Your AI tool knows how to write like you—but it's teaching every other creator to write exactly like you too, and audiences are starting to notice.
I watched this happen to three creators I respect in the last six months. Their newsletters felt slightly off. Not bad, exactly. Just smoothed out. The jagged, specific, occasionally weird qualities that made them worth reading got replaced by something competent and forgettable. When I asked one of them about it, she said she'd been using Claude to draft everything since January. Output tripled. Engagement dropped 31%.
That's the trade you're making when you hand AI your voice without a system to protect it.
The Invisible Commoditization Happening Right Now
AI writing tools are trained on aggregate patterns. When they "learn your style," they're not learning your style—they're learning the statistical overlap between your writing and every other writer in their training data who sounds vaguely similar to you.
If your voice is "sharp financial commentary with blue-collar roots," the model finds every token pattern that fits that description across millions of documents. It gives you back something that hits the general shape of that identity. But the specific details—the way you say "that's garbage and here's why" or your habit of opening with your grandmother's money advice—those get averaged out in favor of what performs broadly across that category.
Morgan Housel built a following on financial writing that reads like Hemingway. Run his essays through an AI analyzer, then ask GPT-4 to "write like Morgan Housel," and you get something that uses short sentences and financial topics. You don't get the philosophical weight, the exact ratio of story-to-insight, the precise moment he pivots from anecdote to implication. The model gives you the average of everyone who influenced him, not him.
At scale, this produces content convergence. Different newsletters, same cadence. Different podcasts, same three-part structure. Different YouTube channels, same hook-conflict-resolution arc. Tools optimize for what gets clicks broadly. Broad optimization kills distinctive positioning.
The Data Paradox: Your Best Work Is the Problem
Here's what most creators miss: the better your existing content, the more damage AI does to your future content.
When you feed an AI your top-performing pieces to train a custom style, you're giving it your most algorithmically optimized work. High-performing content bends toward what audiences respond to broadly. It's your most audience-facing material—which means it's also your most compromised material in terms of raw personal distinctiveness.
Your best essay isn't your most you essay. It's the essay where your voice and audience preference overlapped perfectly. Feed that to an AI, and the model learns to optimize for audience preference with a thin veneer of your stylistic tics on top.
I tested this with my own writing. I gave Claude 12 of my highest-performing articles and asked it to draft a new piece in my style. It produced something competent that got 40% fewer highlights and saves than my average piece. Then I gave it 12 of my lowest-performing but personally favorite pieces—the weird ones, the niche tangents, the essays where I followed an idea somewhere uncomfortable. The draft from that set got 2.3x more comments than the first version.
The model trained on my failures was more me than the model trained on my successes.
Your AI training data should be your most idiosyncratic work, not your most viral work. The pieces where you went too far. The takes that got mixed responses. The format experiments that didn't land. That's where your actual voice lives.
Building a Voice Architecture
The solution isn't avoiding AI. It's building what I call a Voice Architecture—a structured system that uses AI for production tasks while keeping your creative decisions human-generated.
Start with a Voice Document far more specific than typical prompts. Most people write: "Write conversationally, use short paragraphs, be direct." That's a description of half the internet.
Your Voice Document needs to capture:
Sentence rhythm patterns. Not "use short sentences" but "alternate between 6-word punches and 25-word expansions, never three short sentences in a row." Analyze 10 pieces of your actual writing and quantify the patterns.
Specific forbidden phrases. Every creator has verbal tics they'd never use. Mine include "at the end of the day," "in this economy," and "let's unpack that." Add these as explicit exclusions.
Your contradiction signature. What do you believe that your audience finds slightly uncomfortable? Casey Neistat built his brand partly on "hard work beats talent" in a creative space that celebrates natural gifts. That tension is a voice element. Name yours.
Category-specific references. What do you compare things to that no one else in your space compares things to? If you're a marketing creator who always connects campaign strategy to chess, that's a voice anchor. Document it.
Your prompts stop being style requests and start being constraint systems. The difference between "write like me" and "write a 600-word section that opens with a counterintuitive claim, uses data from 2023 or later, avoids the word 'leverage,' and ends before the conclusion feels natural" is the difference between AI-generated and AI-assisted content.
The Authenticity Markup Is Real and Growing
Audiences are developing something I'd call AI taste—a felt sense that content was produced by optimization rather than by a specific human mind. They often can't articulate it, but the behavioral data is clear.
Substack's 2024 creator report showed that newsletters with high comment engagement grew subscribers at 2.4x the rate of those with high open rates but low comments. Comments require something specific to respond to. Optimization produces content that's hard to disagree with—which means it's also hard to engage with.
Sam Sulek built 4 million YouTube subscribers with a camera held wrong, terrible lighting, and zero scripting. His retention metrics shouldn't work by any algorithmic logic. They work because every video is provably him. Patrick Boyle hit 900,000 subscribers using a dry British academic voice that any AI tool would flag as "needs to be more engaging."
The premium is moving toward distinctiveness that seems inefficient. Audiences aren't consciously boycotting AI—they're unconsciously gravitating toward content with friction, specificity, and human fingerprints.
If you're charging for premium access (paid Substack, course, community), your pricing power is directly correlated to how distinctly you you are. Commoditized voice. Commoditized price point.
The Hybrid Production Framework
I use a hard separation between what I call Production Layer and Creative Layer decisions. AI works in production. I work in creative. The two never switch places.
Production Layer (AI-appropriate):
- Research synthesis: turning 40 pages of source material into a structured brief
- SEO metadata: title variations, descriptions, tag suggestions
- Formatting: transcripts into structured sections with headers
- First-pass editing: passive voice, run-ons, spelling errors
- Distribution copy: subject line variations, social clip descriptions
Creative Layer (human-only):
- The opening line. Always.
- Any analogy or metaphor. If I didn't generate the comparison, it doesn't appear.
- The specific data point or example I choose to center an argument on.
- Counterarguments. This is where voice lives most—how you handle opposition.
- The ending. The last thing you say carries your name forward.
Test: if removing it would make the content work equally well under someone else's name, AI can touch it. If removing it would make the content unattributable to you, you write it.
Here's my concrete workflow for a piece on pricing psychology: I freewrite my actual opinion for 20 minutes before touching any tool. I note three specific examples I've personally observed, not read about. I identify the one thing I believe that most pricing consultants would push back on. That all goes into a handwritten note.
Then I ask Claude to pull research on pricing psychology, summarize relevant studies, and flag counterarguments I haven't addressed. It returns structured material. I take my handwritten notes and the research, then write the piece myself. I might use AI to clean up a structurally messy paragraph. I never use it to originate an idea or a sentence in a section that carries an argument.
This takes 2.5 hours for 1,500 words. Before AI, four hours. I got 40% faster without AI touching the content that makes the piece mine.
The creators losing their voice outsourced the 20-minute freewrite. That's the part that feels inefficient. It's also the only part that matters.
Where This Goes
AI writing tools will get better at mimicking individual voices. The analytical gap between AI-generated and human-generated prose will narrow.
What won't narrow: the gap between content produced by someone with a genuine point of view and content produced by an optimization process. A more convincing AI impression of your voice is still an impression. Your specific observations, your personal intellectual history, the particular combination of things you've read and done and failed at—that's not extractable from examples. It only gets generated by you, in real time, having actual thoughts.
The creators winning in a world of abundant AI content will be the ones harder to replicate, not easier to scale.
Scale your production. Keep your voice human.
Follow for more practical AI and productivity content.
