Editing AI drafts properly: eight tells, three passes, forty minutes

There is a version of this problem that everyone tries first: run the draft through a rewriter until the score goes green. It produces worse writing, it stops working when the detector updates, and it optimises for a classifier rather than a reader. Here is the version that works, which is just editing, but editing with a specific list of things to fix.

The list is specific because the failures are specific. Model output is not badly written. It is characteristically written, and the characteristics are learnable. Once you can see them, fixing them takes about the same time as a normal edit and produces something a rewriter cannot: writing that is actually better.

First, an honest statement of the incentive problem

We should get this out of the way because pretending otherwise would undermine everything below.

A writer who uses a model for a first draft and then edits it substantially by hand will often still trip a detector. A writer who runs the same draft through an aggressive rewriting tool frequently will not. Detection currently rewards the less honest behaviour, and researchers have named this directly as an integrity problem rather than a technical one.

The reason to edit properly anyway is that the score is not the actual constraint. Google’s spam policies test value to the reader, not authorship, and explicitly cover “automated transformations like synonymizing, translating, or other obfuscation techniques” where little value is added. Academic publishers have converged on disclosure and human accountability rather than tool bans. Elsevier permits language and readability improvement outright and requires a declaration beyond that. Editors reject writing for being empty, not for being machine-shaped. Every gatekeeper that matters is testing something a rewriter cannot fix.

The eight tells, and what to do about each

1. The hedge stack

Models hedge because hedging is safe. You get sentences carrying three qualifiers where one would do: “This can often potentially lead to improved outcomes in many cases.”

Fix: one hedge per claim, maximum, and only where the uncertainty is real. If you do not know whether something is true, say what you do know instead. “Improves outcomes in the three deployments we measured” is both more honest and more useful than “can often potentially improve outcomes.”

2. The tricolon habit

Three-item lists everywhere: “fast, reliable, and scalable.” Sometimes the third item is real. Usually it is padding generated by the rhythm of the first two.

Fix: for every three-item list, delete the weakest item and see if the sentence lost anything. Most of the time it gains.

3. Abstraction without instance

The dominant failure. Paragraphs that describe a category of thing without ever naming one: “Organisations face significant challenges when implementing these systems.”

Fix: the specificity pass, and it is the single highest-value edit available. Every abstract claim gets a concrete instance: a number, a named example, a date, a case. “Organisations face challenges” becomes “Vanderbilt disabled Turnitin’s AI detector in August 2023 after calculating that a 1% false positive rate across 75,000 annual submissions meant roughly 750 wrongly flagged papers.”

This is also the edit that most reliably moves detector scores, because specificity is genuinely rarer in model output. But that is a side effect. The reason to do it is that the paragraph now contains information.

4. The summary paragraph that repeats the section

Models close sections by restating them. It is a structural habit from training on documents that do this.

Fix: delete every closing paragraph that adds no new claim. Then check whether the section reads better. It almost always does. This alone typically removes 10 to 15% of a draft’s word count with zero information loss.

5. Symmetrical structure

Every section the same length, every list the same shape, every heading the same grammatical form. Real writing is lumpy because some things need more space than others.

Fix: let the important section be twice as long as the others and the unimportant one be two sentences. Structure should reflect what matters, and uniform structure tells the reader nothing does.

6. Terminology drift

Models paraphrase technical terms, and technical terms are fixed collocations that must not be paraphrased. This is exactly the failure that produced the “tortured phrases” catalogued in the scientific literature: “profound learning” for deep learning, “counterfeit consciousness” for artificial intelligence. It now drives real retractions.

Fix: build a term list for the domain and check every instance. In any field with established vocabulary, using the wrong term is a competence signal to every reader who knows the field, and it is the single fastest way to lose a specialist audience.

7. Confident wrongness

The dangerous one. Models produce plausible specifics, statistics and dates and citations and quotes, with exactly the same confidence whether or not they are real.

Fix: verify every number, date, name and citation against a primary source. Not a search summary; the source. Fabricated citations remain the most reliable indicator of unedited AI content in existence, they require no tooling to spot, and they are what editors and reviewers actually catch people on.

This step is not optional and it is not compressible. If you skip it, nothing else in this list matters.

8. No position

Models are trained to be balanced, so drafts present considerations rather than conclusions. The reader finishes informed and unhelped.

Fix: decide what you think and write it. Then keep the strongest counter-argument and answer it honestly. A piece with a defended position and an acknowledged weakness is more credible than one that surveys everything and commits to nothing, and it is the thing a model will not produce for you.

What you should add that no rewriter can

Everything above is subtraction. The value is in what you add, and there are exactly four things a rewriting tool has no access to.

• First-hand observation. What you saw when you ran the thing. What broke. What the customer said. This is the only content that cannot be regenerated from public text, which makes it the only durable differentiator you have.
• Non-obvious causation. Not that A correlates with B, but why, and what would have to be true for it not to. Models are good at the first and weak at the second because the second requires a model of the world rather than of text.
• Disagreement with the consensus, plus reasons. Model output regresses toward the average published opinion. If you think the average published opinion is wrong, that is information, and only you can supply it.
• The relevant negative. What did not work, what you would not recommend, where your own product is weak. This is the most trust-generating content there is and no model will volunteer it about you.

A workflow that actually fits in a working day

Three passes. Roughly forty minutes for a two-thousand word draft once you are used to it.

Pass one, structure, ten minutes. Read only the headings and the first sentence of each paragraph. Does the argument progress? Delete every section that does not advance it. Reorder so the strongest claim comes early rather than in a conclusion nobody reaches. Do not touch prose yet.

Pass two, substance, twenty minutes. The specificity pass. Every abstract claim gets an instance. Every number, date and citation gets verified against a primary source. Every technical term gets checked against the domain list. Add the first-hand material. Take a position.

Pass three, line editing, ten minutes. Hedge stacks. Tricolons. Repeated closing paragraphs. Read it aloud, which catches rhythm problems nothing else does. Cut 10%.

Then, if your context requires it, disclose what you used. That is a one-line statement and it costs you nothing with any audience worth having.

Where a rewriting tool legitimately helps

We build one, so we will be specific about the scope rather than vague about the benefits.

Rewriting tools are genuinely useful for register and register alone: turning a stiff paragraph into something readable, matching a house voice across contributors, tightening a passage that sprawls, adapting one piece of writing for a different audience. Those are real editing tasks with real time savings, and a good tool does them faster than you will.

They cannot verify a fact. They cannot add first-hand observation. They cannot take a position. They cannot restore a technical term they do not know is fixed. And if you run text through them repeatedly to chase a score, they will degrade the meaning. Recursive paraphrasing measurably destroys semantic fidelity while raising the text’s perplexity toward noise.

HumanFlow is built for the first list. If what you need is the second list, no tool is going to do it, including ours, and any vendor telling you otherwise is selling you a score rather than a document.

The point of all this

The editing described here takes forty minutes and produces something a person would choose to read. Running a draft through a rewriter takes ninety seconds and produces something that passes a check.

Only one of those is a durable position. Detectors will keep changing; the requirement that writing be worth someone’s time will not. And the forty-minute version has a property the ninety-second version never will. At the end of it, you know the thing is true, because you checked.