How to Write with AI Without Sounding Like AI: 10 Habits That Give the Machine Away

When a communication professional uses artificial intelligence to prepare a text, three questions appear almost immediately. Will the audience notice? Will they care? And even when nobody notices, is it acceptable to let a machine write on our behalf?

The first question has no universal answer. Some readers recognize AI-assisted writing within a few lines. Others will never think about how the text was produced. The second answer also depends on the context. Discovering that a restaurant used AI to create a decorative image is unlikely to ruin anyone’s dinner. Discovering that an institution used it to answer a sensitive complaint or communicate information affecting people’s rights may provoke a very different reaction.

The third question is the most important because it remains relevant even when the first two disappear. A text can pass as human, contain no mistakes, and still fail to communicate what the organization truly thinks. It can sound professional while saying almost nothing.

A reader may not feel deceived simply because AI was involved. They may, however, disengage when the text feels generic or disconnected from any recognizable human intention. AI usually produces grammatically and conceptually correct texts, but correctness is not the same as excellence or authenticity.

The risk is not simply that people discover the tool. The risk is that the text reveals the moment when the writer stopped thinking.

Before asking AI to write, understand what it is doing

AI is not neutral, omniscient or objective. It has been developed by people and companies, trained on information containing conflicting ideas, and configured to prioritize some forms of knowledge and behavior over others.

A language model does not understand people. It doesn’t think as we think, remember as we remember, connect as we connect, nor reflect as we reflect. It learns patterns in language and produces a statistically plausible response. The fluency of the result can make it appear more knowledgeable or reliable than it really is. But this leads to an important journalistic distinction:

A plausible sentence is not necessarily a true one. A convincing argument is not necessarily a true one.

A model may produce an elegant notice about an interruption to a water supply, but it does not experience the operational pressure on a health centre or the anger of a businessman that cannot open that day. It can imitate an empathetic response because it has learned the linguistic patterns associated with empathy (ELIZA effect). That does not mean it understands the situation behind those words.

Generative AI is remarkably good at producing clean copy. Give it a topic, a tone and a desired length, and it can return a structured text in seconds. The spelling will usually be correct. The transitions will work. The sentences will follow one another with apparent logic.

This creates a dangerous illusion. Linguistic correctness is a minimum standard, not an editorial achievement. Good writing requires a point of view, an understanding of the audience, control over the argument and a reason for every relevant sentence to exist. It requires knowing what must be said and what evidence makes a claim credible.

AI is not a neutral writer

Generative AI does not approach a subject as an independent observer with access to one uncontested body of knowledge. It works with information that contains contradictions, cultural assumptions, political positions, competing definitions and incompatible interpretations of reality.

A brief thought experiment makes this easier to understand. The Bible, the United Nations 2030 Agenda, the Code of Hammurabi, The Communist Manifesto, the Spanish Constitution and the works of thinkers such as Kenneth Grant or Nick Land all circulate online and could, directly or indirectly, form part of the vast textual environment from which an AI system learns. Yet these sources do not describe one coherent vision of society. They disagree on authority, morality, rights, property, freedom, progress, etc… In some cases, they defend plainly opposing ideas. A model cannot treat all those positions as equally valid in every response.

Its behaviour and the selection of the most legitimate ideas is shaped by human decisions: which information is used, how the system is trained, which responses are rewarded, what safety rules are introduced and what the company behind the product considers appropriate. The hierarchy may not always be visible to the user, but it is there. This makes the idea of complete neutrality impossible, which shouldn’t be strictly a problem, but at least we should be aware of that.

The same impossible quest for neutrality applies to human writers, of course. Journalists, public officials and communication professionals also have assumptions, values and blind spots. The difference is that a human author can be questioned about their reasoning and held accountable for a decision. We know what they might believe in, their goals, flaws and what they defend. With AI it remains blurred.

A language model plays the game of imitation

Large language models learn relationships within enormous quantities of text. They identify how words are commonly combined, how arguments tend to be structured, how an official notice usually begins and which expressions often appear in a corporate announcement.

When they respond, they do not retrieve an internal understanding of reality in the way a person recalls an experience. They calculate a plausible continuation based on the patterns they have learned and the context they have received.

That ability is extraordinarily powerful and efficient. But it also creates one of the central risks of generative AI: plausibility can be confused with truth.

A useful way to understand this is through a classical progression of knowledge: ignorance, doubt, opinion and certainty. Ignorance means we do not know. Doubt means we recognize competing possibilities. Opinion appears when we consider one explanation more likely than another. Certainty is the strongest state we can reach as knowers: the conviction that our judgment is correct. But certainty is still not the same as truth. A person can feel entirely certain and still be wrong, because truth exists independently of the confidence with which we describe it.

The same distinction applies to veracity. We often judge a statement as credible because the speaker appears sincere, the reasoning is coherent and the message follows a convincing structure. Yet veracity, understood as the apparent reliability of the relationship between a speaker and a statement, does not guarantee that the statement corresponds fully with reality. A message can be sincere, logically constructed and persuasive while still resting on incomplete information or an incorrect conclusion.

Generative AI complicates this distinction because it can reproduce the language of certainty without experiencing doubt, holding an opinion or knowing the truth. A fluent answer feels trustworthy. A well-structured explanation appears authoritative. A precise quotation looks verified. The model can express hesitation or confidence, but these are linguistic forms, not internal states of knowledge.

The formal quality of an answer therefore does not guarantee the accuracy of its contents. ChatGPT is not a replacement for a search engine, an official source, a legal adviser, a technical specialist or professional judgment. Of course, primary sources can mislead us, experts can make mistakes and our own reasoning is not infallible. But the caution we apply to ourselves and to the people and institutions we trust should not disappear when we speak to AI.

The danger begins when we treat the system as a mystical, all-knowing entity simply because its internal processes are difficult to see. Opacity can create an illusion of wisdom.

The quality of the result begins before the writing

Many disappointing AI texts are blamed on the tool when the real problem began with the instruction.

A request such as “write a post about our new project” leaves almost every meaningful decision to the model and, since the tool has been configured to save us time and seem to understand us better than anyone else, AI will fill those gaps with its most common assumptions. That is how organizations and people with very different identities end up publishing strangely similar content.

A professional prompt should work like a good briefing. It should define the role the model will play, explain the context, establish the objective, provide the available information, specify the format and tone, set clear limits and describe how the result should be evaluated.

The interaction should also remain conversational. The first response is a draft, not a verdict. We can ask the model to explain its choices, identify weak arguments, show what information is missing, propose alternatives and criticize its own version. The best results usually emerge through revision and an ongoing exchange between the writer and the tool.

Ten signs that the machine has taken over the voice

These patterns are not errors in themselves. Human writers use them too, and they are legitimate rhetorical devices.

The problem begins when the same devices appear mechanically in text after text, regardless of the subject, author or audience. They become shortcuts that create the appearance of style without doing the work of thought, gradually weakening the authenticity of the writer’s voice.

The point, then, is not to prohibit these ten resource, but to examine them critically. Are they necessary? Do they add rhythm or meaning? Or are they simply taking up space and making the text sound like countless others because the model has learned that these structures are usually rewarded?

1. Excessive negative contrasts

IT IS NOT X, IT’S Y.

“It is not a cost. It is an investment.”

“It is more than a service. It is an experience.”

“The technology does not think. It processes.”

Generative AI has a particular affection for formulas built around contradiction. But readers usually need to understand what something is before being told what it is not. Negative structures can be useful when they correct a common misconception or draw a necessary distinction: “It is not a search engine; it is a language model.” In that case, the contrast prevents confusion and adds meaning.

The problem is that AI writing tools often use this structure as a default rhetorical move, even when no real misunderstanding exists. The result may sound emphatic without saying anything useful. “The lamp in my bedroom is not a family member; it is a tool” is logically correct, but nobody needed the clarification.

Use negative contrasts when they challenge a genuine assumption, not simply because the formula sounds polished.

2. Everything arrives in groups of three

X, Y, Z

“Clear, agile and efficient.”

“People, purpose and progress.”

“Innovation, sustainability and impact.”

The rule of three has been part of rhetoric for centuries because three elements are easy to remember and can create satisfying rhythm. AI has learned that lesson extremely well. Sometimes too well.

Real ideas do not always organize themselves into identical sets. A project may have two decisive benefits and six relevant risk, an audience may require four different messages. In many cases, one of the three elements adds little or no meaning. It may simply repeat an idea already expressed or act as a synonym for another item in the list.

3. Punctuation that does not belong to the author

—

A punctuation mark can reveal the influence of AI when it appears repeatedly in someone’s writing despite never having formed part of their natural style.

The long dash is a common example in Spanish-language AI outputs. Although it is widely used in English and is perfectly legitimate in literature and certain editorial contexts, it can feel unusual in a routine email or LinkedIn post written in Spanish by someone who would normally use commas, colons or parentheses.

4. Prose begins to look like poetry

X

Y

Z

We had an idea.

It was ambitious.

It was difficult.

But it changed everything.

AI often places each short sentence on a separate line. The layout creates pauses and imitates emotional intensity. It is especially common in social media posts, where empty space can make a text appear easier to read.

Sometimes the technique works. More often, it turns ordinary information into artificial drama. The reader receives a sequence of declarations without enough development to understand why they matter.

Paragraphs exist because ideas have relationships. A sentence introduces a claim, another qualifies it, and a third provides evidence or consequence. Restoring that connection frequently makes an AI draft sound more natural because human thought is rarely a procession of isolated slogans.

5. Decorative lists and predictable emojis

🔷​ X

🔷​ Y

🔷​ Z

AI is skilled at transforming almost anything into a list. Lists are useful when readers need to scan options or follow steps, but become a problem when they are used to decorate information that would be clearer as a paragraph.

Blue diamonds seem to be one of AI’s favourite emojis for lists. They appear repeatedly in AI-assisted social content because they are familiar visual conventions in digital marketing. That familiarity is precisely what makes them feel generic. And, while we are here, be careful with rockets too 🚀.

The issue is not the use of emojis itself, but the automatic use of the same ones regardless of the subject. In many cases, a different symbol would relate far more accurately to the meaning of each item.

6. The machine chooses the emphasis

Generative AI often returns text with selected words in bold. The model is deciding what should attract attention before the author has reviewed the hierarchy of the argument.

Bold type is valuable when it allows a reader to identify a central concept, figure, deadline or action, but too much emphasis has the opposite effect. When everything appears important, nothing is.

Selecting emphasis is a small but revealing act of authorship. It requires deciding what the reader must remember and what merely supports that idea. Even when AI has prepared the draft, the writer should retain control of that decision. The time spent will be very valuable.

Believe us, it is worth spending a few minutes on this. If AI has already saved you time in the drafting process, using part of that time to choose the bold text yourself is a very profitable investment. You will still be working faster.

7. Familiar marketing phrases replace the argument

“Because every X matters.”

“When X, Y…”

“From X to Y, we…”

“Together, we are…”

“And that changes everything.”

“Behind every X, Y.”

“That’s where we…”

This is perhaps the most damaging habit. These phrases are emotionally recognizable and easy to approve because they resemble thousands of messages already circulating online. Their weakness is that they usually make claims without giving the reader a reason to accept them.

Communication becomes credible when it connects conclusions to evidence. A service is reliable because it follows a tested protocol or operates with trained personnel; a project is sustainable because it reduces a measurable impact; A system improves safety because it identifies a defined risk earlier.

We should be the ones building those arguments and feeding them into the AI tool. The model can help organize the facts into a persuasive structure, but it should not be asked to compensate for their absence with enthusiasm, and certainly not to invent them.

This is also where voice input can be especially valuable. Sometimes a carefully written prompt is necessary. In many other cases, speaking allows us to think more naturally: to propose an idea, connect it to another, move from point A to point B, reach point C because of D, and then return to an earlier thought when something no longer fits.

We can think out loud while letting the AI listen. That way, the central reasoning remains ours. The model can then organize, clarify and refine the material, but it is working with a real chain of thought and evidence rather than filling the gaps with familiar, overused phrases.

8. Adjectives add volume instead of meaning

Strategic

Innovative

Comprehensive

Sustainable

Rigorous

Purpose-driven

Clear

None of these words is forbidden. Each can communicate something useful when it is attached to evidence. The problem is that AI frequently accumulates adjectives to give a sentence more weight without making it more precise.

An “rigorous and comprehensive platform” remains an abstraction. A platform that reduces the time required to classify an incident from two hours to fifteen minutes communicates something that can be understood and evaluated.

A useful editing test is to remove or replace adjectives and see what changes. Let’s not use the same 8 adjectives that the AI model gives to every text online.

9. Metaphors arrive before the meaning

‘‘Amid complexity, strategy becomes a compass.’’

‘‘Every decision opens a new horizon.’’

‘‘AI is the next step towards singularity.’’

AI has learned a highly romantic version of storytelling. A good metaphor can make an unfamiliar idea easier to understand, but a weak one can make an already vague idea even less concrete.

Before keeping a metaphor, the writer should ask what explanatory work it is actually doing. Metaphors take a specific situation and connect it to a more familiar reality. That can be useful, but it also narrows the range of possibilities. When AI repeatedly reaches for the same images, different ideas begin to sound identical because of its reliance on a limited set of familiar metaphors that flatten distinctive thoughts into the same predictable language.

10. The language carries traces from somewhere else

A strategic ally for growth
A Strategic Ally for Growth

Un Aliado Estratégico para el Crecimiento
Un aliado estratégico para el crecimiento

AI operates across languages, but editorial conventions do not always transfer correctly. Capitalization is one of the clearest examples.

In English, headings may follow either sentence case or title case, depending on the publication or brand style. In sentence case, only the first word and proper nouns are capitalized. In title case, the principal words begin with capital letters, while shorter articles, conjunctions and prepositions often remain lowercase. Both options can be correct, but they should be applied consistently.

Spanish follows a different convention. In most titles and headings, only the first word and any proper nouns are capitalized. A construction such as Un Aliado Estratégico para el Crecimiento therefore looks as though English title-case rules have been transferred directly into Spanish. The more natural version is Un aliado estratégico para el crecimiento.

The transfer can also happen in the opposite direction. An English organization that normally uses title case may receive AI-generated headings written in sentence case, making the text look inconsistent with the rest of its communications.

The same problem can affect punctuation, sentence rhythm, idioms and levels of formality. A text may be grammatically understandable and still feel culturally imported. The final review should therefore go beyond correcting mistakes and ask whether the language follows the editorial conventions of the language, the organization, the market and the person supposedly speaking.

—

The writer must remain visible in the work

These ten decisions may make AI-assisted writing harder to recognize, but concealment should never be the goal. The goal is to make sure there is still an author: someone who selected the facts, considered the audience, questioned the evidence, removed what could not be verified and accepted responsibility for the final message.

Writing has always been more than a mechanical task. Words can calm, confuse, reassure, mobilize or mislead. . A polished text can still be careless, just as an efficient process can still make a poor ethical decision. Editorial judgment, privacy, context and responsibility matter as much as style.

Of course, the future of writing includes AI. The purpose is neither to continue working as though these tools did not exist nor to use them indiscriminately. Their arrival does not make writers, journalists or communication professionals less necessary. On the contrary, anyone who concludes that AI now makes it possible to remove those roles has misunderstood what the work actually involves, and will eventually face the consequences in weaker judgment and a voice indistinguishable from everyone else’s.

Our work has become more important because differentiation now depends even more on human perspective: the ability to notice what others miss, connect ideas unexpectedly, understand context, take creative risks and speak with a voice shaped by experience. AI can accelerate the process, but it cannot bring the singularity that every person brings to language, even through imperfections.