Key takeaways
- Treat AI output as a rough first draft from a fast writer with no judgement and no stake in the topic.
- Fact-check first. Everything else is wasted effort if a claim in the draft is invented.
- Most AI drafts are 20–40% filler. Cutting is the single highest-value edit you can make.
- The recognisable 'AI voice' comes from a small set of repeated habits — hedged openings, tricolons, hollow transitions — and each has a specific fix.
- A draft needs at least one thing only you could have contributed, or there is no reason for it to exist.
What you are actually starting with
An AI draft is not a bad piece of writing. That is what makes it difficult to edit. It is grammatical, well-organised, on-topic, and almost entirely without a point of view.
The failure mode is not error but blandness — text that covers a subject without ever committing to anything about it. Light copyediting will not fix that, because nothing is wrong at the sentence level. The problems are structural.
What models reliably get wrong:
- Confident specifics that are invented — statistics, dates, citations, quotes, product capabilities.
- No hierarchy. Every point is given roughly equal weight, so the reader cannot tell what matters.
- Padding. Restated premises, throat-clearing openings, summaries of what was just said.
- No stakes. Nothing is argued, so nothing can be disagreed with.
- Recycled phrasing. The same constructions recur because they are statistically likely, not because they fit.
The mindset that helps
Treat the draft as research notes from an assistant who read widely and understood nothing. You are not polishing their work; you are writing your piece using their material.
The editing pass, in order
Order matters here. Every step is cheaper if the ones before it are done first — there is no sense polishing a sentence you are about to delete.
- Verify every checkable claim. Names, numbers, dates, citations, quotes, technical specifics. Models produce fabricated detail in exactly the confident register they use for correct detail, so plausibility tells you nothing. Anything you cannot source, cut or reframe as a general statement.
- Fix the structure before the prose. Read only the headings and topic sentences. Do they form an argument that progresses? AI drafts tend to produce a list of adjacent topics rather than a sequence where each part earns the next. Reorder, merge and delete whole sections at this stage.
- Cut ruthlessly. Delete restated premises, hollow transitions, and any sentence that summarises what the reader just read. Expect to remove 20–40%. If a paragraph can go without leaving a hole, it should.
- Add what only you can add. A specific example from your work, a number from your own data, a caveat you have learned the hard way, a position you will actually defend. This is the step that makes the piece worth publishing.
- Break the rhythm. Model output is metronomic — sentences of similar length, paragraphs of similar shape. Vary deliberately. Follow a long sentence with a short one. Let a paragraph be two lines.
- Then copyedit. Only now. Word choice, transitions, consistency of terminology. Doing this first is how people spend an hour perfecting text they end up deleting.
The specific tells, and how to fix each
The 'AI voice' is not mysterious. It is a handful of habits that recur because they are the statistically safe choice. Once you can name them you will see them everywhere.
| Pattern | What it looks like | The fix |
|---|---|---|
| Hedged opening | 'In today's fast-paced digital landscape…' | Delete the sentence. Start at the first real claim. |
| Rule of three | Every list has exactly three items, every sentence three clauses. | Vary the count. Use two, or five, or one. |
| Hollow transitions | 'Moreover', 'Furthermore', 'It is important to note that' | Cut them. If the logic needs a bridge, write the actual reason. |
| Symmetrical negation | 'It's not just X, it's Y' as a recurring construction. | State Y directly. The contrast is usually doing no work. |
| Abstract nouns | 'leverage', 'utilize', 'facilitate', 'robust', 'seamless' | Replace with the plain verb. 'Use'. 'Help'. 'Strong'. |
| Summary paragraphs | A closing paragraph restating the article. | Delete. End on the last substantive point. |
| False balance | 'While some argue X, others argue Y' with no resolution. | Say which is right, or explain why the question is unsettled. |
Why cutting works so well
Nearly every one of these fixes is a deletion. Model output is padded because padding is low-risk. Removing it is the fastest route to prose that reads like a person wrote it, and it costs no writing skill at all.
Restoring a voice
Cutting gets you to clean. Voice is what gets you to memorable, and it is the part that cannot be automated.
Commit to a position
Models hedge because hedging is safe. The result is text that surveys a topic without ever saying what is true about it.
Find the sentences that describe a disagreement and pick a side. 'Some teams prefer X while others prefer Y' becomes 'X is usually the wrong choice, and here is when it isn't.' A reader can do something with the second version.
Trade abstraction for specifics
'Companies often struggle with onboarding' is true of everything and useful for nothing. 'Our onboarding took eleven days until we removed the manual approval step' is a sentence only you can write.
Every specific you add is something no model could have produced from a prompt, and it is also the reason someone would read your version rather than any other.
Read it aloud
The most reliable test available, and the most underused. Sentences you stumble over are too long. Passages that bore you when spoken will bore a reader silently.
Anything you would not say to a colleague should be rewritten as what you would actually say.
Where tools help, and where they cannot
We make editing tools, so read this section with that in mind. The useful distinction is between mechanical work and judgement.
Mechanical work — flattening robotic phrasing, removing filler, cutting clichés, varying sentence rhythm — is repetitive and rule-shaped, and tooling handles it well. Judgement — deciding what the piece argues, which claim to verify, what to cut, what only you can add — is the actual editing, and no tool does it for you.
Free tools for the mechanical passes:
- Remove fluff — strip padding and filler sentences
- Cliché remover — replace worn phrasing
- Remove robotic tone — flatten stiff, mechanical constructions
- Passive to active voice — tighten sentences that hide their subject
- Sentence rewriter — alternatives for a sentence that will not sit right
For a whole draft, our humanizer rewrites the text with more varied vocabulary and rhythm while keeping your meaning, and returns several versions so you can pick the one closest to how you would have put it. There are no tone or strength controls — you paste text and choose from the alternatives.
Paste a draft to see the alternatives. English only. Free plan covers 500 words per month, 300 per request.
What no tool will do
Verify a fact, decide what your piece argues, or supply the specific detail that makes it yours. If you skip those, you get cleaner filler — which is still filler.
A checklist to run before publishing
- Every number, name, date and citation is verified against a source you have seen.
- The headings alone tell a coherent story.
- The draft is meaningfully shorter than what the model produced.
- There is at least one specific that could only have come from you.
- The piece takes a position a reader could disagree with.
- No paragraph merely restates the previous one.
- Sentence lengths vary noticeably.
- You have read it aloud end to end.
- If your context requires disclosing AI assistance, you have disclosed it.