A draft from a language model is usually correct, orderly and slightly exhausting to read. Readers rarely say why. They notice that every paragraph opens with a connective, that the same handful of grand words keeps turning up, that a clause is wedged between dashes every other sentence, and that the text commits to nothing because every claim is "important to note". Once a reader decides a page was generated, they start skimming, and they trust it less.
The good news is that most of those habits are concrete and countable. This guide names them, shows what fixing each one looks like with real before-and-after text, explains where automatic rewriting goes wrong, and covers what a humanizer can and cannot promise. The examples use the AI Text Humanizer, a rule-based rewriter that lists every change it makes. It is one of the AI writing tools on Toolz.
TL;DR: AI-generated prose reads as machine-written mainly because of a few repeatable habits: inflated vocabulary, hedge phrases like "it is important to note that", sentence openers like "Moreover,", heavy em dash use, typographic quotes, and the "not only... but also" construction. Cutting those makes writing plainer whoever wrote it. The AI Text Humanizer applies six switchable rules client-side, shows each change in a table, and is free, no signup. It does not make text undetectable, and no tool can honestly promise that.
What does an AI text humanizer do?
An AI text humanizer rewrites the specific wording habits that make prose read as machine-generated. There are two very different kinds of product sold under that name.
The first kind sends your text to another language model and asks it to paraphrase. The output changes nearly every sentence, often loses your voice, and cannot tell you what it changed or why.
The second kind, which is what the Toolz humanizer is, applies a fixed and visible list of edits. It finds known patterns, replaces or removes them, and leaves everything else exactly as you wrote it. Every edit shows up in a table with the rule, the text found, what it became and how many times it happened. You can switch any rule off.
The second approach does less, and that is the point. A short list of edits you can audit is a first-pass editor. A full paraphrase you cannot audit is a gamble on your meaning.
Which patterns make writing read as AI-generated?
Six habits account for most of what readers pick up on. Each one maps to a rule in the tool.
| Pattern | Example | What the rule does |
|---|---|---|
| Inflated vocabulary | "leverage", "utilize", "robust", "a myriad of" | Swaps about 50 words and phrases for plain ones |
| Hedge phrases | "it is important to note that", "when it comes to" | Removes the phrase outright |
| Opening filler | "Moreover,", "Furthermore,", "Indeed," at a sentence start | Removes the opener |
| Em dashes | a clause set off with dashes | Replaces each dash, and its spaces, with a comma and a space |
| Curly quotes and ellipses | typographic quotes, the single-character ellipsis | Converts to straight quotes and three periods |
| "Not only... but also" | "not only fast but also cheap" | Rewrites to "fast and cheap" |
None of these is wrong on its own. Plenty of careful human writers use a dash or open a paragraph with "Furthermore". The signal is density: the same few habits, over and over, in a text that otherwise has no fingerprints.
Inflated vocabulary
Models reach for a small set of formal words far more often than people do. The rule swaps them for plain equivalents chosen for meaning rather than from a thesaurus: utilize becomes "use", facilitate becomes "help", commence becomes "start", demonstrate becomes "show", approximately becomes "about", numerous becomes "many". A few words people now associate strongly with generated text, such as delve, tapestry and realm, become "dig", "mix" and "area".
Multi-word phrases run first. Without that ordering, "a myriad of options" would become the ungrammatical "a many of options", because the article belongs to the phrase. With it, the phrase becomes "many options". Capitalization is preserved, so a sentence that began "Robust" begins "Solid".
Hedge phrases
A hedge phrase adds length and removes commitment. "It is important to note that the API is rate limited" says exactly what "The API is rate limited" says, in seven more words. The rule removes fourteen phrases of this kind, including "it is worth noting that", "it should be noted that", "needless to say,", "at the end of the day," and "in today's fast-paced world,". After a removal, the tool tidies doubled spaces and capitalizes the word that now starts the sentence.
Opening filler
"Moreover", "Furthermore", "Additionally", "Notably", "Importantly" and "Indeed" at the start of a sentence are connective throat-clearing. The rule removes them only at a sentence start, so "notably absent" in the middle of a sentence is left alone.
Em dashes
People typing on a standard keyboard rarely produce an em dash, because it takes a key combination most never learn. Models produce them freely. The rule replaces each em dash or en dash, together with any spaces around it, with a comma and a space. That is a blunt fix, and the section on mistakes below covers where it goes wrong.
Curly quotes and ellipses
Typographic (curly) quotes and the single ellipsis character are what publishing software sets, not what a person types into a text box. Converting them to straight quotes and three periods makes pasted text look typed. This rule matters less for meaning and more for text headed into code, CSV files or plain-text systems.
"Not only... but also"
The correlative pair is a construction models overuse. "The tool is not only fast but also cheap" becomes "The tool is fast and cheap". The same rule rewrites the "it's not just X, it's Y" pattern into a plain statement of Y.
A worked example, step by step
Here is a short draft with five of the habits in it:
Additionally, our team will utilize a robust process — it is worth noting that results vary.
With every rule switched on, the rewritten pane reads:
Our team will use a solid process, results vary.
The change table lists five edits: the hedge "it is worth noting that" removed, the opener "Additionally," removed, "utilize" to "use", "robust" to "solid", and the dash to a comma. Each has a count of 1.
Now reread the result. "Our team will use a solid process, results vary" is a comma splice: two independent clauses joined by a comma. The dash rule did what it says and produced a sentence a copy editor would flag. The right human fix is a full stop: "Our team will use a solid process. Results vary." This is the reason the tool lists its changes and why the last step of using any humanizer is to read the output.
A second, simpler example from the tool's own documentation shows how rules interact:
Moreover, it is important to note that we must leverage a myriad of tools.
becomes
We must use many tools.
Four changes: the hedge removed, "Moreover," removed, "a myriad of" to "many", and "leverage" to "use". Untick Inflated vocabulary and only the two removals remain: "We must leverage a myriad of tools."
Where does automatic humanizing go wrong?
A word list cannot read context. Knowing the failure modes lets you catch them in seconds.
Terms of art get flattened. "Landscape" is on the vocabulary list because generated business prose overuses it, so "landscape photo" becomes "field photo" and "landscape orientation" becomes "field orientation". "foster care" becomes "encourage care". If your text uses a listed word in its literal or technical sense, untick Inflated vocabulary or fix the few instances after copying.
Number and date ranges lose their dash. The dash rule also matches en dashes, so 2019–2020 becomes 2019, 2020 and pages 10–12 becomes pages 10, 12. If your text has ranges, switch the rule off or restore them by hand.
Comma splices appear. As in the worked example, a dash that separated two full clauses becomes a comma. Replace it with a full stop or a semicolon.
Quotations get edited. A direct quote that contains "utilise" or "Furthermore," is changed like any other text. Remove quotes from the input or check them in the change table.
Plurals and forms not on the list are untouched. The rule matches whole words, so "realm" is replaced but "realms" is not, and "leverage" is replaced but "leveraged" is not. The change table shows exactly what was matched, so nothing is silently half-done.
Sentence rhythm is not changed. Generated drafts tend toward sentences of similar length. The humanizer does not split, merge or reorder sentences. Varying sentence length is still your job, and the AI Writing Checker measures it for you.
Will a humanizer make text undetectable by AI detectors?
No, and no tool that runs in your browser can honestly promise it. Pages advertising "100% undetectable" output are selling a result they cannot verify.
There are two reasons. First, AI detectors do not reliably establish authorship in the first place. OpenAI withdrew its own AI text classifier in July 2023, citing its low rate of accuracy, as noted in an update to its original announcement. Second, detectors misfire on real people. A Stanford study, GPT detectors are biased against non-native English writers (Liang et al., 2023), found that several popular detectors flagged a majority of TOEFL essays written by non-native English speakers as AI-generated, while classifying essays by US eighth-graders far more accurately.
A test built on unreliable signals cannot be reliably passed or failed. So the honest case for a humanizer is not evasion. It is that the edits improve the writing on their own terms: a sentence without "it is important to note that" is better, whoever produced it.
How should you use a humanizer in an editing workflow?
Treat it as a first pass that removes the obvious habits, then edit like a person. A practical sequence:
- Paste the draft. Text from any model works, and so does your own writing. The word count shows in the pane header.
- Read the change table before the output. It tells you which habits the draft leaned on. A draft with twenty vocabulary swaps and no hedges needs different attention from one with the reverse.
- Switch off rules that fight your content. Technical documentation often needs "landscape" or "foster" in their literal sense. Anything with date ranges needs the dash rule off. Code-adjacent text may want curly quotes converted but nothing else.
- Copy the rewrite and reread it aloud. Fix comma splices, vary sentence length, and add the one thing no rule can add: a specific example, number or detail from your own experience.
The rewrite updates as you type or toggle a rule, so you can compare settings instantly. Clear all and Select all switch every rule at once.
Humanizer, paraphraser or manual editing: which should you use?
Each approach suits a different job.
| Approach | What changes | Can you audit it? | Keeps your voice? | Best for |
|---|---|---|---|---|
| Rule-based humanizer | Only listed patterns | Yes, every edit listed | Yes | A fast first pass on any draft |
| AI paraphraser | Nearly every sentence | No | Often not | Rewording text you do not need to keep |
| Manual editing | Whatever you choose | Yes | Yes | Final drafts, anything published under your name |
| Style checker (grammar tools) | Grammar and clarity suggestions | Yes, one at a time | Mostly | Catching errors a pattern list ignores |
The combinations work well. Run the humanizer to strip the habits, then edit by hand for structure and specifics. If you want to see exactly what changed between the original and your final version, paste both into a text diff checker, which highlights every insertion and deletion.
Is the text uploaded anywhere?
No. The rewrite runs in JavaScript in your browser. Your text is not sent to a server, logged or stored, and the page keeps working with the network disconnected once it has loaded. When you press Copy, the page checks the daily usage allowance and records only the number of changes made, never the text.
That makes it suitable for unpublished drafts, client work under an NDA, and internal documents. The rules and word lists are fixed in the page, so the same input always gives the same output, which also makes results reproducible for a team style guide.
What else signals machine-written text that no rule fixes?
The patterns above are the surface. Some deeper signals need a human editor.
Uniform sentence length. People mix a three-word sentence with a thirty-word one. Generated text settles into a middle length. Break one long sentence in two and join two short ones.
Generic examples. "For example, a business might improve efficiency" is an example of nothing. Replace it with a real number, a named tool, or a specific situation.
Balanced lists everywhere. Three bullet points, each a bolded label, a colon and one sentence, repeated down the page. Write one list as prose, turn another into a table.
No position. Generated drafts present both sides and conclude that it depends. If you have a recommendation, state it early and defend it.
Summaries that repeat the intro. A closing "In conclusion" paragraph that restates the opening adds nothing. End on the most useful point instead.
The same editing principles apply to other kinds of machine output. If you are cleaning up text copied from Claude, the guide on Claude watermarks and hidden characters covers invisible characters that pasted text can carry. And if you want numbers before you edit, the AI writing pattern checker reports sentence-length variation, vocabulary density, hedging and dash rate for any text, with the matches that produced each figure.
When the draft is ready, run it through the humanizer once more as a final sweep. If the green "No formulaic patterns matched" message appears, the listed habits are gone, and what remains is up to your judgment.
