You can humanize AI text manually, and for anything you will have to defend it is the better option than a tool. Editing it yourself leaves a draft history, which is the only thing that actually proves authorship. A tool leaves one block of replaced text.
The point of this page is to make manual editing aimed rather than vague. Detectors read three specific properties, so there are three things worth editing, and a way to check your work without any software. The four named AI patterns are covered separately; this page is the method around them.
Knowing these is what separates aimed editing from rewriting at random.
Sentence-length variation. Generated prose clusters its sentence lengths in a narrow band. Human writing swings: a long sentence, then four words. This is the single biggest contributor, and the easiest to fix deliberately.
Word-level predictability. Models pick likely next words, so the text is unusually easy for another model to anticipate. This is the property you cannot fix by shuffling wording, because a synonym is often just as predictable. The only real fix is below, and it is the first technique for a reason.
Filler and stock phrasing. Connective furniture and favourite vocabulary. Cheap to generate, dense in AI output, and the part whose removal genuinely improves the writing rather than only changing its statistics.
Everything below targets one of these three. If an editing tip does not, it is probably superstition. These are also exactly the properties an automated humanizer works on, so manual humanizing is the same job done deliberately.
The one technique that fixes predictability, and the one most advice leaves out.
A model cannot predict a detail it has never seen. So the most effective manual edit is not rephrasing anything: it is inserting specifics that came from you. A number you measured. The name of the town. What the supervisor actually said in the meeting. The fact that the experiment failed twice first.
Generated prose is almost always abstract where it should be concrete, because the model is averaging over everything it has read. "Stakeholders expressed concerns about the timeline" is the shape of a thousand sentences. "The site manager said eight weeks was optimistic, and he was right" is one sentence that exists once.
No humanizer can do this for you, because it does not have your notes. This works on all three properties at once. It is unpredictable by construction, it changes the sentence rhythm, and it replaces filler with content. It is also, not coincidentally, what makes writing worth reading. If you only do one thing from this page, do this one.
Go back to your notes, your data, your emails, the thing you actually did. If a paragraph has no specifics available to add, that is worth noticing on its own: it may be a paragraph that is not saying anything, and the right edit is to delete it.
A mechanical procedure, because doing this by feel does not work.
Reading for rhythm is hard because you hear what you meant. So count instead. Write the word count of each sentence in a paragraph in the margin. If you see 19, 22, 17, 21, you have found the problem, and no amount of word-swapping will fix it.
Then do two things. Split one long sentence into a long one and a short one. And join two medium ones with a semicolon or a conjunction so a genuinely long sentence exists. The target is a paragraph where the counts are visibly uneven, including at least one sentence under eight words.
Short sentences are the thing generated prose is worst at producing, because they carry no hedging and no qualification. A four-word sentence is a commitment. That is why they read as human.
The oldest advice in editing and still the best diagnostic. Where you run out of breath, the sentence is too long. Where you sound like you are reciting, the rhythm is too even. Where you would not say it that way, do not write it that way. This catches things no count will.
Most of the editing here is subtraction, which is faster than people expect.
The instinct is to swap a flagged phrase for a better one. Usually the right move is to delete it and close the gap. "It is important to note that the results were inconclusive" is better as "the results were inconclusive". The opener added nothing, which is why removing it costs nothing.
Three things to delete without looking for a replacement: transition openers at paragraph starts, where the paragraph break already does the work; hedging stacks, where "may potentially contribute to" means "may contribute to"; and the summary sentence at the end of a paragraph that restates what the paragraph just said, which models produce constantly and writers rarely need.
Expect to cut 10 to 20 percent of a generated draft with no loss of content, which is more than most humanizer passes remove. If you are not cutting, you are probably rearranging, and rearranging moves the properties much less than cutting does.
Three tests you can run on your own draft, and the honest limits of all of them.
The margin test. This is the check a humanizer never shows you. Count the sentence lengths again. If they are still in a narrow band, you edited vocabulary rather than structure and the main property is untouched.
The specifics test. Go through and mark every concrete detail: a number, a name, a date, a thing that happened. If a paragraph has none, it is still averaged prose however much you reworded it.
The attribution test. Could anyone in your field have written this exact paragraph? If yes, it is not yours yet. If it contains something only you would have thought to say, it is.
You can check a passage against our detector for free, with no account, three times a day at up to 1,000 words. Two honest cautions. Detectors disagree substantially with each other on the same text, and they produce false positives on human writing, especially for people writing in English as a second language. So a reading is one instrument, not a verdict, and the false-positive evidence is worth knowing before you treat any score as truth.
Which is also the argument against optimising hard for any single detector, ours included. The properties above are what the category reads in common. Chasing one tool's quirks produces text tuned to that tool and no better anywhere else.
Aim at the three properties detectors read rather than editing at random. Add specifics only you could know, which is the only real fix for word-level predictability and the strongest single move. Break the sentence rhythm deliberately by counting word lengths in the margin and making them uneven, including at least one sentence under eight words. And delete filler rather than replacing it: transition openers, hedging stacks, and the summary sentence that restates the paragraph. Expect to cut 10 to 20 percent with no loss of content.
For anything you may need to defend, yes, and not because the output is better. Editing by hand leaves a draft history that accumulates over sessions, which is the only thing that genuinely evidences authorship. A tool leaves one block of text that arrived at once, which is the shape you least want your history to have. The trade is time. For low-stakes work a humanizer is a reasonable shortcut; for a dissertation it is not.
Adding concrete detail that came from you: a number you measured, a name, what someone actually said, the fact that something failed first. A model cannot predict a detail it has never encountered, so this fixes predictability in a way that rephrasing by hand or by humanizer cannot, and it changes rhythm and displaces filler at the same time. It is also what makes the writing worth reading, which no amount of synonym-swapping achieves.
Three tests. Count the sentence lengths in the margin and confirm they are visibly uneven rather than clustered in a narrow band. Mark every concrete detail in each paragraph and treat any paragraph with none as still unedited in substance. And ask whether anyone in your field could have written that exact paragraph, because if the answer is yes it is not yours yet. Reading aloud catches what the counts miss: where you run out of breath, where you sound like you are reciting.
No, and this is a good example of why chasing a single detector backfires. Detectors disagree sharply on punctuation signals, and some of the major ones weight heavy dash use as an indicator of machine generation rather than against it. So an edit that improves one reading can worsen another. Edit for the properties the whole category reads in common, which are sentence variation, specificity and filler density, rather than for one tool quirks.
No, and you should not expect it to be reliable. Detectors produce false positives on genuinely human writing, so a passage can be entirely yours and still read high, which is particularly common for writers working in English as a second language and for formal academic registers like abstracts and literature reviews. If your own unedited writing scores high, that is a fact about the detector rather than about you, and editing your voice away to satisfy it is the wrong response.
Named patterns to scan for: adjective stacks, transition openers and more.
See the patterns →Why a high score on your own writing is not evidence against you.
Read the evidence →What a draft history evidences that a score never can.
See the steps →The mechanism, if you want to know what you are doing by hand.
Read the explainer →Three detector scans a day at up to 1,000 words, no account, if you want a number on the draft you edited by hand.