BTS: Why I Stopped Trusting AI With My Voice

Tom: A while back I recorded an entire podcast episode using the first script that AI gave me, almost word for word with no real editing. And when I sat down to watch the footage back before it ever went anywhere, I couldn't do it. I couldn't watch my own episode. And it wasn't that the content was wrong, and it wasn't that the points were bad. It just it wasn't me. It didn't say what I wanted to say, and it didn't highlight the things that I wanted to highlight. And that one uncomfortable moment ended up rebuilding the entire way that I make this show. So today's a little different. I'm not covering a leadership topic so much. I'm talking about my behind the scenes process. Because I think how you build something says just as much about your leadership as what you build. So stay tuned. Welcome back to the Lead With More podcast, where we help you lead with clarity, intentionality, and confidence. And today, clarity and tent and intentionality are exactly what this episode is about. It's just pointed inward at how the show gets made. So here's what happened. I was using AI to help me move faster on content and script generation. There's nothing wrong with that on the surface. I mean everybody's figuring out how to use these tools effectively. But on this particular episode I got busy. I was behind and I took the first draft that AI handed me and I basically ran with it. I recorded it, post produced it, got it basically ready to publish. Then I sat down to do a final watch through, same as I do with every episode. And about ninety seconds in I had to stop. Something was off. The words were fine grammatically, structurally, there was nothing wrong with it, but it did not sound like me. It sounded like a competent stranger explaining leadership. I scrapped the whole episode. Full rewrite. And honestly, in that moment, it felt like a waste. Hours of work just gone. But it turned out to be one of the most useful failures that I've had in making this show. That episode taught me something I think matters way past podcasting. Technically correct content and content that sounds like you are two completely different things. AI is very good at hitting that first one, not so good at hitting the second one. There's a real difference between a script that says true things about leadership and a script that sounds like it came from an actual human being who's lived this stuff. The stories, the phrasing, the little imperfections that are actually just how we talk. Once I knew what to look for, I started noticing the signs. There are certain phrases that AI reaches for by default that I don't actually use in my own conversation. There's a tendency to wrap every point in a neat little bow in a rhythm that's smooth in written form for reading, but not in a way that real conversation goes. None of it is wrong exactly. It's just not me. And in this rapidly changing world of AI, that is important. So here is what I've changed. I stopped treating the AI generated drafts as scripts and started treating them as raw material, more like clay than a finished product. Now the process runs in stages with a real checkpoint at each one. I use AI to help generate angles, structure, and even a first pass at the language. That part's really useful. It saves time. But nothing moves forward until I've gone through it and made it sound like something I'd actually say. Not something that's technically true, but sounds foreign coming out of my own mouth. So here's a high-level view of my AI workflow. I first spend time offline making rough notes on the episode, both on the topic and the flow, including the specific points I want to highlight. I also make note of the takeaways. I then upload these notes into AI, and right now that's Claude, and I run this following workflow on it. So step one is that I instruct Claude to help me develop potential angles for the topic. I have it do a critical analysis of the angles at either a mild, medium, or maximum level to look for potential gaps or stumbling blocks. Once I've talked through the angles and made a decision, we move to the next steps. And when I say talked through, I mean I actually go back and forth with Claude. I don't assume that what it generates is the right thing. I push back until I am satisfied. And then steps two, three, and four where I have AI help me generate the title and thumbnail concept for the episode. Most of the time I end up generating a custom thumbnail from the finished product, but these steps actually give me something to fall back on if I don't like what I'm getting when I reach that final stage. Step five of this workflow is where AI generates a potential show outline. Once that is in satisfactory form, I take it to step six where the draft script is created. At this point, I take the process back offline and rework it into something that sounds like me. I then record and edit the episode. From there, I take the full transcript of the final product and give it back to Claude for the remaining workflow steps. This is where Claude helps me generate chapter titles, social media posts, and potential episode hooks and clips. Here is the key. At every step, I make it a practice to take whatever Claude suggests and then modify it as much as I feel like it needs to match what I want to say. That keeps the process genuine and authentic. I don't take what AI gives me verbatim. AI cannot simulate human thinking or emotional impact. That has to be for me. It's like I'm proofreading, but not for grammar and structure, but for tone and overall message. Now the other piece that came out of this failed episode is that I stopped relying on memory to catch things. I've built checklists for pre-production, production, and post-production in Notion. So I have a repeatable process without relying on my memory. Pre-production is about making sure the outline and draft actually reflect how I would naturally walk through the topic, not just a logically ordered list of points. It also includes a step to check for copied or protected material. I mean, after all, AI can only learn from things it's been told or can find online. Post production is where I watch the episode, specifically listening for anything that sounds out of line for me. It sounds simple and it is simple, but it's the difference between a scrapped episode and one that goes live. Now here's a neat little behind-the-scenes nugget. I actually scrapped the original recording of this very episode because I didn't like how it felt after I watched the final recording. I usually have some level of discomfort watching my own episodes, but if they're really bad, I scrap them and redo them. Okay, so back to the my topic. Here's what I actually learned. Just being honest with you, the real lesson wasn't entirely about AI. It was about outsourcing judgment. As a tool, AI didn't fail me. I did. By handing off a decision that could only be made by me. AI can generate a draft. It can't decide what sounds like me or what sounds like something I want to publish. That call has to stay mine every time. And I'd guess this isn't just a podcasting problem. If you're using AI for emails, presentations, performance reviews, anything with your name on it, the question isn't whether the output is good. It's whether it still sounds like you. That one gut check before you hit send or publish catches almost everything. Okay, I know this is short, but in wrapping it up, let me say this. These tools don't replace your voice. They only work when your judgment stays in charge of them. That's true for podcast script, and it's true for a lot more than that. Now, if you'd like to know more about my processes, reach out. I'd love to talk with you about it. If this episode and this content resonates with you, share it with somebody who needs to hear it. And don't forget to like and subscribe. It really helps this episode, this podcast. Thank you for joining me today, and until next time, keep leading with more.