The short version: AI content sounds generic because people type a topic into ChatGPT and hit enter, not because AI can't write in your voice. Feed it your own words first, real emails, real transcripts, real old posts, and the sameness disappears within a few tries. It took me three hours to fix this and I should have done it two years earlier.
The LinkedIn post that made me stop and read my own writing back
Last spring I asked ChatGPT to write a LinkedIn post about small business owners feeling overwhelmed by AI tools. It came back with "In today's fast-paced digital landscape, small business owners are navigating unprecedented change." I read it twice and thought: I would never say that. Nobody I know talks like that. My own mother doesn't talk like that.
I posted it anyway because it was 9pm and I was tired. It got 11 likes. My next post, written by me in about the same amount of time, got 340 likes and four people replied asking to book a call. Same topic, same week, same audience. The only difference was that one sounded like a corporate brochure and one sounded like a 53 year old woman from London who's had a rough five years and isn't pretending otherwise.
That gap bothered me enough that I spent an afternoon working out why it happens and what fixes it, and the answer turned out to be simpler and more annoying than I expected.
Why AI content converges on the same voice
Here's the uncomfortable bit nobody selling AI courses wants to say out loud: the sameness isn't a limitation of the model, it's a limitation of the input. When you type "write me a LinkedIn post about time management for small business owners" into ChatGPT with no other context, it has nothing to pull from except the average tone of every business blog it was trained on. It writes like a McKinsey deck crossed with a motivational poster because that's the statistical middle of business writing on the internet.
Millions of people are typing near identical prompts into the same handful of tools right now. That's why LinkedIn feels like it's been colonised by posts that all start with a one-line hook, a line break, then "Here's what I learned." It's not that AI can't do better. It's that almost nobody bothers to give it better material to work with.
I see this constantly with clients. A founder will show me a blog post their AI tool spat out and say it "doesn't sound like us," and when I ask what they fed it, the answer is almost always: nothing, just the topic. You wouldn't hand a new copywriter a one-line brief and expect them to nail your brand voice on the first try. AI is the same, except it's faster to fix because it doesn't need six months and a probation period.
What I did, step by step
I built what's usually called a custom GPT, though you don't need anything fancy, ChatGPT's "create a GPT" feature or a simple project folder in Claude does the job. Here's exactly what I did:
- Pulled 47 of my own old blog posts and pasted them into a single document, roughly 60,000 words
- Added transcripts from four podcast interviews I'd given, because spoken voice catches things written voice misses, like how often I say "look" and "here's the thing"
- Added twelve months of my email newsletters, since those are closer to how I talk to my audience than polished blog posts are
- Wrote a short instruction file, about 400 words, telling it explicitly what to avoid: no "in today's fast-paced world," no three-item lists disguised as wisdom, no em-dashes, no ending every post with "the choice is yours"
- Tested it on five topics I'd already written about, so I could compare the AI version against my original side by side
The whole thing took three hours on a Sunday afternoon. The output on the first try was maybe 70 percent there. By the third round of tweaking the instructions, it was close enough that I use it as a first draft for newsletters now and only rewrite about a fifth of it.
If you run a business and don't have a spare Sunday, this is one of the more useful hours of work an AI implementation coach can do with you, because most of the value is in the instruction file, not the technology, and getting that right the first time saves you months of mediocre content.
The test I use to check if AI content sounds like a person
Read the first line out loud. If you wouldn't say it to a friend over coffee, cut it. That's the whole test. "Unlocking growth in the digital age" is not something anyone says over coffee. "I lost half my income in 2020 and had to work out what came next" is.
Brands that get this right treat voice as a design decision, not an accident. Look at how Duolingo built its brand voice around being a bit unhinged and self aware rather than corporate, or how Calm's marketing strategy leans into slow, plain sentences because the product itself is about slowing down. Neither of those happened because someone typed a generic prompt. Someone decided on purpose what the brand sounds like, wrote it down, and then applied it everywhere, including anything AI helped produce.
Hotjar's marketing strategy is another good example of a company that writes like actual humans work there, with jokes and specific numbers instead of vague claims. That's the bar. If your AI output could have been written for a rival company by swapping the logo, it's not ready to publish.
The part people don't want to hear about time
This is the bit that makes people's shoulders drop when I say it in workshops: doing this does not save you time in month one. Collecting your old content, cleaning it up, writing the instruction file, testing it against topics you already know the answer to, that's real work, and it usually takes longer than just writing the post yourself would have that day.
Want AI doing the heavy lifting in your marketing?
I build the systems that handle the boring 80 percent, so you get your week back. Done properly, with the human kept in.
The payoff comes later, when you've got fifty topics to cover across a quarter and a system that produces first drafts sounding 70 to 80 percent like you instead of 0 percent like you. If you're only ever going to write three LinkedIn posts a month, don't bother building this, just write them. This is a system for volume, not for the occasional post.
I'd compare it to Dale Carnegie's old lesson about genuine interest being the thing people respond to, not clever technique. AI can help you produce more, faster, but it can't manufacture the interest and the specific detail that made 340 people stop scrolling on my post. That still has to come from you, once, upfront, in the material you feed it.
Where this goes wrong for small businesses
The most common mistake I see is businesses feeding the AI their website copy instead of their real voice. Website copy has already been sanded down by three rounds of "let's make this sound more professional," so you end up training the AI on the flattened version of your brand rather than the actual one. Use your emails, your sales calls transcribed, your voice notes to your team, the messages you send customers when something's gone wrong and you're being human about it. That's where the real voice lives.
The second mistake is never updating the training material. Your voice shifts as your business shifts, and content trained on 2023 you will sound slightly off by 2026. I refresh mine every six months, pulling in whatever I've written most recently that got a strong reaction, good or bad.
The third mistake, and this one costs actual money, is publishing the first draft unedited because it "sounds close enough." Close enough is how you end up with a business blog full of posts that are all fine and none memorable. Readers can tell the difference between fine and specific even if they can't name why. A post with a real number in it, a real client story, a real disagreement with common advice, beats a smooth generic one every time, and that gap shows up in engagement tools like interactive calculators just as clearly as it shows up in blog comments, because the pattern is the same everywhere: specificity gets acted on, smoothness gets scrolled past.
A shortcut if you're starting from nothing
If you don't have years of old content to feed a model, record yourself talking for twenty minutes about your business, why you started it, what annoys you about your industry, what a good week looks like. Get it transcribed, feed that in instead. Twenty minutes of you talking naturally is worth more to an AI tool than five polished blog posts, because it captures rhythm and phrasing that writing tends to iron out. This is the fastest route I know to a voice that doesn't sound like everyone else's, and it costs nothing but the twenty minutes.
Frequently asked questions
Will AI detectors flag content trained on my own writing?
Detectors look for patterns in sentence structure and predictability, not for whether you used a tool. Content trained heavily on your own real voice, with specific stories and irregular phrasing, tends to score lower on detectors anyway because it's less statistically average, which is a side benefit rather than the main goal.
How much of my old content do I need to train an AI on my voice?
Somewhere between 20,000 and 60,000 words gives noticeably better results than a handful of posts. I used 47 blog posts plus a year of newsletters, roughly 70,000 words total, and saw a real jump in quality once I passed about 40,000.
Is this worth doing if I only post once a week?
Probably not. The setup takes a few hours and pays off through volume. If you're producing one post a week, write it yourself and save the system building for when your output needs to scale.
What's the single biggest mistake businesses make with AI written content?
Feeding it polished website copy instead of real, unfiltered communication like emails, voice notes, or sales call transcripts. Polished copy has already lost the specific quirks that make writing sound human, so training on it just produces a smoother version of generic.
Related reading: How to See Who Unfollowed You on Instagram (Real Methods That Work in 2026) and Why Your AI Still Writes Like Everyone Else's (And the 45-Minute Fix That Changes That).
