AI & automation
How do you make AI-written content sound human?
The short answer
Stop asking one prompt to do everything. Write down the specific tells that give AI copy away, keep your writing rules separate from your brand rules, and run every draft through stages: brief, draft, rules check, an AI-filter pass, then a human. Every time a line makes you wince, add it to the rules so the next draft is better.
Why doesn’t “don’t sound like AI” work?
Because the model has no idea what you mean. You’ve handed it a vibe. It’ll agree with you, then write a sentence built on “it’s not X, it’s Y” anyway.
We kept finding exactly that in AI-drafted copy, even with an instruction to avoid it sitting right there in the prompt. That one cliche is the biggest giveaway of the lot, and a general instruction won’t catch it.
What works is getting specific. It should really be as simple as working out what the sentence structures are, what the phrase structures are, what the cadences are, and which particular words give away that it was written by AI. Once you’ve named them, you can check for them.
What goes on your list of AI tells?
Start with the “it’s not X, it’s Y” contrast. Then add the phrases that keep showing up in your drafts. Openers like “here’s the honest answer” were one of ours. Then look at rhythm: AI loves sentences that all march along at the same length. Then the words it reaches for that you’d never say out loud.
Keep two rulebooks while you’re at it. There should be one group of guidelines for content and one group of guidelines for a brand, and they’re different. Your content rules cover how good writing works and the tells to avoid, and they apply to everything you publish. Your brand rules cover your tone of voice, the words you use and the claims you can and can’t make. Mixing them into one giant prompt is how both get ignored.
How should you split the work into stages?
Give each stage one job. This is the order I’d run:
- Brief. Pick the content type, whether that’s a blog, a promo email or a lead magnet, and load the brief for that type so the model knows what it’s making.
- Draft. Write it against your tone of voice guide and brand guidelines.
- Check. Review the draft against your do’s and don’ts guide and make sure it hasn’t made any mistakes or claims you can’t back up.
- Filter. Run a separate AI-filter pass whose only job is to find the tells on your list and strip them out.
- Review. A human reads it and fact-checks it before anything goes live.
It’s also worth trying different models for different stages. The one that’s really good at writing prompts is not going to be the one that’s really good at doing the work. Test a few and keep whichever does each job best.
What do you do when a draft still gets it wrong?
Fix the rule, then fix the draft. If you only fix the draft, you’ll be fixing the same thing next week.
The easiest way I know is one note on your phone. Whenever you read something and think “I can’t have that out there”, jot down what triggered it. Maybe it’s a claim that isn’t quite true. Maybe the tone has drifted into braggy LinkedIn territory, or a phrase screams AI.
That note gets fed back into your tone of voice guide and your compliance documents. From there it feeds the prompts, and it feeds the scan that runs on drafts before you ever see them. Of everything in the setup, the do’s and don’ts guide is probably the most important. Every cycle it gets a bit longer, and the drafts get a bit tighter.
Why does the voice suddenly drift?
Sometimes it’s as simple as the model changing. A new version comes out, it skips part of the memory file or reads your instructions a bit differently, and the copy starts to wobble.
When you notice it, flag it and go back over the prompts. Small changes in a prompt have big effects downstream, so it’s usually a quick fix once you’ve found the line that’s being ignored.
Do you have to review everything forever?
No, but you do at the start. Like all things, at the beginning it’s going to need a bit more hand-holding. I review the first batch myself, line by line. For anything technical, I get someone who knows the subject to check the detail, because I won’t catch every technical mistake on my own.
That first batch is the most important one to get right, because it shows you where the AI is on the mark and where it’s off. Each round of feedback goes into the rules, so the next batch needs less from you.
Once you’re confident it isn’t making things up and the tone is right, you can loosen your grip. We know the tone of voice is right, so let it loose. Spot-check instead of reading every line, and when the odd thing slips through, fix it and add it to the note.