Before we get into audio production — a quick gut check on podcasts and AI.
Before you can script a podcast, you need something worth saying. AI can dramatically speed up the research phase — but the tool you use matters.
Most AI tools are trained on the open web and can hallucinate facts confidently. NotebookLM is different: you upload your own source documents, and it only answers from those. That makes it far more reliable for research you're going to publish or broadcast.
Upload your sources — articles, PDFs, transcripts, or notes
Ask it questions and it answers only from what you gave it
Every answer is cited back to your source — no hallucinations
Use it to pull key quotes, summaries, and discussion questions
You upload specific documents — a research report, a transcript, a policy brief, news articles you've already verified. The AI synthesizes only from those materials. It can't make up facts because it doesn't have access to anything else. This is the right choice when accuracy is non-negotiable — journalism, PR, broadcasting.
Best for: Pulling key facts from a document you trust. Generating interview questions from a topic brief. Creating episode outlines from existing research. Summarizing long transcripts into usable notes.
These tools are trained on large amounts of internet text and can give you broad overviews, brainstorm angles, and suggest what to research. But they can also confidently invent statistics, misattribute quotes, or describe events that didn't happen. The output sounds authoritative even when it's wrong.
Best for: Brainstorming episode angles or interview questions. Understanding the general shape of a topic before you do real research. Getting ideas for what sources to look for.
Once you have solid source material, AI becomes a powerful organizer. Paste in your verified notes and ask it to help you build a narrative arc — an opening hook, the main argument, a counterpoint, a closing call-to-action. The structure is AI-assisted; the facts are yours.
Why this works: You're using AI for what it's genuinely good at — structure and flow — while keeping human-verified facts as the input. The AI organizes; you verify.
Upload your research documents and ask for any of these:
"Based on these articles, give me a 5-segment outline for a 15-minute podcast episode. Include a hook, main points, and a closing question."
"From this briefing document, write 8 interview questions that go from broad context to specific details. Include one follow-up for each."
"Pull the 10 most important facts from these documents. List them as bullets. Cite which document each fact came from."
"What are the three most surprising or counterintuitive findings in these sources? Write each as a one-sentence hook I could open a segment with."
A podcast script is not a news article read aloud. Writing for audio is a completely different skill — and AI needs specific instructions to do it well.
| Element | Written Copy (Web, Print) | Audio Script (Podcast, Radio) |
|---|---|---|
| Sentence length | Can be long and complex — reader can re-read | Short. One idea per sentence. Listeners can't rewind their attention. |
| Numbers & stats | "The study found a 34.7% increase over 18 months." | "The study found a jump of about a third — in under two years." |
| Signposting | Headers and subheads guide the reader | You have to say it out loud: "Let's talk about why that matters…" |
| Punctuation | Follows grammar rules | Written for breath and pause — dashes and ellipses are tools |
| Word choice | Formal vocabulary is fine | Use words you'd actually say in a conversation — nothing you'd stumble over |
| Transitions | Visual — a new paragraph or section break | Must be spoken: "Coming up next…" / "Here's what's interesting…" |
The single most important instruction you can give an AI when scripting for audio: tell it to write for the ear, not the page. Here's what that looks like in practice.
What you get: A dense paragraph written like a news article or Wikipedia entry. Long sentences. Formal vocabulary. No natural pauses. Nobody talks like that — and nobody wants to listen to it.
That sentence is 33 words before it says anything meaningful. On audio, your listener mentally tuned out at word 15.
What you get: Something that actually sounds like a human being talking. Shorter sentences. Natural rhythm. A hook at the top. Something a student could actually record in class without stumbling.
Professional audio scripts use a two-column format: production notes on the left, spoken words on the right. You can ask AI to generate this format directly.
This format makes it easy to hand a script to a host who's never seen it — they know exactly when to pause, when to emphasize, and what to say.
Let it extract key facts, surprising details, and potential discussion angles. Verify anything you plan to say on air.
Paste your verified notes. Ask for an episode outline with a hook, 2–3 segments, and a closing call-to-action.
Prompt for audio-ready copy. Read it aloud before you record — if you stumble, rewrite. Your mouth is the final editor.
This is where the lesson becomes your first real deliverable. Your team will use AI to research your angle, build an episode outline, and write a script — all in one class period.
Final 5 minutes: one team reads their intro aloud for the class.
Your Research Lead opens NotebookLM at notebooklm.google.com and creates a new notebook. Upload all four source files your team prepared before class. Then run the prompts below — one at a time — and share results with your team.
↑ Choose a prompt type to generate it.
Take your research results from Phase 1 and paste them into Claude or ChatGPT. Your Angle Editor enters your approved topic and the key facts your team selected. Use the prompt below to generate a timed episode outline.
↑ Choose a prompt type to generate it.
Your Script Writer takes the approved outline and uses the prompt below to generate an audio-ready intro script. The Audio Editor reads it aloud immediately. Anything that causes a stumble gets rewritten before it goes into the teleprompter.
↑ Choose a prompt type to generate it.
The Audio Editor reads the script aloud and checks every item before the Producer enters it into the teleprompter at susquweb.com/teleprompter.
If you stumble on a word or run out of breath, stop — that sentence needs to be rewritten, not rehearsed past.
One idea per sentence. Long sentences lose listeners. If it can be two sentences, make it two sentences.
"47.3%" → "nearly half." "18 months" → "a year and a half." Numbers that read fine on paper often stall on audio.
Find the moment in your intro where silence lands harder than words. Mark it with a dash —. Actually pause there when you record.
Delete "delve," "landscape," "tapestry," "vital," "revolutionary," and any sentence that starts with "In today's world." These are AI defaults, not your voice.
Your intro should make the listener want to keep listening. Close with a question, a tension, or a promise of what's coming — not a summary of what they just heard.
Key facts, opening hook, and research gaps from NotebookLM — saved in your shared Google Doc.
Timed segment structure with interview questions and a closing call-to-action — ready for the full episode script.
A 60-second, audio-edited intro — approved by your Audio Editor and loaded for recording.
AI voice technology can now produce audio that sounds indistinguishable from a real human host. What does that mean for podcasting — and for trust in audio media?
If a podcast uses an AI-generated voice to host an episode, should listeners be told?
A pop culture recap show uses an AI voice to narrate their weekly summary. The content is accurate. The voice is clear and pleasant. No disclosure is made.
A local news organization uses an AI voice to read their daily news briefing. The stories are reported by human journalists. The voice reads their scripts.
A sports media outlet uses an AI-generated voice that sounds like a known commentator — but isn't them — to produce highlight recaps. The commentator didn't consent.
You use AI to write your entire script, and then use an AI voice tool to narrate it so you don't have to record yourself. You submit it as your podcast.
Using AI to research your script is clearly assistance. Using AI to write every word and narrate every sentence — with no human voice or creative decision in the final product — is closer to full authorship. Where exactly does the line move for you? Does it matter what the context is — class assignment vs. professional publication?
When a listener trusts a voice, they're trusting more than words — they're trusting a person's credibility, judgment, and relationship with them over time. If that voice is AI, the trust is built on something that doesn't exist. Consider: who gets hurt when that trust is broken? The listener? The journalist? The industry? All three?
In journalism, there's a long tradition of transparency about process. Anonymous sources are disclosed. Corrections are published. Why would audio be exempt from that standard?
Written disclosure is relatively easy — you add a note. But audio disclosure is trickier. A verbal disclaimer at the top of an episode? A note in the show description? A standard industry label that streaming platforms display? Brainstorm what a realistic, honest, and listener-friendly disclosure standard would look like for audio media.
Key terms from today's session — AI research tools, audio production, and the vocabulary of podcasting.
A Google AI research tool that lets you upload your own source documents and ask questions about them. Unlike open-web AI tools, NotebookLM only answers from the documents you provide — and cites exactly where each answer came from. This source-grounding makes it significantly more reliable for professional research tasks where accuracy matters.
An AI system that is restricted to answering from a specific set of documents you provide, rather than from its general training data. Source-grounded AI dramatically reduces hallucination risk because the model can't generate information that isn't already in your sources. The tradeoff: it can only know what you've given it.
When an AI model generates information that sounds accurate but is factually incorrect or entirely invented — including fake statistics, fabricated quotes, or non-existent sources. Hallucinations are especially dangerous in broadcasting and journalism because the output reads as confident and authoritative. Every fact from an open-web AI tool requires independent verification before publication or broadcast.
The structural shape of a story or episode — how it begins, builds, and resolves. In podcasting, a strong narrative arc typically includes a hook (why should I keep listening?), a middle that builds tension or complexity, and an ending that either resolves the tension or leaves the listener with a compelling question. AI can help map this structure once you have your research in hand.
The practice of crafting audio scripts to be heard rather than read — using short sentences, conversational vocabulary, spoken-word transitions, and intentional pauses. Audio audiences can't re-read a confusing sentence or scroll back to a statistic they missed. Every scripting decision must account for the limits of the listening experience: one pass, real time, no rewind.
A standard professional audio script format with two columns: production notes on the left (music cues, sound effects, pauses, emphasis marks) and spoken dialogue on the right. The format separates what the listener hears from the technical instructions the producer follows. It's commonly used in radio, podcasting, and broadcast television scriptwriting.
Verbal cues that tell listeners where they are in a podcast and what's coming next — phrases like "Coming up next…", "Here's why that matters…", or "Before we get to that, let me explain…". In print, visual design (headers, section breaks) does this job. In audio, you have to say it out loud. Missing signposts are one of the most common reasons listeners disengage from amateur podcasts.
The written companion to a podcast episode — typically published on the podcast's website or in the app listing. Show notes include a summary of the episode, links to sources mentioned, timestamps for key moments, guest bios, and social media handles. AI is particularly useful for drafting show notes from a script or transcript, since the material already exists and just needs to be reformatted.
Technology that generates human-sounding speech from text, often trained on recordings of real voices. Modern AI voice tools can produce audio that is nearly indistinguishable from a real person. The ethical and legal questions around these tools — consent, disclosure, voice cloning without permission — are among the most active debates in media law and professional broadcasting standards today.
A rough, low-stakes recording made for the purpose of testing a script or timing an episode — not intended for final publication. Scratch tracks help you hear how a script actually sounds before you commit to a professional recording session. They catch problems with sentence rhythm, pacing, and word choice that aren't visible on the page.
The professional and ethical practice of informing your audience when AI tools were used in producing audio content — particularly when AI was used for scripting, voice narration, or research synthesis. There is no universal standard yet for audio AI disclosure, but best practice in journalism and broadcasting is to err toward transparency: if AI meaningfully shaped what your listener hears, they deserve to know.