People notice AI-written text before they can say why. Usually it's not one word but a handful of habits stacked together: an opener about "today's digital landscape", a source nobody can check, three adjectives where one fact should be, and a closing line that could end any article. Each habit is easy to fix once you can name it. That's what this guide is for.
The signs are ordered roughly by how strongly they read as AI. The first few only make sense in a chat window; the last few are style habits that plenty of human writers share.
Assistant leftovers
How to spot it: Lines that only make sense in a chat window: a compliment on the question, an offer to help further, a sign-off aimed at the person who typed the prompt.
Fix: Delete them. Start with the answer and end on your last useful point.
Checked byen.artifact.chat
Model disclaimers
How to spot it: The text talks about its own knowledge: "as of my last update", "as an AI language model", "based on the available sources".
Fix: Check the fact yourself, then state it with its source and date. If you can't check it, cut it.
Checked byen.artifact.disclaimer
Template openers
How to spot it: The first sentence talks about the state of the world instead of your topic. "In today's digital landscape" is the classic; "Let's dive in" and "Without further ado" are its cousins.
Fix: Open with your subject doing something. The first sentence is the one readers judge you on.
Checked byen.structure.signpost, en.hollow.unsourced_everyone
Vague sources
How to spot it: Experts, studies or "many people" agree, but none are named. The claim sounds supported and can't be checked.
Fix: Name the source and date, or drop the claim. Never invent a source to fill the gap.
Checked byen.hollow.vague_source
Fake certainty
How to spot it: "It is clear that", "without a doubt", "needless to say". The sentence declares the point proven instead of proving it.
Fix: Delete the certainty phrase and add the evidence, or soften the claim to what you actually know.
Checked byen.hollow.fake_certainty
The not-just-X frame
How to spot it: "It's not just a tool, it's a shift." A fake contrast that sets up a bigger claim. One per article is a style choice; one per paragraph is a template.
Fix: State the main claim once, with the evidence.
Checked byen.structure.not_just_but
AI vocabulary clusters
How to spot it: Three or four of delve, robust, leverage, seamless, pivotal, landscape in one paragraph. Each word alone is fine. The cluster is the signal.
Fix: Swap each word for the specific thing it stands for. The full list is on the AI words to avoid page. See all AI words to avoid →
Checked byen.lexicon.cliche
Role, testament and journey filler
How to spot it: "Plays a vital role", "serves as a testament to", "embark on a journey", "unlock the full potential". Stock phrases that stand where a verb with a subject should be.
Fix: Say what the thing does, to whom, and with what result.
Checked byen.lexicon.play_role, en.lexicon.journey_unlock
Stacked hedges
How to spot it: Two or more qualifiers on one claim: "could potentially", "may possibly", "it could be argued that".
Fix: Keep one qualifier, or none, and back the claim.
Checked byen.lexicon.hedge_stack
Wordy filler
How to spot it: "In order to", "due to the fact that", "it is important to note that". Phrases that add length and an official tone.
Fix: Cut the phrase or use the short form (to, because). These are safe one-click fixes in the editor.
Checked byen.lexicon.filler_phrase
A transition at the start of every paragraph
How to spot it: Furthermore, Moreover, Additionally, Consequently. Formal connectors that turn paragraphs into an outline.
Fix: Drop the connector, or use the plain word (also, so, but) inside the sentence.
Checked byen.lexicon.transition_stack
Throat-clearing before a list
How to spot it: "Here's what you need to know:", "There are several benefits:". A sentence whose only job is to introduce the next sentence.
Fix: Delete the setup and start with the first real point.
Checked byen.structure.list_intro_empty, en.structure.colon_throat
Em dashes everywhere
How to spot it: Dashes in most sentences, often in pairs, where a comma or period would do. A few are fine; a steady stream reads as a template.
Fix: Keep the dashes doing real work; turn the rest into commas or periods. Check your dashes →
Checked byen.structure.dash_dense
Sentences with the same shape
How to spot it: Neighboring sentences start the same way and run about the same length. "It improves speed. It improves quality. It improves trust."
Fix: Vary the openings. Merge two short sentences or split a long one where the idea turns. Measure your sentence rhythm →
Checked byen.structure.adjacent_parallel
Generic conclusions
How to spot it: The last line could end any article: "In conclusion", "the future looks bright", "only time will tell".
Fix: End with a result, a limit or a next step.
Checked byen.hollow.generic_conclusion, en.structure.in_conclusion
What these signs can't tell you
None of these patterns proves that an AI wrote something. People wrote all of them first. They also show up more in some kinds of human writing:
- Formal and academic prose, which uses transitions like "moreover" on purpose
- People writing in a second language, who often learn from formal templates
- Corporate and legal documents that follow a house style
- Short texts, where one stock phrase can look like a pattern
AI detectors have the same problem at a larger scale: they flag careful human writers and miss lightly edited AI drafts. Use these signs to edit your own writing. Don't use them to accuse someone else.
Questions
How can you tell if something was written by AI?
You can't tell for certain from the text alone. You can spot patterns that are common in chatbot drafts, like the ones on this page, but people write all of them too. Treat them as editing prompts for your own work, not as evidence about someone else's.
Are AI detectors reliable?
Not reliable enough to accuse anyone. They produce false positives, especially on formal writing and on writing by non-native speakers, and their results change with small edits. Naturable doesn't detect AI and doesn't test against detectors.
Which sign matters most?
Assistant leftovers and model disclaimers are the strongest: they only make sense in a chat window. Single words matter least. Most of the time it's a cluster of several signs in one paragraph that makes a draft feel machine-made.