by Larry Magid
This post first appeared in the Mercury News
If you’re looking to impress friends and family, perhaps you should send them a link to a podcast all about you. But what if no one has recorded such a podcast? No problem. Thanks to NotebookLM, an AI tool from Google, you can have a pair of professionally sounding podcast hosts extol your virtues for several minutes.
The AI generated hosts don’t have names, but they’re a male and female team who sound great and did an amazing job describing my attributes in ways that almost made me blush. In my case, I pasted in the web address of my online bio, but you can also load in PDFs, YouTube videos, audio files, Google Docs or Google Slides.
If you load a web address, the two fake hosts not only talk about what they learned from your bio, but they also check out the entire website. I learned this by giving it links to two almost identical bios. The one hosted on Larrysworld.com, my personal site, included references to many of my newspaper articles. The one at ConnectSafely.org focused mainly on my and the organization’s online safety work.
The fake hosts brought some humor into the conversation, kind of making fun of some of the things the AI learned about me from my website. They also went into some tangents that made them sound like live podcast hosts.
Comparison
To get a sense on how it might describe someone in a very different type of job, I downloaded a sample resume of an electrician. The two fake hosts said that they were basing their podcast on his resume. They provided a positive spin on his qualifications while talking a bit about their own struggles with electric problems.
Like my bio, this podcast was based on the material loaded, which, in both cases, was published by the person being described. These fake hosts apparently didn’t do lateral research to find out if the information was true.
Once you generate the podcast, you can download it as an MP3 file and share it with others. If you do, eventually tell the listeners that it’s generated by AI. Aside from ethics and honesty, the same two fake hosts are used for every podcast. They sound real, but eventually listeners will figure out that they’re not.
You can try NotebookLM for free at notebooklm.google.
Practical uses
Besides being a parlor trick, there are some practical uses for this tool. I loaded in a Google slide deck and heard the fake hosts give me an accurate summary. And even the bios – while obsequious, were true to what the AI tool gleaned from the source material, although that doesn’t mean that the source material was accurate.
I also pasted in the address of a PDF from the Pew Research website on a recent survey about the public’s opinion on President-elect Trump’s plans for the country. The AI-generated podcast, which went on for 14 minutes, was based on the facts gleaned from the survey report, but the fake hosts also added their analysis. They covered both positive and negative aspects of Trump’s characteristics, especially as perceived by the public, according to the survey results.
NotebookLM also provided written materials that were more in-depth and useful to anyone wishing to understand the the survey. This included a 115-word summary and AI-generated pages on key topics, post-election “views on Trump and the nation,” “public confidence in a smooth transition,” among other topics.
Based on the PDF, NotebookLM also allowed me to create a FAQ, a study guide, a table of contents, a timeline and a briefing doc. The study guide asked questions about key portions of the document and provided answers based on the PDF. The timeline focused on the events leading up to and following the 2024 U.S. presidential election, based on the information in the document. The 740-word briefing doc covered key themes.
In case you’re wondering, it concluded that “Republicans welcome Trump’s victory with enthusiasm, Democrats view it with deep apprehension.”
Answers your questions
In addition to the material it provides, you can also ask questions about the document. What I like about this is that the answers to your questions are based on what is in the document, unlike many generative AIO tools that pull from the entire web. The fact that the information is very limited is both a bug and a feature. It’s a bug, because there is a good chance you won’t get an answer. It’s a feature, because what you get is based on the document or webpage you’re analyzing, so if you trust that source, you’re more likely to get an answer you can trust. Having said that, NotebookLM does interpret answers if it doesn’t know them, but it clearly discloses the difference between what it knows from the document and what it’s inferring from it.
Needs work
NotebookLM is in its infancy and needs work. For one thing, the user interface on both the web and a smartphone is a bit confusing. I had to click around a few times to find what I needed, including the ability to generate that podcast or find summary information. Also, like all generative AI products, it’s not infallible. There is a disclaimer on every page that “NotebookLM can make mistakes, so double-check it.” Despite its rough edges, this is truly a remarkable tool and a good example of how generative AI can extend beyond the initial offers from OpenAI, Microsoft and Google.
Disclosure: Larry Magid is CEO of ConnectSafely, a nonprofit internet safety organization that receives financial support from Google and other companies, including some of Google’s competitors.