The irony is not lost on me

Yes, I'm writing an article about how to make AI writing sound human. And yes, I used AI tools in the process. And yes, I'm aware of the recursive absurdity. But here's the thing: most of the advice about AI writing is itself obviously AI-written, which is like getting swimming lessons from someone who's drowning. So let me try to do better.

The problem with AI writing isn't that it's bad. It's that it's recognisably AI. After reading even a few AI-generated articles, your brain starts pattern-matching, and once you can see the patterns, you can't unsee them. Every article starts with a broad statement. Every paragraph transitions with "However," or "That said,". Every list has three items of escalating importance. Every conclusion ties everything together with a neat bow about how exciting the future is.

Humans don't write like that. Humans are messy, opinionated, inconsistent, and occasionally wrong. That's what makes human writing interesting. So here's the workflow I use to take AI output and make it sound like an actual person sat down and wrote it.

Step 1: Research with Perplexity (not ChatGPT)

Before you write anything, you need to know what you're talking about. This is where most AI-assisted writing goes wrong from the start. People ask ChatGPT to write an article about a topic and it produces something that's technically plausible but factually squishy. It confidently states statistics that don't exist, references studies that were never published, and presents opinions as facts.

Use Perplexity for research instead. Perplexity cites its sources. You can click through, verify claims, and build your article on facts rather than plausible-sounding fabrications.

Spend 15 minutes with Perplexity understanding the topic. Note the key facts, the current debates, the things that surprised you. These surprises are important. They'll become the interesting parts of your article, the parts that feel like a human discovered something rather than an AI summarised something.

Step 2: Outline by hand (seriously)

I know this feels counterproductive. The whole point of using AI is to save time, and here I am telling you to do something manually. But the outline is where your point of view lives, and AI can't give you a point of view.

An AI outline will be balanced, thorough, and completely devoid of perspective. A human outline has opinions. "This section argues that X is overrated." "This part challenges the common assumption that Y." "Open with the controversial take that Z."

Your outline doesn't need to be detailed. Five bullet points with your main arguments and the rough flow of the piece. That's enough. What matters is that the structure reflects what you think, not what the average of the internet thinks.

Step 3: First draft with Claude

claude is my recommendation for the drafting step, and I'll tell you why. Claude produces more natural-sounding prose than ChatGPT out of the box. ChatGPT's default voice is "enthusiastic American marketing intern." Claude's default voice is closer to "thoughtful British essayist." Both need editing, but Claude's output needs less.

The prompt I use: "Here's my outline for a blog post: [outline]. Write a first draft following this structure. Use a conversational, informal tone. British English. Include specific examples rather than vague generalities. Do not use the phrases 'It's important to note,' 'In conclusion,' or 'Whether you're a.' Avoid starting paragraphs with 'However,' 'Furthermore,' or 'Additionally.' No em dashes. Contractions are good. Opinions are good. Sound like a knowledgeable friend explaining something, not a textbook."

This prompt does a lot of heavy lifting. It pre-empts many of the worst AI writing habits by explicitly banning them. But the output still won't be perfect. That's what step 4 is for.

Step 4: The edit pass (where the magic happens)

This is the most important step and the one most people skip. The edit pass is where AI-assisted writing becomes human writing. Here's my checklist, in order:

Kill every "it's important to note" and "it's worth mentioning." These are AI filler phrases. They add zero information. If something is important, just say it. You don't need to announce that you're about to say something important.

Remove escalating lists of three. AI loves to list three things in order of increasing grandeur. "It's faster, more efficient, and fundamentally transforms how teams collaborate." That third item is always a dramatic exaggeration. Cut it or rewrite it to be specific.

Add contractions. AI underuses contractions. "It is" sounds formal. "It's" sounds human. "Do not" sounds like a warning sign. "Don't" sounds like a person talking. Go through and contract everything that sounds stiff.

Add personal anecdotes. This is non-negotiable. AI can't tell you about the time it spent three hours debugging a CSS issue only to realise it had a typo in a class name. You can. These stories are what make writing feel human. Even a one-sentence aside ("I learned this the hard way last March") adds authenticity.

Break up long sentences. AI loves compound sentences with multiple clauses. Humans don't talk like that. If a sentence has more than one comma, consider splitting it into two sentences. Short sentences have impact. They keep readers moving.

Add opinions. This is the big one. AI output is diplomatically balanced. "Tool X has advantages and disadvantages." Humans have opinions. "Tool X is brilliant for prototyping and terrible for production." Replace every balanced assessment with a stance. You can acknowledge nuance without being neutral.

Remove the conclusion paragraph if it summarises the article. If your conclusion starts with "In conclusion" or "To summarise" or restates everything you just said, delete it. Write a final paragraph that adds something new, or just end. Not every article needs a neat ending. Sometimes stopping is better than wrapping up.

Step 5: Typo check with Grammarly (ignore the style suggestions)

grammarly is excellent at catching typos, spelling mistakes, and genuine grammatical errors. It is terrible at style suggestions. Grammarly's style suggestions will push your writing toward the exact kind of formal, corporate tone you're trying to avoid.

Use Grammarly for red and blue underlines (errors and clarity issues). Ignore the purple underlines (style suggestions). They'll tell you to remove contractions, formalise your tone, and add transitional phrases. The opposite of what you want.

prowritingaid is an alternative if you want more nuanced style feedback. Its "style" report is better than Grammarly's at identifying genuinely problematic patterns (passive voice, repeated sentence starters, overused words) without trying to turn everything into corporate communication.

The red flags checklist

Print this out (or don't, I'm not your mum) and check your finished article against it. If you find more than three of these, your article still sounds AI-generated: - Starts with a broad, sweeping statement about the state of the industry - Uses "landscape" to describe anything other than actual scenery - Contains the phrase "a myriad of" - Any paragraph starts with "However," "Moreover," or "Furthermore," - Uses the word "crucial" when "important" or "useful" would do - Lists benefits in groups of three with escalating importance - Contains the sentence structure "Not only [X], but also [Y]" - Any sentence contains "it's worth noting" or "it bears mentioning" - The conclusion summarises every point made in the article - There are no contractions in any paragraph - No personal opinions or experiences appear anywhere - Every paragraph is roughly the same length - No jokes, asides, or conversational tangents - The tone is consistent throughout (real humans vary their tone)

The meta-point

Here's the thing nobody says about AI writing: the editing process I've just described takes about 30 minutes for a 1000-word article. Writing that same article from scratch would take me about 90 minutes. So AI plus editing saves me about an hour per article.

But that hour is only saved if I do the editing. If I skip the edit pass and publish the raw AI output, I've saved 90 minutes but published something that my readers can immediately identify as AI-generated, which undermines the trust that makes them read my stuff in the first place.

The edit pass isn't optional. It's the whole point. AI gives you a fast rough draft. Your job is to make it sound like you. If you're not willing to spend 30 minutes doing that, you're better off not using AI at all, because bad AI writing is worse than no writing at all. At least silence doesn't actively annoy people.