modern model will give you a useful response to this if you ask. None of them will give you a useful response if you do not ask.
The category that does not exist
There is one category that is endlessly promised and that does not yet meaningfully exist in 2026: the AI that writes in your voice. Personalization at the level of 'this sounds like me' is something every tool claims and none deliver. The technical reasons are interesting and not worth belaboring; the practical implication is that if you have a voice you care about, none of these tools will preserve it. You will, instead, have to do the work of editing the model's voice out of every piece, which is its own kind of writing exercise.
This is not a complaint. It is a constraint. The constraint determines what AI writing tools are good for. They are good for writing where the voice does not matter. They are good for first drafts of forms, summaries of documents, polish on emails, and conversations about ideas. They are not yet good for the kind of writing where the words are the product.
A workflow that works
The pattern I have ended up with, after a lot of false starts:
- 1Think with a thinker. Ten minutes of structured argument before you start writing, with explicit prompts to disagree.
- 2Outline by hand. The outline is the actual thinking. Do not skip this.
- 3Draft yourself. Even if it is slower. The voice is built in the draft.
- 4Polish with a polisher, lightly. Accept the typo fixes. Reject the stylistic edits unless you actively prefer them.
- 5Summarize with a summarizer only if someone else asked for a summary. Otherwise the summary is for nobody and you are wasting the work.
This is, in essence, the workflow of a person who likes writing and uses AI tools to make the parts they do not enjoy faster. The opposite workflow — draft with the drafter, edit yourself, polish with a polisher — produces work that is worse and takes longer, in my experience. Your mileage may vary.
The honest summary
Three years in, AI writing tools have not replaced writing. They have unbundled it. The parts that were always tedious (the email, the recap, the form, the localization) are now mostly free. The parts that were always interesting (the argument, the voice, the surprise) are still done by humans, and the humans who are good at them have not gotten less valuable. They have, if anything, gotten more.
If you treat the tools as junior collaborators with very particular strengths, they make your work better. If you treat them as a replacement for the parts of writing that are genuinely hard, the result is a piece of writing that nobody, including the model, would describe as good. The choice is mostly about which mental model you bring to the tool, not which tool you choose.
The AI category page on dotstore is sorted by usefulness, not by buzz. If a tool fits one of the categories above and does the job well, it is on there. If a tool is mostly press releases and a wait list, it probably is not.
Tags