How to Edit an AI Draft Into Copy That Sounds Like You

What to look for

Hedging is language that softens a claim the writer could simply make. Chat models hedge by default, whether or not there is any uncertainty to express (Leap). Part of the cause is in the training: OpenAI's 2022 InstructGPT paper noted the model "can overly hedge," and the researchers suspected labelers rewarded cautious-sounding answers.

Common offenders, with the fix:

A practical density check: about four hedge and transition words per 100 words reads as a strong signal to detectors that count them (Leap). You do not need to game a detector. You need a number you can act on, and four per 100 words is a reasonable ceiling for marketing copy.

Worked example

Before:

It is important to note that remote work offers many benefits. However, it may not be suitable for every team. Ultimately, the right choice depends on various factors.

That is 36 words. After the cut:

Remote work saves our team two hours of commuting a day. It works badly for onboarding, so new hires spend their first month in the office.

That is 27 words and it says something. Note that one hedge survived in spirit ("works badly for onboarding") because it carries a real limit, and real limits are information. The test for any hedge is whether it tells the reader something. "We only tested this on English text" does. "Perspectives may vary" does not (Leap).

Filler to cut on sight

• Corporate politeness: "Let's take a closer look," "It's important to understand."
• Recap paragraphs at the end of sections that restate the heading.
• Symmetrical sections that could be shuffled in any order. If the article still makes sense with the H2s rearranged, the structure is cosmetic (BrieferCopy).
• Unchecked numbers. Every statistic gets a source or gets deleted.

One caution: do not turn this into a synonym hunt. Scrubbing the obvious model words does not fix a model skeleton. Paragraphs still march along the same tidy arc and conclusions still zoom out to a vague bigger picture (Louis-François Bouchard). That is why pass two exists.

Generic paragraphs are fixed by specifics, not adjectives. The rule: every section gets at least one of four things, and you can see it on the page.

1. A number with its source. Not "customers save time" but "customers cut onboarding from 14 days to 4" with a link or an internal note. If you cannot source it, use one of the other three.
2. A named tool or system. "We run this in HubSpot" beats "using a leading platform."
3. A customer situation. One sentence in the customer's own words: who they are, what broke, what they did.
4. A before and after. Two lines of text, code or copy showing the change. This is the strongest of the four because the reader can check it themselves.

Then verify the numbers you just added. Models confidently invent statistics, study names and URLs that lead nowhere (BrieferCopy), and Google's guidance is explicit that generative models predict words rather than retrieve facts (Google Search Central). A specific number with no source is worse than a vague sentence, because it borrows authority it has not earned.

Here is what the pass looks like in practice, on a paragraph about voice guidelines:

Before:

Establishing a strong brand voice is very important for consistency across channels and helps build trust with your audience.

After:

Write the rule as behavior, not adjective. "Confident" becomes "we state the recommendation in the first sentence and never open with 'we believe.'" Then attach two annotated samples: one paragraph that nails it, one that misses, with a note on why. Glean reports that 95% of companies have brand guidelines but only 25 to 30% actively enforce them (Glean).

Count the additions: a concrete rule, a before and after inside the example, one sourced number. Three of the four kinds of specificity in one paragraph.

You can run the whole pass as a quick diagnostic on any draft, human or machine:

For each section of this draft:
1. Name the one concrete thing (number, tool, customer, before/after).
2. Mark it MISSING if you cannot find one.
3. Flag any number without a source.
Return a table: section | concrete thing | sourced? | missing.

Then fill the gaps yourself. The model cannot supply your customer's story, and that is the point.

Rewrite the opening from scratch

Do not edit the first paragraph. Delete it and write a new one in your own words, using one specific detail that already exists in the draft. Editors who work on AI-assisted submissions at volume put structure before wording: fix the outline, strip default intros and recaps, then tighten language (Louis-François Bouchard).

A usable opening for marketing copy does two things in the first two sentences: it names a situation the reader recognizes, and it makes a claim the rest of the piece supports. "Most AI drafts fail in the first paragraph, not the last" is a claim. "In today's landscape, content is more important than ever" is a mood.

Spot the press release voice

Press release sentences share a shape: passive construction, abstract subject, no actor.

Sentence-level fixes

• Count consecutive sentences that start the same way ("This helps you... This allows you... This ensures you..."). Rewrite until the pattern breaks (BrieferCopy).
• Read the paragraph out loud. If it sounds like a mid-level consultant giving a presentation, it needs rewriting.
• Vary sentence length deliberately. Short, then longer, then short. A wall of even sentences is the strongest rhythm signal there is.
• Keep one hedge per paragraph if it is honest. Cutting every qualifier makes claims you cannot defend, and the FTC requires advertising claims to be truthful, not deceptive and evidence-based (FTC).

Rewrite from scratch rather than paraphrasing. Paraphrasing preserves the skeleton, and the skeleton is what makes the draft sound like a machine. Start from the fact you want to land and write toward it.

If you are running three passes on every draft, the bottleneck is your brief. Two inputs fix most of it.

First, write voice rules as behavior. Optimizely's guidance is blunt: AI is not a mind reader, and if you want consistent voice you have to feed it voice guidelines, guardrails and sample copy, with prompts like "here are examples of our brand voice; rewrite this product description in that style" (Optimizely). Vague adjectives like "friendly" do not survive contact with a model.

Second, keep samples short and annotated. Write the rule, then the sample, then why it works. HubSpot's brand voice feature asks you to populate the sample text that represents your brand, which is the same principle in a product form (HubSpot). Jasper offers a similar voice management layer if you would rather buy than build (Jasper).

There is a maintenance trap worth naming. Whatever you approve most recently becomes the reference for the next draft, so a piece you ninety-percent-fixed becomes the new model of your voice, and the ten percent you ignored quietly compounds (Quora discussion). Keep a small archive of pieces you are proud of and point the model at that archive, not at last month's output.

If your team is producing this volume across multiple content types, the natural next step is to have an expert own the editorial standard rather than dictating it in a prompt each time; Quanta Experts (https://quantaintelligence.ai/agents) read your context and hand back finished deliverables.

Editing changes style. It does not change what you are allowed to publish.

• Facts. Verify every number, quote and tool name against a primary source. Keep a link for each one in the draft. If the draft cites a study, open it.
• Claims. Substantiate what you say the product does. The FTC requires advertising claims to be truthful, non-deceptive and evidence-based, and the same standards apply whether a human or a model wrote the sentence (FTC).
• Disclosure. If AI use would influence a purchasing decision, check whether a label is required in your market. The FTC has confirmed that AI-generated content falls within existing consumer-protection statutes, and the EU AI Act's Article 50 imposes transparency duties on synthetic media, with penalties described in the regulation (Dynamis LLP). For sponsored content, the ordinary endorsement rules still apply and rest with the publisher (FTC).
• Search. Google does not ban AI-assisted content. It rewards helpful, original content and may treat mass-generated pages without added value as scaled content abuse (Google Search Central). Editing for specificity is also the SEO move.

A pre-publish checklist you can actually finish:

If you need to run drafts through several models to compare tone, one API key for text, image and embedding models cuts the plumbing: Quanta Inference (https://quantaintelligence.ai/inference).