I cloned my voice for a podcast. My editor caught it right away.

Jason edits this podcast. I sent him a finished episode and this is what came back:

“So was this AI generated? I think it’s pretty noticeable. The intonation and spacing is weird. Really weird, with punctuation and intonation in the wrong places. Just got some real stick for this on his new episode, people complaining that it was AI, and there’s only a few sentences changed with the AI. But people notice, they tune out, and they unsub. If this episode is every word AI, I’d be hesitant doing this if I was you. Your personality is the best thing about your podcast and what people love. This just zapped it, in my honest opinion. Are you sure you want me to put it out?”

I wrote back: yo bro, I trust your call. I like it as a test, but could you tell right away?

“First I thought you were sick, not yourself, in a low mood. Then as I got to the end of sentences it just seemed odd. Then I was convinced you were reading a script and it’s not your usual style. By the end I’d already twigged it was AI before you even said. I always said the best thing about your show is your personality.”

I don’t know what the hell a twig is.

How I made it

I asked Anna, our COO at AppSumo, to send me everything we’re doing with AI so I could share it on Twitter. She sent a document. I told Claude to use my voice and some of my setup to turn it into a Twitter article, I edited it, and I put it out. That became the Claude for All post. I was proud of it. It was good, it took less time, and I went through every line.

So, could I do the same for the podcast? I have a second daughter now. My hours are much, much less. Tell Claude to turn the article into a podcast script, upload it to ElevenLabs, and maybe I put out a good episode monthly, possibly weekly, possibly daily.

The script came back and I thought it was great. It didn’t have the Rabbi Can’t Lose stuff, so I added that in. On ElevenLabs you upload your own audio, so I gave it two hours of me from older shows. $11 a month. I gave it the script and hit go.

When I listened back, it sounded like me. A slightly measly Jewish voice. Hey guys, it’s Noah Kagan. It wasn’t me, but it was kind of me. I was going to put it out.

Everything in it was true, because it came from Anna’s document. We were spending $56,000 a year on AI tools: individual Anthropic accounts, OpenAI subscriptions, Cursor, a dozen things floating around and nobody sharing anything. We killed every subscription and put the whole company on one Claude Team plan, $24,000 a year, every teammate gets a seat. Three days of training, a Slack channel called #ai_daily, shared plugins that sync through GitHub so when one person builds a skill everyone has it. Dusty’s deal of the day went from 25 minutes to 60 seconds. Angelina, our payroll specialist, had never opened Terminal before and in her first week automated an audit she used to do by hand twice a year. $32,000 saved, three weeks to full adoption.

All of that is real. And the whole thing still sounded like a guy who was sick.

The coffee shop

The same day I was at a coffee shop with my buddy Joe from Unchained and a guy named David comes up. David Van Podcast is what I called him. Hey man, I love your podcast.

Funny timing. A lot of us want to do a little work and get great results. I was hoping to automate the whole thing and still put out episodes you’d like. And this guy walks up because of the version where I put in the work, the effort, and the personality. That part is what’s valuable, and it’s required.

Copying a template and sending it, AI is amazing at. Creative work, video thumbnails, audio, a personality: it’s not there yet. It’s general. It’s trying to learn you by pulling from everything else to recreate you. Maybe in time the models get so good you can’t tell the difference. I keep thinking AI should be like spell check. Before spell check, things had a lot of spelling errors. The question is how you work with it to make yourself better, so instead of a ton of errors you sound superior. Iron Man, iron woman, whatever. Superpowers on your own abilities.

Jason ended the thread like this:

“I think this AI working creative roles is kind of a race to the bottom sometimes. It’s great for organizing and admin. All the things you mentioned in the episode are really cool things that AI does well. They’re just no good at being a person, or a creative one at that. Something a lot of your peers in the tech world might need reminding of before we spin into a black hole of bland homogenized guff.”

Hell yeah, Jason.

I don’t know if the moral is hold yourself to a high standard, or just ask honestly, is this actually as good as it could be? If you put out something and people know it’s not as good as it could be, or it’s not really you, they’re going to know.

Going forward I most likely won’t use AI on this show. If I do, I’ll tell you it’s a test. The fake episode is in the audio of this one. Go listen, message me on X at noahkagan with whether you could tell, and I’ll Venmo the first five people five bucks.

From episode 386 of my podcast, Noah Kagan Presents. Listen on Spotify or Apple Podcasts.