With AI writing, you get usable text when you give the language model a full brief, rewrite the first draft paragraph by paragraph and check every fact yourself. Writing is one of the uses of AI where the brief matters more than the tool.
A language model writes on the basis of what you give it. The goal, the reader, the core message and the source material decide whether the text is about your situation or about an average one.
Text from a language model often gives itself away through formal words, long sentences and fixed structural patterns. With a sample of your own style and fixed style instructions, the text sounds more natural.
Facts, figures and sources in an AI-written text are not reliable without checking. How much checking a text needs also decides which texts you leave to AI and which you write yourself.
How do you write a text with AI?
To write a text with AI, you work in four steps: brief, first draft, rewriting paragraph by paragraph, and final edit. If you have a text written in a single prompt and use it straight away, you skip the two steps where the quality is created.
I used to have a fully automatic content engine in Python that wrote and published articles, and the quality was poor. Now I write with Claude and read every text myself; only the publishing is still automated.
When rewriting, give one targeted instruction for each paragraph, for example “make this paragraph shorter and put the example first”. Ready-made ChatGPT prompts for work give you a starting point, but adapt them to your own text and reader.
What input does AI need to write a good text?
To write a good text, AI needs five kinds of input: the purpose of the text, the reader, the core message, the source material and an example of the style you want. Wherever you leave input out, the model fills the gap with generalities.
Source material is the most important input, because according to Google a generative model does not retrieve facts but predicts a likely sequence of words. So paste the data, quotes and figures that belong in the text into the prompt yourself. When you give context to an LLM, the quality of that information counts for more than the quantity: the more targeted the context, the less the model fills in on its own.
A style example works better than a description of the tone. A paragraph from an earlier text you liked shows sentence length, word choice and form of address in one go.
How do you make AI writing sound natural?
You make AI writing sound natural by giving fixed style instructions and deliberately removing the recognisable AI patterns. Those patterns include formal words such as “moreover”, long sentences full of subordinate clauses, lists of three and announcements like “it is important to note”.
In my experience, Claude follows instructions more strictly than ChatGPT, especially in complex prompts with several conditions. A side effect is that the output sounds less like AI, provided you set up the instructions well.
Save those instructions once instead of typing them again for every text. I work with one project per client, with the client context in separate files. With Claude Projects, style rules, sample texts and banned words are ready for every new text in that project.
How do you check a text that AI has written?
You check an AI-written text on four points: facts, figures, sources and names. Look up each of those points in a source you trust yourself, and delete anything you cannot find.
A recurring pattern I saw with ChatGPT was made-up sources and quotes: studies and reports that turned out not to exist when I checked. A text that builds authority through external references is particularly at risk.
In its guidance for Search, Google recommends checking all AI content manually for accuracy before publication, including titles, meta descriptions and alt text. Finally, read the text aloud: the sentences you stumble over are often the ones that sound like AI.
Which texts is AI suitable for?
AI is suitable for texts with a clear structure and facts you can supply: business emails, product descriptions, social posts, summaries and first drafts of longer pieces. With these texts, the brief provides most of the content and the model writes the form.
Rewriting is a second strong use. Turning an existing text into plainer language, adapting it for a different audience or producing a shorter version works well, because the content is already fixed in the source.
When is it better to write a text yourself?
It is better to write a text yourself when the content consists of your experience, your point of view or confidential information. A personal message to a client, advice that reflects your own judgement or a text about a sensitive situation loses its value when a model fills in the content.
In those cases, AI is still useful for the finishing touches: spelling, sentence length and structure. The content and the judgement then come from you, and the model only adjusts the form.
Can Google tell that a text was written with AI?
Whether Google can recognise it is not the test: according to Google Search Central, using generative AI to create content is allowed. Mass-generating pages without added value does fall under its spam policies.
AI text detection tools are not a reliable test. On 20 July 2023, OpenAI withdrew its own detector because of its low accuracy. In OpenAI’s own evaluation, the detector identified about a quarter of AI-written texts and wrongly flagged 9 per cent of human-written texts.
What is the best AI tool for writing?
For writing, use a general AI chatbot such as Claude, ChatGPT or Gemini, ideally in an environment where you can save style instructions and sample texts. Claude has Projects for this, ChatGPT has projects and the free Gemini app has Gems. Google is replacing Gems with skills: on 17 November 2026 Gems become read-only, and on 17 December 2026 they disappear from the Gemini app for personal accounts.
Frequently asked questions about AI writing
Alongside the writing steps, AI writing quickly raises legal and practical questions.
Do you have to disclose that a text was written with AI?
Under Article 50(4) of the EU AI Act, disclosing that a text was written with AI has been mandatory since 2 August 2026 if you publish the text to inform the public on matters of public interest. That obligation does not apply if a human has reviewed or edited the text and someone holds editorial responsibility for the publication. For texts that are not about matters of public interest, such as a product description, Article 50 does not mention this obligation.
Can AI write a text for you for free?
Yes, you can have AI write a text for free with the free versions of ChatGPT and Claude (you need an account) or the free version of Gemini. Those free versions have limits: ChatGPT restricts messages with uploads, among other things, and Claude restricts usage per five hours and per week. For the occasional email or short text that is enough, but for daily writing work with long documents you will soon hit those limits.
Who owns a text that AI writes for you?
According to the terms of OpenAI and Anthropic, a text that AI writes for you belongs to you, as both assign any rights they have in the output to you. OpenAI points out that output is not necessarily exclusive to you and that other users may receive similar text. The terms say nothing about whether the text is protected by copyright, and they only transfer rights to the extent that the provider holds them.
Where can you learn AI writing you can stand behind?
You can practise responsible AI writing in LearnLLM’s Claude Course. You set up a Project with fixed instructions and write a system prompt that makes Claude work consistently in your style. Each module ends with a practical assignment based on your own work.
