Twelve tells of AI writing, and the edit that fixes each one

We see a lot of drafts. Watching what the Humanizer changes most often — the green words in the result pane — gives a fairly clear picture of what makes prose read as machine-written. It is not one thing. It is a dozen small habits that language models have because they were trained on a great deal of competent, cautious, average writing, and they reproduce its averages.

Here are the twelve we see most, each with why it happens and a specific edit. You can apply these by hand or let the tool take the first pass and then read the green.

1. The throat-clearing opener

Tell: "In today's rapidly changing world, …" / "In the modern era, …" / "Throughout history, …"

Why: The model is easing into the topic the way a school essay does. The sentence contains no information.

Fix: Delete the clause and start with the subject.

Before: In today's rapidly changing world, remote teams need clear written communication.

After: Remote teams need clear written communication.

2. "It is important to note that"

Tell: "It is important to note that", "It is worth mentioning that", "It should be noted that".

Why: A hedge that announces a point instead of making it.

Fix: Cut it. If the point is important, saying it is enough.

Before: It is important to note that the sample was small.

After: The sample was small.

3. Transitions in marching order

Tell: Furthermore. Moreover. Additionally. In addition. Consequently. In conclusion.

Why: Models learned that paragraphs begin with connectives, so they supply one every time, usually from the same short list.

Fix: Remove most of them. Ideas that follow logically do not need a signpost; when they do, prefer a plain one — "but", "so", "then" — or make the connection inside the sentence.

Before: Moreover, the second trial produced similar results.

After: The second trial produced similar results.

4. Nouns doing the work of verbs

Tell: "the utilization of", "the implementation of", "facilitate the improvement of".

Why: Nominalisation sounds formal and models associate formal with correct.

Fix: Find the verb hiding inside the noun and use it.

Before: The utilization of feedback facilitates the improvement of drafts.

After: Feedback improves drafts.

5. The tidy triplet

Tell: "clear, concise, and compelling"; "students, teachers, and administrators"; three examples, always three.

Why: Three is the rhythm of persuasive prose and models over-learn it.

Fix: Use the number the content actually has. Two if there are two. One good example beats three adequate ones.

Before: The tool is fast, accurate, and reliable.

After: The tool is fast, and its results hold up.

6. Sentences all the same length

Tell: Every sentence lands between fifteen and twenty-five words. Nothing short. Nothing long.

Why: The model is minimising risk. Medium-length sentences are rarely wrong.

Fix: Vary deliberately. Follow a long sentence with a short one. Let a genuinely complex idea take forty words if it needs them, then stop.

7. The paragraph that ends by restating itself

Tell: A final sentence beginning "Overall," or "In summary," that repeats the paragraph's first sentence.

Why: Essay structure — topic sentence, support, restatement — applied to every paragraph.

Fix: End on the last new piece of information. If the paragraph needs a conclusion, make it say something the paragraph has earned rather than something it already said.

8. Adjectives that cannot be checked

Tell: "cutting-edge", "robust", "comprehensive", "seamless", "innovative", "powerful".

Why: They fill the slot where a specific would go, and the model has no specific.

Fix: Replace with the fact. If there is no fact, delete the adjective.

Before: Our robust, cutting-edge detector delivers comprehensive results.

After: The detector returns a score for the passage and highlights the sentences that drove it.

9. "Delve", "tapestry", "landscape", "realm"

Tell: A small vocabulary of words that appear far more often in model output than in ordinary writing. "Delve into" is the famous one; "the landscape of", "a rich tapestry", "in the realm of", "navigate the complexities" travel with it.

Why: These were over-represented in the text models learned to prefer, and they now function as fingerprints.

Fix: Use the plain word. Delve becomes look at or examine. Landscape becomes field, or just the thing itself. Navigate the complexities becomes deal with.

10. Hedging everything equally

Tell: "may potentially", "can often", "in many cases", "it could be argued" — attached to claims that are not in doubt as well as to ones that are.

Why: Uniform caution is safe. The model cannot tell which claims deserve it.

Fix: Hedge the uncertain claims and state the certain ones. The contrast is what makes the hedges meaningful.

Before: Regular sleep may potentially improve memory in many cases.

After: Regular sleep improves memory. How much, and for whom, is less settled.

11. The explanation nobody asked for

Tell: A definition of a term the audience already knows. "Email, a form of electronic communication, …"

Why: The model does not know who the reader is, so it writes for everyone.

Fix: Decide who you are writing for and cut what they already know.

12. The balanced non-conclusion

Tell: "Ultimately, both approaches have their advantages and disadvantages, and the best choice depends on individual needs."

Why: Refusing to conclude is never wrong. It is also never useful.

Fix: Say what you think, and for whom. If you genuinely cannot choose, say what would change your mind.

Before: Both tools have strengths, and the right one depends on your needs.

After: Use the detector on anything under a page, where the humanizer has little to work with; use the humanizer on longer drafts and check them afterwards.

What these have in common

Almost every tell is a way of avoiding commitment: to a specific word, a specific fact, a specific reader, a specific opinion. Machine prose is prose with the risk removed. Human prose takes risks — some small, like a five-word sentence, and some large, like a conclusion.

That is also why "humanizing" a draft is not really about fooling anyone. A draft with the twelve tells removed is a better draft. It says more in fewer words, it can be checked, and it sounds like someone was there when it was written. When Kinzo's Humanizer marks a change in green, this list is, in effect, what it is pointing at — and reading the green against these twelve is the fastest editing lesson we know.