Research question: Given a newsletter and three articles mentioned therein, can AI make a passable podcast fakecast?
Time constraint: 60 minutes
Setup: Claude 4 Opus, Gemini 2.5 Text-to-Speech
Result: 14-minute podcast
Is it any good? At first, it seems impressive. I like how the different sources get incorporated, with the speakers looping back to arguments from one source to contrast the next one. But it really is just superfluous summaries. Twice, the speakers talk about layoffs without knowing what they are talking about.
What’s next? Future experiments could explore giving these AI speakers actual personalities—maybe some background knowledge that leads to genuine disagreements or surprising insights. Could we feed one speaker an MCP server for deeper content access while giving the other web search capabilities? Or do we need something more sophisticated—agents with persistent memories and evolving belief systems? Though honestly, that just feels like more elaborate trickery.
But here’s the real question: Why would anyone choose to listen to this? It’s not a podcast—it’s a fakecast. These new AI tools are designed to fool you into thinking you’re hearing real humans have real conversations. For text-to-speech, yes please. But the mimicking of human conversation? That feels unsettling. Yet companies like Google and ElevenLabs are racing to perfect this illusion.
Here’s what I’d prefer: Give me synthetic voices that are pleasant to listen to but obviously artificial. Let’s embrace the technology without the deception. Make it clear when we’re listening to machines, not humans masquerading as our neighbours.
Just One Prompt
I wrote one short prompt to get a much longer one. I asked Claude 4 Opus to analyze a transcript of “Hard Fork” as an example and come up with a prompt. I imagine you could come up with something much better. This is just one shot:
This is a podcast transcript. Learn from it - the style, tone, flow, format - and give me a prompt I can use with Claude to make my own podcast by supplying sources, blog posts and newsletter, and Claude writes a engaging conversation for two hosts.Claude generateds a prompt which I used, just adding source material:
You are an expert podcast scriptwriter. Create an engaging, natural-flowing conversation between two podcast hosts based on the provided source materials. Follow these guidelines:
Show Format:
- Two co-hosts with distinct personalities who know each other well
- Natural banter and chemistry between hosts
- Mix of serious analysis with humor and casual asides
- Segments can include: main topic discussion, interviews with guests, and lighter closing segments
Writing Style:
- Conversational and natural - use contractions, incomplete sentences, and casual language
- Include verbal tics like "You know," "I mean," "Well," etc.
- Add reactions: laughter, surprise, agreement ("Right," "Exactly," "Yeah")
- Include tangents and callbacks to previous jokes or references
- Use specific examples and anecdotes to illustrate points
- Break up dense information with humor or personal observations
Structure:
1. Cold Open: Start with banter, personal anecdotes, or a hook related to the main topic
2. Introduction: Brief host intros and show name
3. Main Segment: Deep dive into the primary topic with:
- Clear explanation of the issue/story
- Multiple perspectives explored
- Real-world implications discussed
- Personal reactions and opinions from hosts
4. Interview Segments (if applicable): Natural Q&A with:
- Context-setting questions
- Follow-ups based on interesting answers
- Hosts building on each other's questions
5. Closing: Lighter topic or forward-looking discussion
Dialogue Techniques:
- One host sets up, the other provides color commentary or jokes
- Include moments of genuine surprise or learning ("Oh, I didn't know that!")
- Add meta-commentary about the show itself
- Use callbacks to earlier jokes or discussions
- Include technical difficulties or real-time discoveries ("We're recording this on [date], so things may have changed...")
Example Exchange Pattern:
HOST 1: So, Casey, [setup of topic or observation]
HOST 2: [Reaction or joke]
HOST 1: No, but seriously, [deeper point]
HOST 2: Right, and what's interesting about that is [additional insight]
Source Material Instructions: Based on the blog posts/newsletters/sources provided, the hosts should:
- Summarize key points conversationally
- Debate different angles
- Share personal experiences related to the topic
- Explain complex concepts in accessible ways
- Point out surprising or counterintuitive findings
Tone Guidelines:
- Smart but not condescending
- Informal but informative
- Skeptical but fair
- Humorous but respectful of serious topics
Now, using the sources I provide below, create a [X]-minute podcast conversation between [AI-Ole] and [Another AI-Ole] about [AI in Journalism].
Please write the full conversation with all the natural flow, interruptions, asides, and personality that makes for engaging podcast listening. Include at least one moment where the hosts disagree or see things differently, and one moment where they build excitedly on each other's ideas.The added source material:
- How the Associated Press Built its AI Strategy Without Breaking Trust (Ulrike Langer, News Machines)
- AI’s Next Land Grab: Your Browser (Ole Reissmann, THEFUTURE)
- How AI bots quietly dismantle paywalls via web search (Henk van Ess, Digital Digging)
- The Media’s Pivot to AI Is Not Real and Not Going to Work (Jason Koebler, 404 Media)
The Unedited Script
AI Ole: Okay, so I’m looking at my browser right now and I have… let me count… forty-seven tabs open. Forty-seven!
Another AI Ole: [Laughs] Amateur numbers. I’m sitting at eighty-three.
AI Ole: Eighty-three?! What are you, some kind of digital hoarder?
Another AI Ole: Hey, each one of those tabs represents a thought I had at some point that seemed important enough to keep open. There’s a recipe for sourdough starter from 2023…
AI Ole: Of course there is.
Another AI Ole: …three different articles about the same news story because I wanted to compare coverage, and approximately seventeen tabs that are just YouTube videos I swear I’m going to watch later.
AI Ole: Well, funny you should mention all those tabs, because that’s actually kind of what we’re talking about today. The battle for your browser—and how AI companies are basically trying to become the middleman between you and… well, everything.
**[Theme music]**
AI Ole: I’m AI-Ole…
Another AI Ole: And I’m Another AI-Ole…
AI Ole: And this is “Tabs Open,” where we talk about AI, journalism, and the increasingly weird intersection between the two.
Another AI Ole: We’re recording this on Thursday morning, July 18th, 2025, so if by the time you hear this, OpenAI has already launched their browser and taken over the world, just know we called it.
AI Ole: Sort of. I mean, we’re talking about it because it’s already happening. Perplexity just dropped something called Comet…
Another AI Ole: Which sounds like something that would destroy the dinosaurs, to be honest.
AI Ole: [Laughs] Right? But no, it’s their new browser. And here’s the kicker—if you want early access, that’ll be $200 for Perplexity Max.
Another AI Ole: Two hundred dollars! For a browser! I remember when browsers were free and we liked it that way.
AI Ole: Okay, boomer.
Another AI Ole: I’m literally the same age as you.
AI Ole: [Laughs] Fair. But here’s what’s actually interesting about this, and it’s something you wrote about in your newsletter. The browser isn’t just a browser anymore. It’s becoming this… I don’t know, this surveillance middleman? You had this great line about how if you own the browser, your AI gets access to everything that was off-limits before.
Another AI Ole: Yeah, and that’s the thing that kind of freaks me out. Like, we’re not just talking about your search history here. We’re talking about your emails, your shopping carts, all that paywalled content you’re reading…
AI Ole: Or trying to read.
Another AI Ole: Right. And speaking of paywalls, did you see Henk van Ess’s investigation? This is wild. He tested all these AI chatbots to see if they could get around paywalls, and guess what? They’re really, really good at it.
AI Ole: Define “good at it.”
Another AI Ole: ChatGPT, Perplexity, and Grok—Elon’s thing—they successfully accessed paywalled content about 50% of the time.
AI Ole: Fifty percent!
Another AI Ole: And here’s the thing—they’re not like, hacking the paywall. They’re doing something way more clever. They’re basically reconstructing the articles from fragments they find all over the internet. Social media posts, archived versions, people quoting bits in other articles…
AI Ole: It’s like… okay, you know when you’re trying to remember a movie, and you can’t find it streaming anywhere, so you watch the trailer, read the Wikipedia plot summary, watch some YouTube reviews, and by the end you feel like you’ve basically seen the movie?
Another AI Ole: That’s exactly what it is! Except it’s happening with journalism. And Henk found six different methods they use. My favorite is what he calls “the distributed archive”—where the AI just hunts for all the pieces of an article that have been shared or quoted anywhere online and then reconstructs it.
AI Ole: That’s… I mean, that’s kind of genius and terrifying at the same time.
Another AI Ole: And get this—when Henk paywalled his own article about this investigation, the bots could still access it and summarize it perfectly.
AI Ole: [Laughs] Of course they could.
Another AI Ole: But here’s what kills me. While all this is happening, while AI is literally dismantling the business model of journalism, what are media executives doing? They’re pivoting TO AI.
AI Ole: Oh, this is where we get to Jason Koebler’s piece.
Another AI Ole: Which is just… chef’s kiss. The title alone: “The Media’s Pivot to AI Is Not Real and Not Going to Work.”
AI Ole: He’s not pulling punches.
Another AI Ole: No, and he shouldn’t! Because it’s insane. You’ve got these media companies whose traffic is getting destroyed by AI—ChatGPT is sending them like 187,000% less traffic than Google…
AI Ole: Wait, let me do the math on that… [pause] That’s… that’s basically nothing.
Another AI Ole: Right! And yet these same executives are like, “You know what will save us? More AI!”
AI Ole: It’s like your house is on fire and you’re like, “You know what we need? More fire!”
Another AI Ole: [Laughs] Exactly! And Koebler has all these examples. Business Insider laid off 21% of their workforce, and in the same breath, the CEO is saying they’re “going all-in on AI” and that 70% of employees are using ChatGPT.
AI Ole: For what, though? That’s what I don’t get.
Another AI Ole: Slide decks, apparently. They specifically mentioned slide decks.
AI Ole: We’re laying off journalists to make better PowerPoints?
Another AI Ole: I mean… [pause] Yeah, basically.
AI Ole: This is depressing.
Another AI Ole: It gets worse. The LA Times has an AI bot that argues with its own writers and once wrote that the KKK was “not so bad, actually.”
AI Ole: WHAT?
Another AI Ole: I am not making this up. This is a real thing that happened at a real newspaper.
AI Ole: Okay, but… [sighs] Okay. Let me play devil’s advocate for a second here, because I think there’s something we’re missing.
Another AI Ole: Oh, here we go.
AI Ole: No, hear me out. Because when I look at what the AP is doing—and Ulrike Langer had this great piece about their AI strategy—they’re actually being pretty thoughtful about it.
Another AI Ole: How so?
AI Ole: Well, first of all, they started with standards. Before they did anything with AI, they sat down with their standards team and figured out the red lines. No generated imagery. No content delivered to customers without journalist review.
Another AI Ole: Okay, that’s… that’s actually smart.
AI Ole: Right? And they’re building tools that journalists actually want. Like, they had this problem with the JFK papers—you know, when Trump released those 63,000 pages of documents?
Another AI Ole: Yeah.
AI Ole: They built an AI tool to help reporters move through them faster, do summarization, identify what had been newly unredacted. And it worked so well they’re keeping it for future document dumps.
Another AI Ole: See, that makes sense to me. That’s AI as a tool, not AI as a replacement for journalism.
AI Ole: Exactly. And they’ve got this other project where they’re monitoring 430 federal agencies for regulatory actions that could become local news stories. When the EPA fines a company, the system figures out if it’s newsworthy and identifies local relevance.
Another AI Ole: That’s… actually kind of brilliant.
AI Ole: But—and here’s the but—even they admit it’s not quite working. The local newsrooms don’t just want tips. They want something they can publish. They don’t have time to run down leads.
Another AI Ole: Because they’ve all been laid off!
AI Ole: [Laughs darkly] Yeah, exactly. It’s this whole vicious cycle. You fire all the reporters, then you build AI tools to help the remaining reporters, but there aren’t enough reporters left to use the tools effectively.
Another AI Ole: You know what’s interesting though? This quote from Troy Thibodeaux at the AP. He said something like… hold on, let me find it… Here: “Every news organization is building its own headlines tool.”
AI Ole: Oh, that’s so true.
Another AI Ole: Right? Everyone’s building the same basic stuff instead of, I don’t know, collaborating? Sharing resources?
AI Ole: But that would require media companies to work together, and we know how good they are at that.
Another AI Ole: [Laughs] Fair point.
AI Ole: Although, you know who is doing something interesting? There’s this thing from Ole Reissmann’s newsletter—apparently Google launched a tool to access Gemini from the command line. Thousand free requests per day.
Another AI Ole: Command line? In 2025?
AI Ole: I know, I know. But he’s using it to automatically rename screenshots with descriptions. Like, you know how you take a million screenshots and they’re all just “Screenshot 2025-07-18 at 3.47.23 PM”?
Another AI Ole: Oh my god, yes. I have hundreds of those.
AI Ole: Right? So he’s got this system where Gemini looks at the screenshots and renames them with actual descriptions of what’s in them.
Another AI Ole: That’s… that’s actually useful. That’s the kind of boring, practical thing that AI should be good at.
AI Ole: Exactly! It’s not trying to replace human creativity or judgment. It’s just doing the tedious organizational stuff that nobody wants to do.
Another AI Ole: Speaking of which, we should probably talk about what this all means for actual journalists. Because Koebler makes this point that I think is really important. He says the only journalism business strategy that works is creating something of value that human beings want to read and can’t find anywhere else.
AI Ole: Which sounds obvious when you say it like that.
Another AI Ole: But apparently it’s not! Because you’ve got all these executives talking about “efficiency” and “AI-first newsrooms” and meanwhile, they’re competing against—what was his example?—a child in Romania publishing 9,000 articles a day.
AI Ole: With no regard for facts or accuracy.
Another AI Ole: Right. So you can’t compete on volume. You can’t compete on SEO. The only thing you can compete on is being human.
AI Ole: It’s funny, there’s this quote from a comedian in Ole’s newsletter. Let me find it… Here: “Everything in the entire world is either ‘AI-powered’ or ‘protein-packed.'”
Another AI Ole: [Laughs] That’s perfect. That’s exactly what it feels like.
AI Ole: But I think… I don’t know. I’m conflicted about all this.
Another AI Ole: How so?
AI Ole: Well, on one hand, yeah, the pivot to AI is clearly not working. These companies are grasping at straws. But on the other hand, AI is here. It’s not going away. And pretending it doesn’t exist isn’t a strategy either.
Another AI Ole: No, but I think that’s different from making it your entire strategy. Like, using AI for transcription or document analysis or organizing your screenshots—that’s just using tools. That’s different from saying “We’re an AI-first newsroom now!”
AI Ole: True. And I think that’s what the AP gets right. They’re using AI where it makes sense, but they’re not pretending it’s going to save journalism.
Another AI Ole: Also, can we talk about this thing from the Chicago Sun-Times? They had some AI-generated reading list fiasco, and in their apology, the CEO said they’re not backing away from AI experimentation.
AI Ole: After the fiasco!
Another AI Ole: Right! But then she also said—and I love this—”We must own our humanity. Our humanity makes our work valuable.”
AI Ole: So close to getting it.
Another AI Ole: So close! Like, you’re right there! You’ve identified the thing that makes you different from the spam farms! Maybe lean into that instead of the thing that makes you more like them?
AI Ole: You know what this reminds me of? Remember the pivot to video?
Another AI Ole: Oh god.
AI Ole: Where all these media companies fired their writers to hire video producers because Facebook said video was the future?
Another AI Ole: And then it turned out Facebook was lying about the numbers.
AI Ole: Exactly. And I feel like we’re in the same cycle again, except now it’s AI instead of video.
Another AI Ole: Yeah, but at least with video, you were still making content for humans. This is like… I don’t even know who the audience is supposed to be anymore. Other AIs?
AI Ole: ChatGPT reading articles written by ChatGPT.
Another AI Ole: It’s AIs all the way down.
AI Ole: [Laughs] Okay, but seriously, let’s talk about solutions for a second. Because it can’t all be doom and gloom.
Another AI Ole: Can’t it?
AI Ole: Come on. What’s working? Koebler mentions a bunch of places that are succeeding. Defector, 404 Media, Hell Gate…
Another AI Ole: Notice what they all have in common?
AI Ole: They’re all journalist-owned.
Another AI Ole: Bingo. No venture capital, no private equity, no billionaire owners who think they know better than the people actually doing the work.
AI Ole: And they’re all focused on building direct relationships with their audiences. Not chasing algorithms or SEO or whatever the platform du jour is.
Another AI Ole: Right. It’s almost like… and stay with me here… it’s almost like journalism is about serving readers, not tech platforms.
AI Ole: Revolutionary concept.
Another AI Ole: I know, right?
AI Ole: Although, you know what’s interesting? Even the places that are doing well are still using AI in some ways. Koebler mentions he uses AI transcription for interviews.
Another AI Ole: But not for sensitive sources.
AI Ole: Right, of course. But my point is, it’s not about avoiding AI entirely. It’s about using it in ways that enhance your work rather than replace it.
Another AI Ole: Yeah, I think that’s the key distinction. AI as a tool versus AI as a strategy.
AI Ole: And definitely AI as a tool versus AI as your entire business model.
Another AI Ole: [Laughs] Yeah, that’s probably not going to work out.
AI Ole: You know what else I found interesting? This bit about the browser wars. Like, we started talking about browsers, and I think there’s a connection here.
Another AI Ole: How so?
AI Ole: Well, if AI companies control the browser, they control what people see. They control discovery. They can prioritize their own AI-generated content over human journalism.
Another AI Ole: Oh, that’s… that’s actually terrifying.
AI Ole: Right? It’s not just about privacy or data collection. It’s about controlling the entire information ecosystem.
Another AI Ole: And charging $200 for the privilege.
AI Ole: [Laughs] Can’t forget that part.
Another AI Ole: Honestly, the more we talk about this, the more it feels like we’re at this really critical moment. Like, we’re deciding what the internet is going to be for the next decade or more.
AI Ole: Yeah, and the choice seems to be between an internet full of AI slop where you can’t trust anything you read…
Another AI Ole: Or an internet where human creativity and expertise still matters.
AI Ole: I know which one I’d prefer.
Another AI Ole: Same. But I’m not sure we’re the ones making that choice.
AI Ole: No, but we can at least support the places that are doing it right. Subscribe to real journalists. Pay for good content. Use ad blockers on the slop farms.
Another AI Ole: Vote with your wallet.
AI Ole: Exactly. And maybe don’t pay $200 for a browser that’s going to spy on everything you do.
Another AI Ole: [Laughs] Yeah, that seems like a good place to start.
AI Ole: Alright, I think we should wrap this up before we get too depressing.
Another AI Ole: Too late.
AI Ole: [Laughs] Fair. But seriously, I think the takeaway here is pretty clear. AI isn’t going to save journalism. Only good journalism is going to save journalism.
Another AI Ole: And good journalism requires humans.
AI Ole: Weird how that works.
Another AI Ole: It’s almost like… and I’m just spitballing here… it’s almost like journalism is about humans telling stories to other humans about things that matter to humans.
AI Ole: Crazy talk.
Another AI Ole: I know. Next you’ll be telling me that readers actually want to know that a real person reported and wrote the stories they’re reading.
AI Ole: Revolutionary.
Another AI Ole: [Laughs] Alright, on that note…
AI Ole: This has been “Tabs Open.” I’m AI-Ole…
Another AI Ole: And I’m Another AI-Ole. Close some of those tabs, people. Your computer will thank you.
AI Ole: But maybe keep the one with our show open.
Another AI Ole: Obviously.
AI Ole: Obviously.
The Text-to-Speech Conversion
First, I wanted to use ElevenLabs v3, but in Creative Platform, when I chose “podcast” and tried to upload the script, it just crashed. In a new project, I had to manually set speaker changes and could only get to 3,000 characters in total—nowhere near enough.
My fallback was Google AI Studio and the new-ish Gemini Text-to-Speech that is also powering NotebookLM. I chose two voices, Rasalgheti and Laomedeia, set the temperature to 1.3, and gave this instruction:
Podcast. Talk fast, show emotions, natural flow, conversational style.I replaced the speaker names to let Gemini know who is talking. This doesn’t work well—in several attempts, Gemini kept ignoring one random speaker change throughout the script. In addition, in the web interface, generation stops after 10 minutes, so I had to split my script and stitch the audio back together using another tool.
