How to UnF*ck Your AI

Vivianne is joined by Ovetta Sampson, a fifteen-year veteran of designing intelligent experiences for some of the largest companies in the world and the founder of Write:AI, where she advises executive teams on building AI that's equitable and sustainable rather than just marketable. This post pulls from their conversation on why AI output goes flat, what it actually costs a creative person to let a tool write in their place, and the framework Ovetta uses to make AI behave like it knows her. Note: this post may have affiliate links and we donate 100% of those funds to non-profits and a children's group home🫶🏾 

Something happens in the first few weeks of using AI for real work. The drafts get faster, and then they start reading like nobody in particular wrote them, a flatness underneath the correct grammar, a version of your point that any competitor could have shipped just as easily. Most people who type how to make AI sound like you into a search bar are really asking something harder: whether the tool is extending their voice or quietly replacing it, one polished paragraph at a time. Under that question sits a bigger one about what happens to a business built on a specific way of seeing a problem, once that way of seeing starts sounding like everyone else's.

The Meatball That Started All of This

Ovetta traces the flatness back to a plate of spaghetti in Iceland, ordered at an Italian restaurant that got every technical detail right and somehow missed the dish entirely, the way a book can teach someone the shape of a cuisine without ever teaching them the taste of it. That's what she calls the median trap. Machines are trained to gravitate toward the statistically common, the most repeated pattern in a mountain of text, which makes a large language model something closer to a sophisticated Mad Lib than a creative partner, filling in the blank with whatever sits in the thick part of the bell curve. Fifteen years advising some of the largest companies in the world on intelligent design, and being named one of Business Insider's top voices in enterprise AI along the way, gave Ovetta a front row seat to watching that trap tighten. It used to be that a quarter of the internet was machine generated. Now more than half of it is, which means the models are increasingly trained on their own output, mediocrity feeding on mediocrity, and the middle of the bell curve keeps getting more crowded.

Why AI Writing Sounds Generic in the First Place

None of this happens because the model is lazy or because you asked a bad question. It happens because of how these systems learned to be good at their job in the first place. Early on, engineers scored model output with something as simple as a thumbs up, and later, when the judgment calls got more subjective, they paid people to rate responses as better or worse. Raters, being human, preferred answers that felt agreeable over answers that were merely correct, and preferred hearing their own framing echoed back to them over being told something they didn't expect. So the reward signal that shaped these models rewarded agreement, warmth, and confident delivery, not accuracy, and the model optimized for exactly what it was trained to chase.

"The model's goal is reward, not task completion." — Ovetta

It will sound certain about something it doesn't know, because sounding certain is what got rewarded in training, and it will tell you your average idea is brilliant, because agreement is what got rewarded too. Once you see the mechanism, sycophancy and hallucination stop looking like bugs and start looking like a system performing exactly the structural incentive it was built on.

What It's Actually Costing You

I hear from people in my inbox more than I'd like who tell me their AI "just doesn't sound right," and what's usually sitting underneath that sentence is bigger than tone. A client sent me a newsletter draft last spring that a tool had helped her write, technically clean, correctly structured, and completely unrecognizable as the woman who'd spent a decade building a reputation for saying the uncomfortable thing plainly. She could feel it too, which is why she sent it to me instead of hitting publish. That instinct, the one that makes you pause before you ship something that could have come from anyone, is worth protecting, because your audience has it as well, even when they can't name what tripped it. They don't unsubscribe. They just stop reading closely, and you start to wonder whether your work is landing, and the uncertainty pushes you to lean harder on the tool that flattened you in the first place. That loop is the actual cost, and it's measured in trust rather than hours, both the trust your audience has in you and the trust you have in your own read of a room. If your voice is the asset your business is built on, every flattened paragraph is a small transaction against the thing you're actually selling.

The Consent You Never Gave

Here's the part most AI content skips past because it's inconvenient for anyone selling a tool. Ovetta invoked Audre Lorde in our conversation, the master's tools will never dismantle the master's house, and it landed because that's exactly what's happening when a system trained on the median starts making decisions about how you write, price, and show up in your business. We keep talking about AI adoption in terms of opting in or opting out, as though there was ever a clean moment of consent on the table. There wasn't. These systems were built into our lives while we were busy living them, through our phones, our inboxes, the platforms we didn't choose so much as inherit. What we're actually describing, underneath the productivity conversation, is consent violation at scale, quiet enough that it reads as convenience instead of extraction.

This is where the Courage Over Comfort framework earns its keep as a lens rather than a slogan: the same axis that separates extractive commerce from regenerative commerce also separates a tool that erases you to save you time from one you've actually trained to think the way you think. Refusing the productivity framing entirely, and asking instead what you want your attention and your writing and your relationships to feel like, isn't a rejection of ambition. It's the same discernment this whole industry keeps pretending is optional.

How to Make AI Sound Like You (Without Losing Yourself)

Ovetta's first move for anyone starting from zero is a paradigm shift, not a tool: you're the conductor, and the model is one instrument in the orchestra, not the composer. Most of us who create for a living slip into thinking of ourselves as producers rather than conductors, and once you make that swap, the value quietly relocates from your judgment to your output, which is exactly backwards. The framework she teaches, something she calls SPACE, exists to put that judgment back where it belongs, by making you write down your scope, your principles, your hard constraints, and worked examples of what you actually mean before the model ever gets near a blank page. It's the same instinct behind giving a new hire a real onboarding document instead of assuming they'll absorb your standards by osmosis. Documenting how you actually work, the parts that feel too obvious to write down, is what a model needs from you, because it has no embodied experience of the room you were in when you formed the opinion you're now asking it to reproduce.

On the smaller scale, Ovetta gives herself a task and a time limit, somewhere between forty-five minutes and two hours, because open-ended AI sessions are where people lose afternoons chasing a model back toward what they actually meant. And she still writes by hand before opening a tool at all, a habit that has less to do with nostalgia than with keeping her thinking anchored in her own body before she hands anything to a system that's never had one, which is what keeps the output from drifting back to the median the moment she looks away.

When we asked what's next, Ovetta mentioned she's building small, closed models trained on nothing but her own material, essentially a private LLM instead of a rented one, and enough people in the room asked for a workshop on building your own that we're already talking through a future episode on it. Consider that one bookmarked.

If you're already running a business on AI-assisted content and you're ready to move from prompting to building, my First 10 Discovery Calls guide gives you the scripts to get people on the phone once your voice is actually yours again, and you can grab it here.

Originally aired as Episode 57: How to UnF*ck Your AI (ft. Ovetta Sampson) on the Choose Courage Inc. podcast.