I Fed ChatGPT 500 of My Blog Posts...

Last month, I decided to run an experiment that's been nagging at me for weeks.

Could AI actually learn to write like me?

Not just mimic my sentence structure or vocabulary, but capture the rhythm, the humour, the weird metaphors that make my writing... mine?

So I fed ChatGPT every blog post I've written over three years. Over 500 pieces. Roughly 300,000 words of my unfiltered thoughts, rants, and ideas.

The results were fascinating and frustrating in equal measure.

What I discovered about AI's ability to replicate human voice completely changed how I think about using these tools for content creation.

P.S. - The one element AI absolutely cannot replicate (even with perfect training data) explains why your authentic voice just became your most valuable business asset.

The Two Roads: Easy Mode vs. Deep Training

Before diving into what worked and what failed spectacularly, I want to share something you might have missed:

There are two completely different approaches to teaching AI your writing style, and most creators choose the wrong one first.

Approach #1: Custom Instructions (The 20-Minute Method)

This is the accessible route, no coding, no technical knowledge, works immediately.

Here's the process:

1. Gather 3-5 writing samples that represent your authentic voice

2. Ask ChatGPT to analyse your style patterns

3. Convert that analysis into reusable instructions

4. Save it in your custom instructions or prompt library

What the AI typically identifies:

• Sentence length preferences (short and punchy vs. long and flowing)

• Vocabulary choices and recurring phrases

• Formality level (academic, conversational, casual)

• How you structure arguments and transitions

• Your opening and closing patterns

Example instruction it might generate:

"Write with a confident yet informal tone, as if you're a knowledgeable friend giving advice. Use personal anecdotes when relevant. Favour short, punchy sentences over elaborate prose. Occasionally deploy unexpected metaphors."

Advantages: Immediate, free (beyond ChatGPT subscription), easily updated

Limitations: Takes up context window space, captures only surface-level patterns, requires you to paste instructions into every conversation

Approach #2: Fine-Tuning via API (The Technical Deep Dive)

This involves actually retraining the model's parameters on your specific writing, creating a customised version that has internalised your patterns.

Requirements:

• Dataset of 50-500+ prompt-response pairs in JSON format

• Access to fine-tuning APIs (OpenAI, Google, or open-source models)

• Understanding of training parameters and costs

• Time investment: multiple hours to days

The conventional wisdom on dataset size:

• Minimum viable: 10-50 examples for basic style adaptation

• Recommended: 240-480 examples for solid accuracy

• Optimal: 1,000-2,000+ examples for capturing nuanced personality traits

One academic study found that fine-tuning with 240+ examples achieved 94.5% expert-rated accuracy in matching personality traits in scholarly writing.

Advantages: Model "thinks" more like you from the ground up, requires shorter prompts, potentially captures subtler patterns

Disadvantages: Expensive, time-consuming, requires retraining as your style evolves, technically demanding

Most creators should start with custom instructions.

Move to fine-tuning only if you're producing massive content volume and have technical resources.

What AI Actually Learns (And What It Desperately Fakes)

After training ChatGPT on my writing using both methods, I tested it across 50 different prompts, some matching topics I'd written about before, others completely new.

The Wins: What AI Captured Surprisingly Well

Surface-level patterns were nearly perfect.

The AI nailed:

• My preference for em-dashes (I use them constantly—probably too much)

• Sentence length distribution (I alternate between short punches and longer explanatory flows)

• Vocabulary choices (specific terms I favour, industry jargon I use naturally)

• Paragraph structure (I typically open with hooks, build through examples, and land on insights)

• Formatting habits (how I use bold, bullets, subheadings)

Tone and formality translated well.

The AI successfully distinguished between my casual blog voice and more professional business writing. When I asked it to write in "blog mode" vs. "client proposal mode," it adjusted appropriately.

Domain-specific language was absorbed completely.

All my marketing terminology, AI concepts, and business frameworks appeared naturally in the AI's outputs. It understood my typical reference points.

One test actually fooled two colleagues who regularly read my work. They thought an AI-generated piece about content strategy was actually mine.

That felt like success... until I ran the next set of tests.

The Failures: Where AI Falls Embarrassingly Short

Here's what the research actually shows:

A comprehensive study analysing 40,000+ AI generations across 400+ authors found that LLMs struggle significantly with informal, personal writing styles, exactly the kind found in blogs and personal essays.

Read the full study here: Catch Me If You Can?

The more subtle and implicit your style (the "you" that emerges between the lines rather than in word choices), the harder it is for AI to replicate.

Rhythm and cadence remained completely elusive.

Human writing has musicality, a rhythm in how ideas flow, how sentences build momentum, how paragraphs breathe.

I write with intentional pacing. Sometimes short sentences hit hard.

Then I'll stretch into longer exploration, building layers of context before landing the insight that makes it all click.

The AI? It maintained a technically correct variety but missed the why behind the rhythm. The strategic pause before a punchline. The building tension. The unexpected release.

Comedy writers studying this phenomenon found that AI fundamentally fails at timing, which is essential to both humour and compelling prose.

Authentic humour stayed completely beyond reach.

This was the most disappointing failure.

I use observational humour, irony, and situational comedy throughout my writing. The AI could generate puns and basic wordplay, but it completely whiffed on:

• Observational humour that requires cultural context

• Sarcasm (it often took my sarcastic examples literally)

• Ironic juxtapositions

• Callbacks to earlier jokes in the piece

Research on professional comedians confirms this: AI-generated jokes lack the "momentum" and contextual awareness that makes humor actually land.

One researcher captured it perfectly: "Humour thrives on defying expectations through surprising connections, while LLMs are trained to produce expected patterns—making them the antithesis of genuinely funny."

Personality quirks got smoothed into generic averages.

My writing has weird verbal tics. Unusual metaphors. Odd analogies. Distinctive ways I combine seemingly unrelated ideas.

These elements—the ones that make readers think "that's so [your name]"—rarely transfer authentically.

AI tends to smooth out irregularities, defaulting toward more average, generic patterns.

Emotional authenticity remained AI's Achilles heel.

A powerful essay in WIRED described a writer using GPT-3 to write about their sister's death.

The AI could acknowledge the loss but kept missing the emotional truth, substituting generic grief narratives even with extensive prompting.

The subtle emotional register, the vulnerability, the raw honesty that connects writers to readers—these emerged inconsistently at best.

The "AI Slop" Problem Nobody Warns You About

Perhaps most concerning: AI-trained models often introduce their own stylistic tics that can actually overwrite your original voice.

The telltale AI phrases that crept into outputs:

• Overuse of "delve," "moreover," "furthermore," "it's important to note"

• Everything becomes a list of three examples (the infamous "rule of three")

• Mechanical formality: "In conclusion, it is essential to highlight..."

• Excessive transitional phrases and over-explanation

• Diplomatic, non-committal language that avoids strong stances

One writer whose polished human writing consistently scores as "AI-generated" by detection tools captured this irony perfectly: "I don't write like AI—AI writes like me."

The convergence toward "correct" patterns in both AI training and formal writing instruction creates a bland middle ground that lacks personality.

Even after training on my specific voice, I had to constantly edit out these AI-isms.

The Real-World Test: Can Readers Actually Tell?

After generating 50 pieces of "AI writing in my voice," I ran an experiment.

I mixed 10 AI-generated articles with 10 of my actual articles and sent them to my email list with a simple question: "Which ones did I actually write?"

The results:

For structured content (how-to guides, tutorials, listicles): Readers correctly identified authorship only 62% of the time. Essentially a coin flip.

For personal essays and opinion pieces: Readers correctly identified authorship 89% of the time.

The more personality-driven and informal the writing, the easier it was to spot the AI.

As one reader put it: "The AI version is technically good, but it feels like you on autopilot. The real ones have that spark of 'I can't believe they just said that.'"

This matches the academic research perfectly:

Structured writing (news, emails, tutorials) shows reasonable AI attribution accuracy. Informal writing (blogs, essays, personal narratives) sees AI-generated content frequently detected as different.

The Hybrid Approach That Actually Works

After three months of experimentation, here's what I've learned:

Don't try to make AI your clone. Make it your collaborator.

The most effective workflow:

1. Use AI for structural scaffolding

• Feed it bullet points outlining your argument

• Let it draft connecting prose

• Edit heavily for voice and personality

2. Employ it for polishing, not creation

• Write your rough draft in your authentic voice first

• Use AI to tighten language, improve transitions, catch issues

• Think of it as an intelligent editor, not a ghostwriter

3. Accept the 90/10 rule

Tiago Forte who has published 500 blog posts and 11 books, developed a comprehensive AI system that now drafts "90% of my content."

But that 90% still requires hours of editing, rearranging, and adding the touches that make it truly his.

That final 10% the personality, the surprising connections, the emotional resonance, remains irreplaceably human.

4. Focus on your unique value-add

AI increasingly handles:

• Research synthesis

• Structural organization

• Initial drafting

• Grammar and polish

Your unique contribution becomes:

• Taste and editorial judgment

• Creative connections nobody else would make

• Emotional intelligence and vulnerability

• The distinctive perspective from your personal experience

What This Means for Your Content Strategy

If you're building a content-driven business, several practical implications emerge:

Start with prompting, not fine-tuning.

Unless you have extensive technical resources or thousands of writing samples, begin with custom instructions. Immediate value without infrastructure investment.

Build systems, not one-off prompts.

Create reusable prompt templates, maintain updated style guides, and develop workflows that consistently incorporate AI at the right stages.

The compound returns from systematic AI integration far exceed sporadic use.

Recognise the authenticity gap.

While AI can approximate your style, truly authentic writing, especially humour, vulnerability, and personality quirks, still requires direct human authorship.

Use AI for volume. Reserve purely human writing for pieces requiring maximum authenticity.

Embrace the evolution.

Writing in the age of LLMs isn't about AI replacing writers. It's about writers who use AI effectively, replacing writers who don't.

The skill shifts from generating every word yourself to directing, curating, and refining AI-assisted output with your distinct editorial vision.

Before You Go

If this changed how you think about AI and authentic voice, share it with another creator still debating whether to use these tools.

Also, here is a free n8n template to use AI to automate the generation of blog content :

New to AI but curious about what's possible?

Subscribe here for weekly tutorials that actually make sense.

No jargon, no hype, just step-by-step guides you can follow.