AI Text Generator

Google's published guidance turns on whether a page is helpful and made for people, not on which keys produced the words; automation is called out where it is used to churn out pages at scale with nothing to add. The risk in practice is the workflow rather than the tool — publishing generated text unread, on a subject you cannot vouch for, at a rate no human could have reviewed. A draft you have checked, corrected and added your own knowledge to is not different in kind from a draft you typed. Treat what comes out of here as raw material for a page rather than the page itself.

Not reliably, in either direction. Detectors score how predictable a text looks, and plain, competent, well-organised writing scores as predictable whoever wrote it — which is why they misfire on non-native English speakers, on technical documentation and on anyone taught to write simply. Light editing defeats them as well, so the two errors run in opposite directions at once. Where a school, journal or publisher has a rule about declaring AI use, that rule is about disclosure and applies whether or not a detector could tell. If you want the text to be genuinely yours, the work is the same work that makes it worth reading: your examples, your data, your sentences.

Your brief leaves your device. The page posts it to a route on this site, which forwards it to Anthropic's API to be answered and streams the words back as they are written — unavoidable for a tool of this kind, since the model doing the writing is not in your browser. What the route does not do is keep a copy: it writes nothing to storage, logs only its own failures, and the result lives in the page's state until you reload, which empties it. Requests are counted per address to stop a script running up a bill, so a rapid burst of runs can be refused for a few seconds. Keep client names, unreleased products and other people's personal details out of the brief.

Because producing plausible text is the whole of what it does, and a plausible statistic looks exactly like a real one from the inside. There is no lookup step behind this box — no search, no sources, no documents — so a figure in the output was produced by the same process that produced the sentence around it, and arrives in the same confident register whether or not it corresponds to anything in the world. Numbers, dates, names, citations, quotations and anything legal or medical are the categories to distrust by default. The safer shape of request is the one where you supply the facts in the brief and ask for them to be written up.

Give it the things it cannot guess. A word count, because there is no length control anywhere on the page, and without one you get whatever length the model considers reasonable. The reader, in a phrase — a district nurse, a sceptical finance director, the parent of a six-year-old. The two or three points that must appear, and anything that must not. Any facts of your own: names, figures, the real features of the real product. There is room to do this properly, because a single message can run to 40,000 characters, which is enough to paste in your notes, a transcript, or an older version to work from.

This is unsettled and varies by country, so take the following as background rather than advice. The United States Copyright Office has held that material produced by a machine without human creative input is not protected, while a work combining generated material with substantial human authorship can be protected in respect of the human part. Other jurisdictions have landed differently. Two practical consequences follow: a page of untouched generated text may be nobody's property to defend, and the more you rewrite, restructure and select, the more of it is recognisably yours. Client contracts and publisher agreements increasingly say something on the subject as well, so read them before you deliver.

A run is capped at 4,000 tokens of output — tokens being word fragments rather than whole words — so the practical ceiling is a few thousand words, and a run that reaches it stops mid-sentence with a bracketed note explaining why. There is also a sixty-second budget for the whole request. Asking for a five-thousand-word guide is therefore the wrong shape of request, while asking for its introduction and then each section in turn works. Since runs do not remember one another, a section written separately has no idea what the previous one said, so repetition between sections is the thing to watch for when you stitch them together.

Ask for it in the brief. The two dropdowns cover format and tone and nothing else, so spelling, house style, banned words, sentence length, whether to use the Oxford comma and whether to address the reader directly all have to be stated in your own text. Say British English explicitly if you want colour and organise rather than color and organize, since these models are trained mostly on American text and drift back towards it. A style rule stated once tends to hold through a short piece and to fray in a long one — another argument for generating in sections, and for reading the tail of a long output as carefully as the head.