Does AI sound like AI? Control style without saying “write more humanly”

When you read a text generated by ChatGPT and after three sentences you already know: this sounds like AI. Everything is correct, smooth, and unsettlingly polite. Every paragraph is about the same length. Phrases like “key aspects,” “the dynamically changing world,” and the final “in summary” keep appearing.

So you ask: “Write more humanly.”

The model improves the text. Instead of “In today’s dynamic world,” you get “Hey! We live in really interesting times.” In theory, it’s more relaxed; in practice, it still doesn’t sound like something you’d write to a client, an employee, or a newsletter audience.

The problem is not that ChatGPT doesn’t understand natural language. The problem is an imprecise instruction. “Humanly” can mean factual, conversational, warm, direct, witty, or concise. Until you specify what you expect, the model has to guess.

Why “write naturally” usually isn’t enough

Style is not a single slider with “robot” and “human” values. It consists of many elements:

• sentence length and rhythm,
• word choice,
• level of formality,
• how the writer addresses the reader,
• amount of detail,
• emotional tone,
• text structure,
• acceptable humor,
• industry phrases and expressions to avoid.

The instruction “write naturally” doesn’t define any of these. ChatGPT therefore falls back on a safe, averaged style. Such a text won’t offend anyone, but it often won’t interest anyone either.

The good news: you don’t need a prompt the length of a mortgage contract. It’s enough to give the model a few pieces of information that actually affect the result.

Five prompt elements that change style

The most useful style instructions answer five questions.

1. Who will read the text?

“Customer” is too vague. You write differently for the owner of a local beauty salon, differently for a sales director in a corporation, and differently again for someone buying an online course for the first time.

Instead of:

Write a post for entrepreneurs.

Better:

Write a post for owners of small service businesses. These are people who handle marketing themselves, have little time, and don’t know specialized advertising language.

The model then gets guidance on vocabulary, level of detail, and examples.

2. What should the tone be?

“Professional” doesn’t add much. Professional can be both a bank’s terms and conditions and a direct email from a friendly advisor.

Combine tone with the author’s behavior:

Write in a factual and direct way, but not coldly. The author knows the topic but is not trying to impress with jargon. Address the reader as “you.”

You can also define the tension between two traits:

Calm and competent tone, with a slight sense of ease. No excessive enthusiasm and no forced jokes.

That is much clearer than “write nicely.”

3. What does a good example look like?

An example can be more effective than a whole list of adjectives. If you have an email, post, or paragraph whose style you like, show it to the model.

Model the rhythm of this fragment: “You don’t need another tool. First check whether you’re actually using the ones you already pay for. In most companies, the answer is: not really.” Don’t copy the sentences or arguments. Keep the same simplicity, directness, and varied sentence length.

The last caveat matters. The point is style features, not rewriting someone else’s text word for word.

4. What should the model not do?

Restrictions are especially useful when the same language habits keep coming back in later responses. You can create your own blacklist.

For example:

Don’t use phrases like “in today’s times,” “key,” “it’s worth noting,” “extremely important,” “game changer,” or “in summary.” Don’t start several consecutive paragraphs with questions. Don’t use exclamation marks. Don’t create a fake contrast like “It’s not X. It’s Y.”

But don’t add twenty bans just in case. Add only the ones that solve a real problem. If the text is meant to be a short sales post, the list of restrictions may differ from an internal team instruction.

5. How will you know the text is good?

This is often overlooked, but it’s very practical. Instead of judging the result with “something feels off,” define criteria.

The text should be understandable after one read. Each paragraph should add new information. Use at least one concrete example. Sentences should vary in length. Remove sentences that only repeat an earlier idea.

Criteria help both the model and you. It becomes easier to point out what needs revision.

Before/after: a post promoting a service

Let’s assume you run a company that implements CRM systems for small sales teams.

General prompt

Write a natural LinkedIn post about why it’s worth implementing a CRM.

Typical result

In today’s dynamic business world, effective customer relationship management is the key to success. A CRM system helps organize sales processes, increase team efficiency, and better respond to customer needs. It’s not just a tool, but a true game changer for any growing company. If you want to take your sales to the next level, it’s worth considering CRM implementation.

The text is correct, but it could promote any system in any company. There’s no real-life scene, no clear audience, and no reason to stop scrolling.

Prompt with style instructions

Write a LinkedIn post for owners of small B2B companies who manage a team of 3 to 10 salespeople. Many of them still keep customer information in spreadsheets, email inboxes, and notes.

Tone: factual, direct, and slightly relaxed. The author is a practitioner who knows the realities of a small business. Don’t scare people and don’t use corporate language.

Start with a concrete situation: a salesperson goes on vacation, and the rest of the team doesn’t know what stage their conversations are at. Show the cost of the chaos, and only then explain the role of CRM.

Don’t use phrases like “in today’s world,” “the key to success,” “game changer,” or “take it to the next level.” No exclamation marks or emojis.

Criteria: 900–1200 characters, short paragraphs, one example, one question at the end. Each paragraph must add new information.

Result after changing the prompt

A salesperson goes on vacation for a week. On Monday, a client calls, but no one knows what was agreed on or whether the offer has already been sent.

Someone searches the inbox. Someone checks the spreadsheet. After a few minutes, the sentence appears: “Let’s wait until they’re back.”

This is not a team engagement problem. The information simply lives in five different places.

A CRM makes sense when it ends this kind of searching. It shows the contact history, the next step, and the person responsible. It won’t sell for the salesperson, but it will make sure the client doesn’t depend on someone’s memory or private notes.

Before choosing a system, check one thing: how many times in the last month has your team spent more than two minutes looking for customer information?

The difference doesn’t come from a magic command. The model was given the audience, the scene, the tone, the constraints, and the quality criteria.

Before/after: an email to a client

AI style shows up not only in posts. Sometimes a simple email sounds so formal it feels like it concerns international treaty negotiations.

Instruction

Write a more human email saying the project delivery will be delayed by three days.

Result

Good morning,

we would like to inform you that, due to circumstances beyond our control, the project completion date will be slightly postponed. We are making every effort to finish the work as quickly as possible. We apologize for any inconvenience and thank you for your understanding.

It sounds polite, but it hides the most important information. What happened? What is the new deadline? Does the client need to do anything?

Better prompt

Write an email to a regular client with whom we use formal “Mr./Ms.” language, but without an official tone. State clearly that we will deliver the project on Thursday instead of Monday. The reason is the need to recheck the payment integration after finding a bug in testing.

Take responsibility, don’t blame subcontractors. Explain what we’re doing and when the client will get the next update. Don’t use phrases like “due to circumstances beyond our control,” “we are making every effort,” or “inconvenience.” Maximum 120 words.

Result

Good morning Ms. Anna,

we will deliver the finished project on Thursday, three days later than planned. During testing, we found a bug in the payment integration, and we want to check the whole process again instead of handing over a site with a risk of order issues.

Today we’re fixing the integration, and tomorrow we’ll run additional tests. On Wednesday by 2:00 p.m., I’ll send a short update and confirm Thursday’s handover.

I apologize for the change in schedule. No action is needed on your side right now.

Best regards,
Marek

Naturalness here doesn’t come from adding “Hi” and a smiley face. It comes from specificity, responsibility, and a clear order of information.

A universal prompt template for controlling style

The following framework works for emails, offers, articles, product descriptions, and posts:

Write [type of text] for [specific audience and their situation].

Text goal: [what the reader should understand, feel, or do].

Tone: [2–3 traits]. The author sounds like [role or behavior], not like [undesired association].

Context and facts: [information the model should not invent].

Model these style features from the example: [fragment or style description]. Don’t copy the wording.

Don’t use: [banned phrases, constructions, habits].

Criteria: [length, structure, number of examples, CTA, level of detail].

Before answering, check whether the text meets all criteria. If data is missing, ask at most three questions.

You don’t always need every field. For a short message, audience, goal, tone, and restrictions may be enough. For an important sales offer, it’s worth adding an example and evaluation criteria.

Improve the instruction, not just the finished text

A common workflow looks like this:

1. the user enters a general prompt,
2. gets an average answer,
3. asks for “shorter,”
4. then “more natural,”
5. then “less salesy,”
6. after the sixth revision, rewrites everything manually.

It’s better to treat the first version as a diagnosis. Mark the specific problems and turn them into rules for the next response.

Instead of:

It still sounds like AI. Try again.

Write:

This version has three problems: a generic opening, repeating the same benefit, and abstract words like “efficiency.” Rewrite the text. Start with an observation from a salesperson’s work, support each benefit with an example, and replace abstract terms with visible results.

Even better, ask the model to analyze before revising:

Point out five fragments that sound formulaic. Briefly explain why each one. Then propose a new version of the text while keeping all facts.

That way, you don’t just get a correction — you also see the mechanism behind weaker style.

How not to overdo the “human” style

Trying to escape AI language can lead to the opposite extreme: broken sentences, too many parentheses, forced anecdotes, and phrases like “Alright, listen up.” Natural text doesn’t have to imitate a coffee-break conversation.

Match the style to the situation:

• a complaint email should be specific and calm, not buddy-buddy,
• a B2B offer can be simple but still precise,
• an expert post doesn’t need a joke in every paragraph,
• a product description should show use, not show off the author’s personality,
• a sales message doesn’t become natural just because it starts with “Hey.”

So the best question is not “Does this sound human?” but: “Does this sound credible for this author, audience, and situation?”

Learn to design prompts instead of guessing

Style control is part of a broader skill: formulating prompts that produce predictable results and evaluating responses consciously. It’s especially useful for marketing and sales professionals, small business owners, and anyone who creates emails, offers, posts, or team instructions every day.

If you want to practice this process on concrete tasks, check out the course Prompt Engineering – The Art of Talking to ChatGPT. It’s a workshop designed for non-technical users. It shows how to build better prompts, assess answer quality, and improve results step by step instead of typing yet another version of “try again.”

For this professional group, it has very practical value: even a small improvement in prompts shortens the editing time for dozens of repetitive texts. The gain is not just a nicer text. It’s also fewer revisions, a more consistent brand voice, and greater control over what ultimately reaches the client.

Quick checklist before generating text

Before pressing Enter, check:

• Did you define a specific audience, not just a broad group?
• Did you state the text’s goal?
• Did you describe the tone through behavior, not just one adjective?
• Does the model know the facts and not have to invent them?
• Did you show an example or name the features of good style?
• Did you indicate phrases and constructions you want to avoid?
• Did you set measurable criteria: length, structure, number of examples, call to action?

If the result still sounds artificial, don’t immediately ask for “more human.” Point out which sentences are too general, what repeats, and what information is missing. ChatGPT responds much better to a diagnosis than to a five-word expression of dissatisfaction.

“Write more humanly” expresses intent, but it doesn’t give direction. The audience, tone, example, banned phrases, and quality criteria provide that direction. And then the text not only feels less like a generator’s answer. It starts to sound like a message from someone who knows who they’re talking to and why.