How I edit AI-generated copy and why it takes so long

AI-generated copy drafts shift work from writing to editing.

Disclosure: This post was written by me, a human, with an assist by a team of rogue AI agents.

I guess it was inevitable. In my day job as a marketing writer, I'm starting to get more requests to write first drafts with AI. I’m not entirely thrilled.

While I believe AI is most effective as a research and ideation tool, and that’s how I like to use it in my writing process, growth leads often think it’s the fastest (and cheapest) way to build the content they need to get to market.

But using AI to write copy doesn’t always eliminate work. It often just shifts the workload from writing to editing.

I’ve tracked my time on writing projects for a month now and, on average, I only save an hour or so overall by using AI to write copy drafts. Of course, actual time saved (or lost!) can vary a great deal, depending on the subject matter.

For me, an AI-generated draft is most helpful when I’m writing about a topic I know really well. In this case, I can feel confident that I’ll catch any hallucinations quickly.

If I’m not familiar with the subject matter, then I’ll have someone who is review every single technical passage and claim. Depending on the expert’s availability, this can add hours or even days to a project.

And then there’s AI’s writing style.

In my experience, AI writing typically has two modes: bland and cringe. The former is flat and emotionless, and the latter cobbles together every marketing hook and clever transition known to humankind into a single paragraph. Either result is unusable as-is. Someone will have to edit it.

Now, it is possible to significantly improve the quality of AI drafts by providing writing samples and a style guide along with your prompt. AI can also be discouraged from injecting its typical writing patterns. And you could even fine-tune a model to speak (mostly) in your voice.

But even the most carefully crafted AI writing framework with well-designed prompts and a data source will only get you about 80% of the way to your final draft. Human editing will still be required.

Skipping this step is not an option. While audiences may not mind if AI is part of your writing process, especially if you disclose it, they will care if your writing sounds like AI.

🔎 A field guide to AI writing patterns

Most AI-generated copy sounds the same because it's trained on the same massive datasets of online content, which include blogs, marketing sites, press releases, and social media posts. It works by predicting the next most likely word based on patterns it has seen, which means it often overuses commonly used phrasing, popular structures, and clichés.

Without strong human direction or editing, AI tends to default to safe, predictable patterns or, in what it thinks is marketing copy, an overenthusiastic tone full of buzzwords and fake excitement.

Here are some of the more common patterns, with real examples:

Too many em-dashes

AI tends to overuse em-dashes for emphasis. This has been extensively documented on LinkedIn and right here on Substack.

AI output:

“Marketing automation helps teams scale their efforts — saving time and resources — while maintaining personalization — all without adding headcount.”

Human edit:

“Marketing automation helps teams scale their efforts and save resources while maintaining personalization. No additional headcount required.”

Of course, some human writers do use em-dashes and use them well. Emily Dickinson was a fan.

But too many em-dashes can be distracting and make your audience wonder, “Was this written by AI?”

Set-up with punchline

One common approach to transitions often found in marketing copy is a short, dramatic phrase or question followed by a quick resolution or reveal. It's a technique many human writers have used to make copy sound “punchy,” and now it's widely used by AI.

AI outputs:

• “The thing is, data quality matters more than quantity.”

• “Here's the kicker: most companies don't realize..."

• “The problem? Integration complexity.”

• “The solution? A unified platform approach.”

Human edits:

• “Data quality matters more than quantity.”

• “Most companies don't realize...”

• “Integration complexity blocks progress.”

• “A unified platform solves this.”

These constructions can help break up longer copy blocks. But, since they’ve become synonymous with AI outputs, it's better to avoid them.

The false contrast formula

AI tends to use contrasting parallel constructions to create drama when discussing two related concepts. It often follows the pattern, “It's not about X, it's about Y,” or “It's not just X, it's also Y.”

AI outputs:

• “It's not about having more data, it's about having the right data.”

• “It's not just about messaging, it's also about sales.”

Human edits:

• “The right data matters more than simply having a lot of it.”

• “Effective messaging drives sales.”

Overuse of explainer verbs

AI uses words like dive deep, explore, delve, and unpack because they're common in blog posts, explainers, and educational content, the kinds of sources AI is heavily trained on. These verbs signal thoroughness and insight without committing to a specific point of view.

AI output:

“Let's dive deep into why customer retention metrics matter. We'll explore the key indicators, delve into calculation methods, and unpack the strategic implications.”

Human edit:

“Customer retention metrics matter because they predict revenue. Here are the three that actually move the needle.”

My SHARP editing framework

I like to use a structured editing framework when working with AI drafts. By walking through the framework step-by-step every time, I can dramatically reduce (but not eliminate, I’m human!) the probability of missing an erroneous fact or robotic passage.

Ironically, the acronym for this framework was suggested by a team of AI agents I was testing on a copywriting workflow.

Here are the steps:

• Spot hallucinations: Fact-check stats, names, quotes, and anything that you're unsure of.

• Hone your voice: Edit or remove clichés, overused words and constructions, and vague explainer language. Adjust the style for consistency with your voice or your brand.

• Add unique insights: Infuse tone, perspective, and nuance that reflect your viewpoint and cannot be replicated by an LLM.

• Read it aloud: Improve pacing, tighten paragraphs, and ensure logical progression.

• Perform one final review: Take a moment to review your final draft and ask yourself if you honestly feel good about it.

  ---

1. Spot hallucinations

Verify every statistic and quote. Click on links and make sure they go to the right place. Validate any citations AI has provided.

How to verify:

• Find the original source document, not just a citation

• Check the publication date (AI may cite 2019 research as current)

• Read the methodology (survey of 50 people ≠ universal truth)

• Verify quotes word-for-word (AI paraphrases aggressively)

2. Hone your voice

Use the find function in your text editor to systematically identify AI markers:

How to remove the AI-isms:

• Find all em-dashes. Replace most with periods or commas.

• Search for “thing is,” “Here's the,” “The problem?” Delete them all. Start those sentences with the actual point.

• Look for "dive into," "explore," "delve into," "unpack." Replace with specific verbs like analyze, examine, or review. Or just present the information directly.

• Rework the false contrasts. Unless you're correcting a genuine misconception, delete every "It's not X, it's Y" construction.

• Look for high-urgency words like "crucial" and "essential" used more than once.

Once this is done, edit the copy to make it sound more like you (or your brand).

3. Add insights AI can't offer

Make sure your content is adding value above and beyond what an AI model can pull out of its training data. This often means adding data points or observations grounded in recent, original research or real-world experience.

Examples include:

• Contrarian takes based on real experience

• Specific vendor comparisons with pros/cons

• Implementation gotchas you've actually encountered

• Current market dynamics AI's training data missed

• Actual human quotes (with permission)

Read it aloud

Read the edited piece out loud to another person. If you stumble over sentences or feel embarrassed by certain phrases, those parts still need work, whether or not they were written by AI.

Generally speaking, human writing sounds natural when spoken. AI writing can sound either featureless and bland or try-hard and cringe.

Perform one final review

Once you've finished the previous four steps, let the draft sit for at least an hour or two. Work on something else and then come back to it. Read it again with fresh eyes. Ask yourself some hard questions:

• Would I stake my reputation on its accuracy?

• Does it provide value a reader can't get elsewhere?

• Is it free of obvious AI patterns?

• Does it enlighten, entertain, or actually help solve a real problem?

More editing in our future?

The more we rely on AI-generated drafts, the more editing we'll have to do. Period.

Ultimately, whether it makes sense to enlist AI as a writing partner depends on your subject matter and your own strengths. If your topic is highly technical, fast-changing, or brand-sensitive, the editing workload may absolutely outweigh the initial time savings.

And if you're more comfortable writing from scratch than fixing awkward phrasing and fabricated facts, AI might slow you down rather than speed you up.

But if you have a solid process for reviewing and rewriting, AI can be a useful drafting tool…as long as you stay SHARP. 😄