AI copywriting that still sounds like you

What AI is good at in copywriting, and what it isn’t

AI is a very fast, very well-read copywriter with no experience of its own and no knowledge of your business. Almost everything you need to know about AI copy follows from that.

My rule of thumb: AI is the ticket in, not the job. It takes away the blank page, not the decision about what you want to say. And because trust is a brand’s most important asset, no AI copy goes out from me without a human check.

The voice brief: your most important prompt

An AI assistant only knows what you give it. Positioning, tone and no-gos belong in a brief that sits before every copywriting prompt or lives permanently in the project (ChatGPT project, Claude project knowledge, Gemini Gem).

A ready-made template for this brief is in our 50 marketing prompts. In practice, the biggest jump in quality doesn’t come from cleverer prompts but from real sample copy in the brief.

Four frameworks that work well with AI

Copywriting frameworks suit AI perfectly: they give a clear structure the model reliably follows. What matters is that you supply or check the content of every step.

[Brand brief]. Write landing page copy for [offer] using AIDA. Attention: headline (max 10 words) and subhead. Interest: 3 sentences on the audience’s problem in their words: [customer quotes]. Desire: 3 benefits with proof from [proof]. Action: a button label and one sentence that lowers the barrier. No invented numbers.

Write 5 variants of a Meta ad for [offer] using PAS: problem in one sentence, agitation in one sentence without fearmongering, solution in one sentence, call to action. Max 125 characters of primary text. No income or health claims.

Here are the facts of a project: [starting point, action, result with numbers]. Write a case study using BAB (Before, After, Bridge) in max 200 words. Use only the facts provided; where something’s missing, write [check].

Score these 10 headlines on 4U (useful, urgent, unique, ultra-specific) at 1–4 points each: [headlines]. Suggest a better version of the 3 weakest. Only use urgency where it’s factually true.

Copy types: what AI handles for each

Topic and benefit: [topic]. Write 20 headlines, 4 each as: benefit promise with proof, question, number from [data], “How to …”, contrarian. Max 60 characters. Mark the 3 best for a test.

Here’s the spec sheet: [data]. Write a product description (max 120 words): first sentence = the main benefit for [audience], then 3 features framed as benefits, then who it isn’t for. Facts from the spec sheet only.

Here are my notes on a customer enquiry: [notes, without names or contact details]. Write a reply (max 150 words): answer the question directly, suggest a next step, friendly, no filler. No commitments that aren’t in my notes.

Here’s my newsletter topic with a personal observation: [note]. Write 3 openings (max 50 words) that start with the observation, not “In this issue”. Add 5 subject lines (max 45 characters).

Removing the AI sound: clichés to cut

Readers don’t spot AI copy with a detector; they spot patterns. When several of these come together, a text feels generic, whoever wrote it:

More to add to your brief’s banned list: “unlock the power of”, “navigate the landscape”, “elevate”, “seamless”, “robust”, “in the realm of”. German has its own set, such as „In der heutigen schnelllebigen Welt“ and „Tauchen wir ein“.

Check this text for AI-typical patterns: empty openers, unproven superlatives, “it’s important to note”, the rule-of-three reflex, “not only … but also”, empty endings, emoji bullets. Mark each one and suggest something more concrete. Don’t change content or facts: [text]

Rewrite this so it sounds like an experienced person talking: short and long sentences mixed, one fragment per paragraph allowed, no run of equal-length sentences, active verbs. Facts unchanged: [text]

AI detectors: why you shouldn’t rely on them

People often ask whether AI copy can be detected. The honest answer: not reliably. Two solid pieces of evidence:

• OpenAI itself released a classifier in January 2023. In its own evaluation it correctly flagged 26 % of AI-written text and wrongly flagged human text as AI 9 % of the time. Since July 20, 2023 it’s been unavailable because of its low accuracy.
• A study by Weber-Wulff and colleagues (2023) tested 12 publicly available and two commercial detection tools. The verdict: neither accurate nor reliable, biased towards classifying text as human-written, and simple obfuscation makes detection much worse.

For your marketing: don’t write for a detector, write for readers. Google says it judges quality, not the production method. Copy that helps, is specific and makes claims you can back up is the goal, with or without AI.

Copyright and advertising law

Who owns AI-written copy?

In the US, the Copyright Office concluded in its January 2025 report on copyrightability that, with current technology, prompts alone don’t make the user the author of AI output; human-authored expression in a work stays protected. In the EU, protection requires the author’s own intellectual creation; German law (§ 2(2) of the Copyright Act) protects only “personal intellectual creations”. Practical upshot everywhere: raw AI output is weakly protected at best, so a competitor may be able to reuse it. One more reason to build in your own experience, data and phrasing.

Advertising law: claims have to be true

Misleading claims about a product’s features, price or benefits are prohibited in the US (FTC Act) and the EU (unfair commercial practices rules; in Germany § 5 UWG). This matters especially for AI copy, because models invent plausible but false details: an award, a customer count, a deadline. And ads must be recognisable as ads, even when AI wrote them.

Tools around writing

For the writing itself, a chat assistant is enough. Two kinds of tools save noticeable extra time: dictation for your own raw material, and email assistants.

“Try” links are affiliate links (advertising): I earn a commission at no extra cost to you.

Pricing and free plans: best AI marketing tools.

The process: from brief to finished copy

For a short piece the process takes 15 to 30 minutes, for a landing page about half a day. Most of the time goes into step 4. That’s not a flaw in the process; it is the process.

[Voice brief]. Here’s my raw material (dictated, unordered): [text]. Turn it into 3 versions of a [copy type] with different openings. Keep my wording where it works. Add no facts.

Cut this text by 40 % without losing key points, proof or examples. Then list what you removed: [text]

Adapt this text for readers in [US/UK], not word for word: fitting examples, tone, units, date format. Flag anything that only applies legally in the original market: [text]

List every checkable claim in this ad copy (numbers, awards, prices, deadlines, customer counts, guarantees) and flag which could be misleading without proof. Don’t judge yourself whether they’re true: [text]

Emails and subject lines with AI

Email is where AI copy stands out fastest, because readers expect a person there. Three rules help: one goal per email, a personal opening from your own day, and a subject line that says what’s inside. For product recommendations, the ad label belongs right next to the link.

Here’s my newsletter: [text]. Write 12 subject lines (max 45 characters) in four styles: benefit, curiosity, question, number (only numbers from the text). Add preview text for each (max 80 characters) that complements rather than repeats. No all-caps words, no spam triggers.

[Voice brief]. I recommend [product] (affiliate link, I earn a commission). Write an email with one goal and one link: the problem, how I solve it, who the product suits and who it doesn’t, the price with the date checked, and “Ad · affiliate link” right next to the button. No fake scarcity, no invented experiences.

Pre-publication checklist

We run through these eight questions before any copy goes out. They take five minutes and catch most of the mistakes that make AI copy expensive:

1. Is every number right? Checked against the source, not the AI’s memory.
2. Is there a point only we could make? An observation, an example, an opinion of our own.
3. Does the first line sound like us? No “In today’s …”, no “Let’s delve in”.
4. Can every promise be backed up? No awards, customer counts or deadlines without proof.
5. Are ads labelled? Affiliate links and paid recommendations visibly marked.
6. Is the tone consistent? Same level of formality throughout.
7. Did the copy get shorter? The first AI draft is almost always too long.
8. Would we sign it with our name? If not, it isn’t finished.

The most common mistakes

1. Writing without a brief. You get the average of everything on the internet.
2. Taking the first draft. Three to five variants cost seconds and are almost always better.
3. Missing invented proof. Awards, customer counts, deadlines: anything that wasn’t in your raw material is suspect.
4. Inventing urgency. Fake scarcity reads as cheap and can be misleading.
5. Writing for detectors. Detectors are unreliable; readers and Google care about quality.
6. Not labelling ads. AI-written advertising is still advertising.