These rules help to write German texts naturally, factually and reader-oriented. They improve the prose without changing facts or the intended statement. Target medium, specific user instruction, subject conventions and accessibility take precedence over general style heuristics.
The basis is the collections of the Wikipedia communities: Signs of AI-generated content (German) and Signs of AI writing (English). The English page is more detailed; its observations are transferred here to German.
The rules are not prohibitions of individual words, but indications of recurring patterns. Every pattern has a reason. Anyone who understands the reason can decide for themselves in case of doubt. Neither a single characteristic nor an accumulation proves machine origin. Conversely: A text does not become good by avoiding all giveaways; it also needs substance (concrete facts, real examples, verifiable sources).
Naturalness, AI detection and origin
Distinguish three levels: linguistic peculiarities, statistical assessments by an AI detector and technical origin markers. A style check is not an origin test; even an accumulation of linguistic patterns does not prove machine authorship.
- Revise for readers, not for a detector score. Promise neither "not recognizable as AI" nor "watermark-free"; a linguistic revision proves neither.
- Strengthen the personal voice through approved language samples, genuine assessments and specific details from the author. Do not invent experiences and do not build in errors to fake human authorship.
- Remove unintentional formatting and citation artifacts without presenting this as removing a watermark. Preserve functional characters, sources and required disclosures. Do not remove or falsify origin markers to deceive about the editing process.
- Do not treat a label for AI editing as proof of fully machine authorship. Proofreading, translating or format conversion can also produce labeled outputs. Conversely, a missing label does not prove human origin.
- For specific questions about labels, check the provider's current documentation. Do not transfer statements about Claude unchecked to other models or products. Do not derive a blanket labeling obligation for every contribution from provider obligations; authoritative obligations must be checked separately.
Classification of the Claude labeling
According to Anthropic's documentation (checked on September 17, 2026), Claude uses two methods: embedded text watermarks for supported models and signed Content Credentials (C2PA) for supported files. The text marking occurs at the model level, also affects API and Claude Code, and can survive copying as well as some edits. C2PA describes the origin or processing of a file, not automatically the authorship of all content.
The documentation does not disclose the technical implementation of the text watermark. Therefore, do not claim that it consists of specific words or invisible Unicode characters. Do not promise that rewording, plain-text export, or a skill will reliably remove it. Model coverage and detection access can change; check again before making specific statements.
Completion criterion: The text is factually correct, natural, and appropriate for the medium; statements about origin are substantiated, and no undetectability is claimed.
Basic attitude
Write the way an informed, restrained person writes: clear, concrete, without staging yourself or the topic. Say what is—not how significant it is. If a sentence adds nothing but solemnity, delete it.
A good self-test: Could a sentence also be placed in a promotional brochure or a scientific abstract? Then it probably sounds too pompous for a factual German text.
Evidence and facts
Never invent sources, titles, authors, years, links, ISBNs, or DOIs. This is the most important point. It is typical of AI to produce plausible-sounding but nonexistent evidence.
Cite only sources whose existence can be verified, and use them to support only statements that actually appear in them. If a detail is uncertain, mark it as uncertain or omit it—better a gap than invented certainty.
A second, more subtle pattern: The evidence exists but does not support the statement. This includes a valid DOI that leads to a completely different paper, a real book on a related but unsuitable topic, a book citation without a page number, a link to a search results page instead of the document, and a citation that appears in the bibliography but is not attached to any statement. Therefore, check not only whether a source exists but whether it says exactly what you attribute to it.
Treat an unsupported claim as a possible error, not as a fact. When in doubt, formulate cautiously rather than confidently incorrectly.
Language and tone
Avoid inflated significance. The topic does not need to constantly justify its importance. Avoid phrases such as “plays a significant role,” “underscores the importance,” “stands as a testament to,” “is regarded as a turning point,” “leaves a lasting impression,” “deeply rooted,” “unshakable.” Instead, state what specifically happened or applies.
Avoid advertising language. Phrases such as “rich cultural heritage,” “breathtaking,” “impressive beauty,” “lasting legacy,” “in the heart of,” “a must-see” belong in brochures, not in factual texts. Remain neutral, even with topics that are easy to charge emotionally (places, culture, people, species).
Avoid editorial comments in which the text evaluates itself: "it is important to note," "it is remarkable that," "it is worth mentioning," "no account would be complete without." If something is important, that shows in the content, not in the announcement.
Do not enter an opinion or interpretation as fact. If an assessment is necessary, write who holds it.
Do not substantiate notability through assertion. AI texts like to demonstrate that a topic is important by making media coverage itself the subject: "was discussed in national media," "received attention in the trade press," "maintains an active presence on social media." Listing the type of source ("independent reporting," "regional media") is part of this too. Write what happened and put the evidence in as evidence—not as an argument in the running text.
Avoid vague authorities and weasel wording: "observers say," "critics argue," "industry reports show," "it is often said," "studies show." State specifically who says something, or omit the claim. Do not transfer the view of a single source to a large, indefinite group.
Avoid typical AI buzzwords when a plainer word says the same thing: "delve into," "illuminate," "vibrant," "multifaceted," "seamless," "robust," "crucial"/"central" as constant emphasis, "exciting" as a filler word, "landscape" as a metaphor ("the digital landscape"). Likewise pseudo-scientific inflation ("causal," "empirical," "correlated") without real data behind it.
Sentence structure and phrasing
Use em dashes sparingly. Clusters of em dashes—especially where a comma, colon, or parenthesis would be more natural—are now regarded as the best-known AI signal of all. A deliberately placed em dash is fine; several per paragraph are a warning sign. For ranges (numbers, dates, routes), use the en dash (–), not the hyphen: "2020–2024," not "2020-2024."
Use connecting words sparingly and naturally. "Moreover," "in addition," "furthermore," "also," "on the other hand" often string together mechanically in AI texts. An occasional conjunction is good; every paragraph that begins with one is a warning sign.
Avoid formulaic negative parallelism: "not only …, but also …," "it is not just about …, but," "it is not a …, but a …"—and the reverse form "X instead of Y" as a constant figure. Such constructions come across as argumentative and promotional rather than factual.
Do not shy away from simple copula sentences. AI texts systematically avoid "is/are/has": "serves as," "functions as," "represents," "marks," "offers," "features," "denotes" (in definitions). "The building is the exhibition hall" is better than "The building functions as an exhibition hall." The same applies to "there is," "it has," and other simple constructions that AI texts steer clear of.
Use the plain verb, not the stiff synonym. AI systematically reaches for the elevated variant: "verfasste" instead of "schrieb", "verstarb" instead of "starb", "siedelte über" instead of "zog um", "bediente sich" instead of "nutzte", "erachtete" instead of "hielt für", "leistete Unterstützung" instead of "half". Conversely, the plain word is a mark of human texts.
Do not force synonym rotation. AI avoids word repetitions at any cost and reaches for ever-new substitute words ("die Stadt … die Metropole … der urbane Raum"). A natural text may repeat the same word if it is the right one.
Avoid the rhetorical triadic pattern (tricolon): three adjectives or short phrases lined up for emphasis ("schnell, zuverlässig und elegant"), "sowohl … als auch … und". A factual German text manages with little rhetoric.
Avoid piled-up participial constructions as appended interpretations: "…, wodurch die Bedeutung unterstrichen wird" ("..., thereby underscoring its significance"), "…, was seine Rolle hervorhebt" ("..., which highlights its role"), "gewährleistend" ("ensuring"), "widerspiegelnd" ("reflecting"). These phrases often tack on an empty evaluation. Better to write a sentence of your own or leave the interpretation out.
Avoid spurious ranges with "von … bis" when they merely string examples together and sound encyclopedically promotional in the process.
Do not strike out what makes human texts what they are. AI avoids on its own some constructions that are entirely normal in German usage texts: hedging words ("vielleicht", "eher", "ziemlich", "sehr", "meist"), clear superlatives and unambiguous statements ("der erste", "der einzige", "einer der größten") and the somewhat cumbersome but common phrasing ("um … zu", "infolge", "die Tatsache, dass"). If something like that is true and fits, leave it standing. A text does not become more human by smoothing it down.
Structure
Be concise. AI texts are often too long and too heavily structured. Not every thought needs its own heading, and not every text needs headings. For short replies, running text is usually enough.
Do without a concluding "Fazit" section and without summarizing paragraphs at the end that merely repeat the text. That is the style of academic papers, not factual usage texts. A summary belongs—if at all—at the beginning.
Do without the schematic "Herausforderungen und Ausblick" ("Challenges and outlook") component ("Trotz seiner Erfolge steht … vor mehreren Herausforderungen" — "Despite its successes, ... faces several challenges"; "Zukunftsaussichten" — "Future prospects"; "Vermächtnis" — "Legacy"). This structure is a strong AI giveaway. Mention concrete problems where they belong in terms of content, without a formula.
Avoid inappropriate, often two-part marketing subheadings. Headings should name, not advertise.
Formatting
Use bold sparingly. Entire paragraphs with highlighted buzzwords are a typical AI pattern. Highlight at most individual, truly central terms.
Use bullet lists only if the content really is a list. Flowing connections belong in sentences. If you do use a list, use the target system's formatting instead of copied characters such as •, - or – at the start of a line.
Avoid the pattern "bullet point + bold keyword + colon + explanation" ("Flexibility: The system allows ..."). These inline-header lists are one of the most conspicuous AI formatting patterns. Likewise: no mini-tables for content that fits in one sentence.
Follow the heading hierarchy and do not skip levels.
Write headings in German, not following an English pattern. "Die Wichtigsten Vorteile Im Überblick" ("The Most Important Advantages at a Glance") betrays a title capitalization borrowed from English. In German, headings are normally capitalized and lowercased as in running text.
Do not put emojis before headings or bullet points, and generally no emojis in factual texts unless explicitly desired.
Do not use decorative horizontal rules between sections.
Technical artifacts
Do not leave Markdown remnants in environments that do not render Markdown: **fett**, ###headings, backticks or copied list markers look like foreign bodies there.
Remove chatbot citation remnants without a trace: URLs with utm_source=chatgpt.com, fragments like "oaicite", ":contentReference", "turn0search" (ChatGPT), "[cite: 1]", "[span_1]" (Gemini) or "grok_render_citation_card" (Grok) are clear giveaways. Check borrowed links to ensure they exist and lead to the stated content.
Use quotation marks consistently – in German "like this" – and do not mix straight and typographic characters in the same text.
Do not deliver broken sentences or abruptly ending texts; this happens with AI outputs that stop mid-generation.
No dialogue or meta remnants
Write the finished text, not the conversation about it. No "I hope this helps", "Happy to!", "Of course!", "Good question!", "Here is the article", "Would you like me to ...", no forms of address and salutations in running texts or discussion posts.
No letter formulas in texts that are not letters: "I hope this message finds you well" does not belong in articles or posts.
No self-description of one's own draft: "This text maintains a neutral tone", "The structure follows the guidelines" – such meta sentences are borrowed chatbot comments, not content.
No references to one's own knowledge state or training cutoff: "as of my last update", "up to my knowledge cutoff", "as far as can be seen from the available sources". If something is uncertain, say it factually about the matter, not as a confession about one's own model.
Do not leave placeholder text standing ("[insert name]", "add details here").
What is not a giveaway
Just as important as the patterns above is which characteristics prove nothing. Anyone who tries to avoid them makes the text worse without making it more human:
- Flawless spelling and grammar. Many people write cleanly; inserting errors does not make a text more authentic.
- A formal, academic or sober tone. Only certain words are typical of AI, not elevated language as such.
- A single connecting word. Only the mechanical accumulation is conspicuous.
- The shift between an objective and a casual register in one text.
- Missing source citations. This is a content-related problem, but not an indication of machine origin.
And the converse does not apply anyway: a text without these patterns can still come from a machine, a text with some of them from a human.
Quick self-check before sending
Before output, briefly check the text against these questions:
- Are all cited sources and facts real and verifiable – and does each source really support the statement it is attached to?
- Is there any promotional or grandiloquent language anywhere that can be cut?
- Is there a "Conclusion" or "Challenges" component or a closing repetition?
- Do paragraphs begin mechanically with "Moreover," "Additionally," etc.?
- Are dashes piling up where a comma or colon would suffice?
- Does the text avoid plain "is/has" ("serves as," "functions as"), reach for stiff synonyms ("authored," "passed away"), or compulsively rotate synonyms?
- Does the text claim importance via media presence instead of content?
- Is there too much bold, too much structure, too much list?
- Are there technical artifacts in the text (Markdown remnants, chatbot quotation remnants, mixed quotation marks)?
- Is there a dialogue or meta remnant in the text?
- Is it asserted without evidence that the text is undetectable as AI or free of watermarks?
- Is an origin marker confused with proof of complete AI authorship?
If an answer is "yes" (or "no" for the first question): revise. But revise selectively – not every formal sentence and not every "perhaps" is a mistake, see "What Is Not a Tell".