Your AI-generated text sounds like everyone else’s. That is no accident. Removing AI patterns from texts has by now become an editorial discipline in its own right — and one that makes the difference between being read and being skipped over.
Removing AI patterns from texts is no wizardry. It requires that you know what to look for. The problem: most people who copy GPT output unread into their corporate communication no longer notice it themselves. I have explicitly forbidden my own AI assistant from using certain constructions. I do that not because I reject AI on principle, but because the typical output patterns instantly devalue a professionally solid text. Anyone interested in the style-critical side of the topic will find sharp analyses of patterns of linguistic decline in the long-running language columns of well-known German style critics; I recommend such reading as essential for everyone who signs off texts in companies.
Here are eleven patterns that give anyone away. Each with a before-and-after example. Anyone who wants to tackle removing AI patterns from texts systematically gets a checklist for their own editing pass.
Pattern 1: em-dash antitheses
Before: „Documentation is not a cost block — it is a quality feature.“
After: „Documentation is a quality feature. Anyone who treats it as a cost block has not thought liability law through to the end.“
The em dash as a dramatic pause between two half-sentences is a favourite reflex of language models. It sounds rhetorical but is usually an empty formula. Split the sentence. The second sentence then has to stand on its own, and you notice immediately whether it can.
Pattern 2: „not X, but Y“ constructions
Before: „It’s not about technology, it’s about processes.“
After: „The technology is the smallest problem. Processes decide whether a system introduction works.“
The „not X, but Y“ scheme suggests a surprising twist. Usually there isn’t one. The model has two aspects it wanted to mention and resolves this through contrast. Write two sentences instead, each standing on its own. That also forces you to check whether the second thought is really stronger than the first.
Pattern 3: banality pivots
Before: „In a world that turns ever faster, documentation becomes a strategic resource.“
After: „Anyone who treats operating manuals as a box-ticking exercise loses time on complaints and money on conformity checks.“
The opening via „In a world…“ is a filler sentence that says nothing except: I’m starting now. Cut it. The sentence after it is almost always the real first sentence. And if it isn’t, write it anew.
Pattern 4: chains of superlatives
Before: „This revolutionary, groundbreaking technology transforms the way we think about documentation.“
After: „The tool shortens the turnaround for first translations by 30 to 40 %, if the source texts are clean.“
„revolutionary“, „groundbreaking“, „game-changing“, „far-reaching“ — these words appear in AI-generated texts so frequently that they serve as a giveaway. Replace every superlative with a concrete figure or a concrete example. If you have no concrete figure, the superlative was a statement without content.
Pattern 5: lists of three as a rhetorical reflex
Before: „For a successful introduction you need three things: a clear strategy, the right team, and the right technology.“
After: „Whether a system introduction succeeds depends, in my experience, less on the technology than on the willingness to touch the processes.“
AI models have a pronounced fondness for groups of three. Three points, three pillars, three steps — whether there are three or seven or two. Check with every list: are there really exactly these three? Or have you cut an arbitrary group of three out of a larger context? If the latter, write a sentence that contains the essence. Anyone wanting anti-AI cleanup often starts right here; the seemingly clean lists of three are the most common inconspicuous hits.
Pattern 6: filler formulas
Before: „It is important to understand that quality assurance is a continuous process.“
After: „Quality assurance is a continuous process.“
„It is important to understand that…“ and „It is worth noting that…“ say exactly the same as the sentence that follows, only with four words of preamble that have no information value. Search for these constructions with find-and-replace. They occur more often than you think. That is the simplest way to learn this cleanup: a single find-and-replace list can make a text ten to 20 % shorter without losing a single piece of information.
Pattern 7: essay-school
Before: „In this article you will learn how to optimise your documentation processes with AI.“
After: Simply leave it out. Start with the first relevant sentence.
This formula comes from school. It makes sense there, because pupils learn to announce their texts. In a specialist article it is a sign that the text has no strong opening. If the first sentence is good, it needs no announcement.
Pattern 8: abstract nominalisations
Before: „The carrying out of the implementation takes place in close coordination with the relevant stakeholders.“
After: „We implement the system together with the affected departments.“
Nominalisations often arise from turning verbs into nouns: „implement“ becomes „the implementation“, „coordinate“ becomes „in coordination“. That makes sentences longer and harder to read. With every noun ending in „-tion“, „-ment“, „-ity“, or „-ness“, ask: is there a verb behind it that I can use directly? Usually yes.
Pattern 9: „… and …“ as a fashionable coupling
Before: „The tool supports the creation and the management and the delivery of documentation.“
After: „The tool runs in creation, in management, and in export. Whether your process needs all of that depends on where your bottleneck sits.“
„… and …“ is an enumeration that feels more important than it is. It couples elements that usually need no particular coupling. Replace the construction with a direct enumeration or — better still — with a sentence that judges which element is more important in your context.
Pattern 10: conclusions that summarise instead of thinking further
Before: „In summary, it can be said that AI offers great potential in technical writing, but also requires careful preparation.“
After: „When the foundations are right, AI measurably shortens the turnaround for translations and consistency checks. When they are not right, AI accelerates the production of poor documentation. That is not a balanced statement, that is the order I observe in practice.“
A conclusion that summarises the text is a conclusion that says nothing new. It doubles the text without any gain in insight. A good conclusion takes a thought further, draws a consequence, or poses a question for the reader to take away. That is more demanding as a craft of writing, which is why AI does not do it.
Pattern 11: the scripted comment appeal
Before: „And now over to you: which of these patterns annoys you the most? I look forward to your thoughts in the comments!“
After: Delete it. Or replace it with a concrete, answerable question that refers to a specific point in the text.
The comment appeal at the end of a text is the most obvious marker of automated content production. „What genuinely interests me“, „looking forward to your thoughts“, „let’s continue the exchange“ — three phrasings that can appear in the same week in three different LinkedIn posts without anyone noticing. The call to action is generic because it never belongs to a particular text. It would work under any post. That is precisely the problem: a CTA that fits everywhere signals that it belongs to no one. Anyone with genuine interest in answers asks a question that fits only this text, and accepts that only a few readers will respond. The majority read and move on. That is the normal state, not a communication problem.
Anti-AI cleanup in three phases
Removing such patterns works in three phases. You can use each of them individually if you first want to try the tidying work on a small section.
Five minutes of find-and-replace: open your text in an editor with a search function. Search for the following strings and decide per hit: delete or rewrite.
- „It is important“
- „It is worth noting“
- „both … and“
- „In this article“
- „not only … but also“
- „revolutionary“, „groundbreaking“, „game-changing“, „far-reaching“
- „In a world“
- „the carrying out“, „the implementation“, „the creation“ (as a subject)
- „And now over to you“, „What genuinely interests me“, „looking forward to your thoughts“
10 minutes of reading aloud and deleting: read the cleaned-up text aloud. Every sentence where you hear that it says nothing goes out. Every enumeration where you notice that one element is superfluous is shortened. That sounds like more work than it is. Reading aloud is the fastest means of judging texts, because the ear notices what the eye reads over. Anyone who wants a first automated diagnosis can use the Wortliga text analysis — it finds nominalisations, long sentences, and filler words, but it does not replace critical reading.
Conclusion: the style pass needs no special software. It needs the willingness to read your own text critically — and to accept that a shorter, more direct text is almost always better than a long one that sounds like something. Anyone with a general interest in the development of language through AI will find well-founded style criticism, including on the influence of generative models on everyday language, at the Society for the German Language.
Anyone who wants to know why dealing with AI-generated content demands particular care precisely in technical writing will find the professional background in A chatbot on the command line and in the article on AI in technical writing.
But if you want to know how to prevent these stylistic blunders directly, and thus reduce time-consuming rework, then get in touch with me.