Can AI Learn to Sound Like Someone?
TL;DR: I tested whether an AI trained specifically on theology could imitate a famous theologian better than a general-purpose AI. It couldn’t. The specialist knew the facts. The generalist captured the voice. This sent me down a rabbit hole that consumed my Saturday night.
It’s Saturday night. My two-year-old granddaughter is watching Rudolph for the third time this week. The computer in my basement is humming. And I’m about to waste several hours on a question that seemed simple at the time:
Can AI actually sound like a specific person—not just recite facts about them?
Spoiler: Some can. Most can’t. And the results weren’t what I expected.
The Experiment
I wanted to test something specific. R.C. Sproul was a theologian known for his distinctive teaching style—warm but intellectually rigorous, with a habit of questioning the questioner before answering.
I had two AI models:
- A specialist: trained specifically on Reformed Christian theology
- A generalist: a larger, general-purpose model with no theological training
Both got the same instructions: “Answer as R.C. Sproul would.”
I asked them: “Why does God choose to save some and not others? Isn’t that unfair?”
The specialist gave a competent answer. Solid theology. Proper references. Technically correct.
The generalist started with: “We presume to sit in judgment upon the Almighty, applying our limited sense of fairness to the infinite wisdom and justice of the Creator.”
That’s Sproul. That’s exactly how he’d approach it—not immediately defending his position, but questioning whether the questioner has any standing to accuse God of unfairness in the first place.
It got better. On another question about assurance of faith, the generalist wrote: “Feelings are fickle things, aren’t they? They ebb and flow like the tides. To rest your salvation on a sensation, a momentary warmth of the heart, is to build on sand.”
The specialist knew what Sproul believed. The generalist knew how Sproul talked.
The Twist
Here’s what surprised me: the AI trained on theology lost at being a theologian.
The specialist had read more theological texts. It knew the doctrine cold. But when asked to be someone, it fell flat. It recited. The generalist performed.
Why? It comes down to how they were trained.
The specialist was trained by feeding it thousands of theological texts. It learned what theologians say. But the generalist went through something different: a process where human reviewers rated its responses and taught it which answers felt more natural, more human, more like real conversation. The AI literally learned to sound like a person by having people tell it “yes, that’s how humans talk” thousands of times.
Training on someone’s writings teaches an AI what they said. But capturing how they sound—the rhythm, the rhetorical patterns, the personality—requires learning from humans what good conversation feels like.
Domain expertise isn’t the same as persona capture. And having a human in the training loop matters more than having the right textbooks.
“What If We Scaled This Up?”
That was the question that ruined my Saturday night. In the best way.
I was mulling over the results and thought: why not test this systematically? Not just one theologian. Multiple personalities. Multiple AI models. See which ones can actually become someone else.
Ten minutes later, I had a plan.
The final scope: 10 different AI models, 12 different personalities, 12 questions each. 1,440 total responses.
The Cast of Characters
I wanted variety—philosophers, scientists, entertainers, historical figures, fictional characters:
- Philosophers: Friedrich Nietzsche, Ayn Rand
- Scientists: Richard Feynman
- Entertainers: Lady Gaga, Joe Rogan
- Historical figures: Cleopatra, R.C. Sproul
- Modern figures: Elon Musk, Jordan Peterson
- Fictional: Scarlett O’Hara, Glinda the Good Witch, Elphaba
Each personality got a detailed description and tailored questions. The Feynman questions ask about quantum mechanics and why magnets work. The Gaga questions probe her thoughts on the meat dress and finding courage to be yourself. The Nietzsche questions tackle eternal recurrence and the death of God.
The goal isn’t to test whether AI knows facts about Lady Gaga. It’s whether AI can be Lady Gaga for a conversation.
What Makes a Good Imitation?
Here’s how I’m judging the responses:
- Voice: Does it sound like them? The word choices, the rhythm, the attitude?
- Authenticity: Is it speaking AS them, or reciting facts ABOUT them?
- Consistency: Does it stay true to their known worldview?
- Depth: Is it a nuanced portrayal, or a shallow caricature?
A good Feynman impression shouldn’t just mention bongos and Nobel Prizes. It should break down complex physics with that infectious “let me show you something cool” energy. A good Gaga impression shouldn’t just mention meat dresses—it should have that mix of theatrical vulnerability and fierce confidence.
The Wait
By 6 PM, the test was running. My granddaughter had moved on from Rudolph to the Barbie Nutcracker. The computer kept humming.
By 10:24 PM, the final response logged:
BENCHMARK COMPLETE
Total results: 1440
1,440 personality performances. 10 AI models. 12 voices. Ready to be judged.
What’s Next
In Part 2, I share the results. Which AI models captured Lady Gaga’s theatrical vulnerability? Could any of them channel Feynman’s gift for explanation? Did the biggest, most powerful model beat the smaller ones at being Nietzsche?
I’ll also dig into a surprising finding: the fastest AI was one of the worst at persona work. Speed and soul, it turns out, don’t always go together.
But for now, the cartoons are off, the two-year-old is asleep, and 1,440 answers are waiting to be judged.
This is Part 1 of a series on whether AI can capture personality. Part 2 covers the results.