Audio Overview with NotebookLM
Fact-checking the VP debate
In previous posts, I’ve focused on ChatGPT. In this edition, I’ll introduce a new tool from Google: NotebookLM. While this tool offers many features, I'll highlight its standout capability: Audio Overview. This feature transforms any document into an engaging, in-depth audio discussion between two AI hosts.
For example, I take the recent vice presidential debate transcript, generate an Audio Overview, and critique the output. It’s an impressive state-of-the-art AI demo, but we’ll also see where it misses the mark and why we should approach AI cautiously and skeptically.
Example: The VP Debate
Click the New Notebook icon.
Adding Source Documents
The process is straightforward. You can add one or more “sources.” They can be web URLs, Google Docs, PDFs, YouTube videos, or plain text files. In practice, you can achieve excellent results with just one source.
The Notebook opens to the Add Sources dialog.
Click the “Website” button at the bottom center of the dialog.
Paste the URL for a CBS News transcript of the VP debate into the “Paste URL” textbox.
According to Google, your sources are not used to train NotebookLM and remain confidential.
Audio Overview utilizes Google's multimodal Gemini 1.5 Pro AI model. This model can comprehend and reference images, charts, and diagrams in your sources.
Generating the Audio Overview
Press the blue “Generate” button and wait a few minutes for the process to complete.
The only input was a URL to a page with the transcript. The "send to" share icon provides a link to share the audio, which can be made public with a private link.
Tailoring the Podcast
Other than adding sources, the other optional configuration is a text box giving instructions to the hosts. I generated with and without the custom instruction of:
fact-check the debaters and specifically highlight any false or exaggerated statements.
Evaluating the Output
Here’s an audio file of
Host 1: [Vance] goes all in on the whole Trump playbook, strong borders, deportations. He even claims that the Biden administration is separating families at the border.
At 4:10, the fact-checking Audio Overview gives a false explanation:
Host 1: Vance also said something really shocking. He claimed that Homeland Security has lost over 320,000 children, basically implying they were separated from their families at the border. That's a serious accusation. What's the truth there?
The 320,000 number
Host #1: “Get those BS detectors warmed up…”
Ironically, this is good advice for users of Audio Overview. I don’t trust AI fact-checking, and neither should you…
Cringe-worthy Moments
The witty banter of the hosts was, at times, cringe-worthy:
AI Host #1: Asks Vance about this idea of holding parents legally responsible for
Don’t trust AI; always review the outputs. The training datasets for the models contain toxic sewage scraped from the internet. The models are “intelligent,” but you must still think for yourself. The safe assumption is that any AI output is potentially racist and sexist.
The Future
Listening to today's somewhat realistic voices suggests a future where telling humans and computers apart will be challenging. I shared a personal Audio Overview with friends, and several of them mistook it for an actual podcast about me. The AI podcasts are intriguing yet creepy but still somewhere in the