How to Make AI Writing Sound Human: 15 Giveaways to Delete (With Before and After)

You can tell within a sentence. The email "hopes this message finds you well", the post promises to "delve into" something, and a simple offer is described as "not just a service, but a partnership". Nobody talks like that. Your customers have read enough of it by now to recognise it, and they are not impressed: in a June 2026 survey of 408 consumers by Clutch, 33% said the use of AI made them think worse of a brand, against 16% who thought better of it.

The answer is not to stop using AI. It is to stop publishing its first draft. This guide shows how to make AI writing sound human: the 15 giveaways to delete, a five-minute edit with a real before and after, the prompts that prevent the problem, and the truth about detectors and Google. It was written with the same rules it recommends.

How we checked this guide

1. A public catalogue of the signs. The giveaways draw on the list that Wikipedia's editors maintain to spot AI-written articles, read on 4 October 2026.
2. Published research on vocabulary and detection. The studies are named where they are used.
3. The companies' own statements. What OpenAI and Google have said comes from their announcements and guidance.
4. A warning that comes with the list. These are patterns, not proof. People wrote this way before AI, and some writers still do.

The 15 giveaways

The vocabulary in row 2 is measurable. A 2025 study in Science Advances examined more than 15 million biomedical abstracts and found a sudden rise in certain style words after ChatGPT appeared. The authors estimate that at least 13.5% of abstracts published in 2024 were processed with a language model, and the word that gave the paper its title was "delving".

Why AI writes like this

A language model is trained to produce the most likely next word. The most likely phrase is, by definition, the one everybody has already used. Ask for "a friendly email to a customer" with no other detail and you get the average of a million friendly emails: pleasant, complete and nobody's.

Two more habits follow from how the tools are built. They are trained to be helpful, so they cover every angle and hedge every claim. And they are trained to be agreeable, so they praise, summarise and offer more. None of that is a fault you can prompt away entirely. It is the starting point you edit from.

The five-minute edit

Take any AI draft and do these five things in order.

1. Delete the first sentence and the last. The first is usually a warm-up and the last a summary. The message is in between.
2. Swap one general claim for one specific fact. "We offer great value" becomes "A boiler service is £95, the same as last year".
3. Cut every word from row 2 and row 7. If the sentence falls apart without "vibrant", it was not saying anything.
4. Break the rhythm. Make one sentence short. Join two others. Humans do not write in even beats.
5. Read it aloud. Anything you would not say to the customer's face comes out.

Before and after

A real first draft, the kind every assistant produces:

In today's competitive market, choosing the right electrician is a pivotal decision. At Carter Electrical, we pride ourselves on delivering not just exceptional workmanship, but a seamless customer experience. Our team of dedicated professionals is committed to quality, reliability and transparency, ensuring peace of mind for every client. Get in touch today to discover how we can illuminate your home.

After five minutes:

I am Jamie Carter, a qualified electrician in Leeds. I rewire houses, replace consumer units and fix the faults other people could not find. Every job comes with a written quote, a certificate and a 12-month guarantee. Call 0113 496 0000 or message me a photo of the problem.

The second version is shorter, makes four claims a customer can check and contains no word you would not say out loud.

Prompts that stop it at source

Editing is faster when the draft starts closer. These five instructions do most of the work:

• Show it your writing. "Here are three emails I wrote. Match the tone, the sentence length and the sign-off."
• Ban the tells. "Do not use these words: delve, pivotal, seamless, vibrant, testament. No long dashes. No sentence starting with 'In today's'."
• Name the reader. "Write to a landlord who has two minutes and wants a price."
• Ask for facts first. "Before you write, ask me for the three facts this message has to contain."
• Set a ceiling. "Under 90 words. One idea per sentence."

The punctuation ban is easier than it used to be. In November 2025 Sam Altman announced that ChatGPT would finally follow an instruction not to use long dashes when it is saved in custom instructions. Save your rules once, in custom instructions or a saved skill, and you stop typing them. Our 30 prompt examples and 18 AI email writing tips go further.

Do not trust AI detectors

If the worry is that a customer or a client will run your text through a detector, the evidence should reassure you and warn you in equal measure.

• The makers gave up. OpenAI launched its own AI text classifier in early 2023 and withdrew it on 20 July 2023 because of its low accuracy. In OpenAI's own tests it correctly identified 26% of AI-written text and wrongly labelled 9% of human text as AI.
• Detectors punish plain English. A Stanford study published in the journal Patterns in 2023 ran seven detectors on essays written by non-native English speakers. On average, 61.3% of those human-written essays were flagged as AI.
• People are not much better. Wikipedia's own guidance tells its editors not to rely on detection tools, and cites research in which people identified AI text correctly only 57% of the time.

So a detector score proves nothing, in either direction. Write for the reader. If a human would enjoy it, the question of who typed it stops mattering.

What Google actually says

Google does not rank a page down for being written with AI. Its guidance, updated on 1 October 2026, asks for accuracy, quality and relevance, and tells site owners to fact-check AI output before publishing. What it treats as spam is producing many pages with little value, by any method.

That fits everything above. The giveaways are not a problem because a machine wrote them. They are a problem because they take up space and say nothing. Remove them and the page gets shorter, more specific and more useful, which is what both readers and search engines reward. The 10-step method for SEO and AI search builds on the same idea.

When to say you used AI

There is no general UK rule that an AI-assisted email or web page must carry a label. Three situations are different:

• A customer is talking to a bot. Say so. People are entitled to know whether they are talking to a person.
• Reviews and testimonials. Fake reviews, including AI-written ones, have been banned in the UK since 6 April 2025.
• Facts someone will rely on. If AI helped with prices, legal wording or advice, a person must check it and stand behind it.

Google adds that telling readers how content was made "can help give your readers more context". Use your judgement: nobody needs a label on a tidied-up email.

How an AI agent helps you sound like yourself

ChatGPT, Claude, Gemini, Microsoft Copilot, Manus and Meta's Muse all write the same first draft if you give them the same vague request. The difference comes from what the tool knows about you before it starts.

• Learns your voice once. On Jobbit you can keep your writing samples and your banned-word list in a project, so every draft starts from them.
• Asks before it writes. The agent can interview you for the facts instead of inventing filler.
• Lets you choose the model. Some models write plainer prose than others, and you can pick per task.
• Finishes the job. The email goes into a document, the page goes on the website, the post gets its image.
• Starts free. 100 credits a day with no card. Paid plans begin with Go at £7 a month.

For everyday use, the 25 ChatGPT tips apply to any assistant.

Frequently asked questions

How do I make AI writing sound human?

Delete the first and last sentences, replace one general claim with a specific fact, cut the stock words, vary the sentence length and read it aloud. Then prevent the problem by giving the AI samples of your own writing and a list of words to avoid.

What words give away AI writing?

Delve, tapestry, testament, pivotal, intricate, seamless, vibrant and "serves as" are among the best known, along with "not just X, but Y" and long dashes. A study of 15 million abstracts found these style words rose sharply after 2022.

Can AI detectors tell if I used ChatGPT?

Not reliably. OpenAI withdrew its own detector in July 2023 for low accuracy, and a Stanford study found detectors wrongly flagged 61.3% of essays by non-native English writers as AI.

Does Google penalise AI-written content?

No. Google asks for accurate, useful, original content however it is made, and penalises pages produced at scale with little value. A well-edited page written with AI help is fine.

Should I tell customers I use AI?

Tell them when they are talking to a bot, never publish AI-written reviews, and check anything they will rely on. For ordinary emails and web copy there is no general duty to label.

What is the quickest way to stop sounding like a robot?

Give the AI three things you wrote and ask it to match them. Try it free on Jobbit with 100 credits a day and no card.