Grok writes in a deliberately looser register than most assistants: more asides, more contractions, the occasional joke. That makes it an unusual input for a Grok humanizer, because the obvious AI tells that other models leave behind are already partly absent.
It does not follow that the output reads as human. Casual prose has its own tells, and they are the ones detectors are least often described as catching. This page is about what actually remains to be fixed, which is narrower and more specific than the general advice.
Which changes what is left to do, rather than removing the need to do it.
The standard AI tells are a narrow band of sentence lengths, heavy connective filler, and vocabulary that is unusually predictable to another model. Grok's register already disturbs the first two to some degree: it varies sentence length more than a default assistant reply, and it uses less of the "it is important to note that" scaffolding.
What it does not change is predictability at the word level. That is a property of how any language model picks the next token, and a conversational tone does not alter it. So a Grok passage can read informally and still sit high on a detector, which is the result people find most confusing.
Forced informality. Contractions and asides placed evenly rather than where a writer would actually have paused. The rhythm is still regular; it is just regular in a chattier way.
Rhetorical questions as transitions. "So what does that mean?" used to move between paragraphs. One is natural. Three in a page is a pattern.
Punchline sentence fragments. A short fragment after a long sentence, for emphasis. Genuinely a human habit, which is why it is used heavily, and it becomes a tell when it appears at a fixed interval.
These are harder to spot by reading than the stock-phrase tells, because they are imitating good writing rather than bad. The practical consequence is that you cannot eyeball Grok output and conclude it is fine.
And the one thing we are not going to put a number on.
A rewrite redistributes sentence lengths, substitutes less predictable word choices, and removes filler. On Grok input the filler-removal has less to do, so most of the movement comes from the other two. We score the passage before and after with our own detector and label the reading as ours, and we highlight which words changed.
What we will not tell you is how much a rewrite moves a Grok passage on average. We have not measured Grok output specifically, and publishing a figure we have not produced is how this category got its reputation. Run your own paragraph and read the two numbers; that is a measurement you can actually trust because you made it.
Nor can anyone tell you what Turnitin, GPTZero or Originality will say about your document. Those services do not expose their verdicts to third-party tools. A vendor advertising a bypass rate against a named checker is quoting a number it chose.
Light varies rhythm and cuts filler and leaves your vocabulary alone, which suits Grok input well, since the vocabulary is usually not the problem. Balanced edits phrasing more broadly. Maximum replaces most of the wording and is a paid mode; the free tier covers Light and Balanced per our published plans. On casual prose, Maximum tends to flatten the voice that made the passage readable, so heavier is not obviously better here.
The limitation most relevant to this particular input, stated plainly.
If you chose Grok partly for its tone, understand the trade before you rewrite: a rewrite optimises statistical properties, and it has no concept of whether a line was funny. Timing depends on word order and on the exact length of the setup, and both are what a rewrite changes. The aside that landed will often come back as a plain clause.
That is not a bug we are going to fix with a setting. It is what the mechanism does. If the humour is the point, rewrite the surrounding exposition and leave the lines you care about alone, which is also why working a paragraph at a time beats pasting a whole piece.
The same applies to anything load-bearing. A rewrite can soften a claim, shift a number, or add a detail that was not there. In testing for these pages one pass returned "Roughly a year into it" when nothing in the source mentioned a timescale. Read the highlighted changes, not the prose. The check is here.
Yes. You can rewrite Grok output here with no account: 3 humanizations a day at up to 260 words each, with the AI score before and after from our own detector and the changed words highlighted. It resets daily rather than being a one-off trial.
It can, and people find that surprising because the tone reads informally. A casual register disturbs two of the three properties detectors read, sentence-length variation and filler density, but it does not change word-level predictability, which is a property of how any language model picks its next token. So Grok text can read chatty and still score high. We have not measured Grok specifically and are not going to publish a rate we did not produce.
Three that a casual register leaves behind. Contractions and asides spread evenly rather than placed where a writer would actually have paused, so the rhythm is regular in a chattier way. Rhetorical questions used as paragraph transitions, which is natural once and a pattern three times. And short punchline fragments after long sentences, appearing at regular intervals. These are harder to spot than stock-phrase tells because they imitate good writing rather than bad.
Partly, yes, and it is worth knowing before you paste. A rewrite changes word order and sentence length, which is exactly what comic timing depends on, and it has no concept of whether a line worked. Asides frequently come back as plain clauses. If the tone is the point, rewrite the exposition around the lines you care about rather than the whole passage, and work a paragraph at a time.
No, and nor can anyone else. Turnitin does not expose its verdict to third-party products, so no external tool can measure its own output against it. The same is true of GPTZero and Originality. We score with our own detector and label the reading as ours. A vendor publishing a pass rate against a checker it cannot query is publishing a number it invented, which is a reason to distrust its other claims too.
Yes, the mechanism is the same regardless of which model produced the text, because it operates on properties of the text rather than on knowledge of the generator. We keep separate pages where a model genuinely changes the advice: Perplexity output carries citations that a rewrite can break, and Notion AI output is block-structured. For Grok the specific thing is the casual register, which is why this page exists rather than redirecting you to the general one.
Citation-bound output, and the claim-source link a rewrite can break.
Read the warning →Rewriting Gemini output, and what moves the score.
See the guide →How to catch a drifted claim or an invented detail.
See the method →Three rewrites a day, 260 words, changed words shown.
Run a paragraph →Three a day, no account. Read the highlighted changes to see what the pass did to your asides.