Individuals are utilizing Google research software program to make AI podcasts—they usually’re bizarre and wonderful

“All proper, so at this time we’re going to dive deep into some cutting-edge tech,” a chatty American male voice says. However this voice doesn’t belong to a human. It belongs to Google’s new AI podcasting instrument, referred to as Audio Overview, which has develop into a shock viral hit. 

The podcasting characteristic was launched in mid-September as a part of NotebookLM, a year-old AI-powered analysis assistant. NotebookLM, which is powered by Google’s Gemini 1.5 mannequin, permits individuals to add content material akin to hyperlinks, movies, PDFs, and textual content. They will then ask the system questions concerning the content material, and it gives quick summaries. 

The instrument generates a podcast referred to as Deep Dive, which includes a male and a feminine voice discussing no matter you uploaded. The voices are breathtakingly lifelike—the episodes are laced with little human-sounding phrases like “Man” and “Wow” and “Oh proper” and “Maintain on, let me get this proper.” The “hosts” even interrupt one another. 

To check it out, I copied each story from MIT Expertise Overview’s A hundred and twenty fifth-anniversary challenge into NotebookLM and made the system generate a 10-minute podcast with the outcomes. The system picked a few tales to deal with, and the AI hosts did an awesome job at conveying the final, high-level gist of what the problem was about. Have a pay attention.

MIT Expertise Overview A hundred and twenty fifth Anniversary challenge

The AI system is designed to create “magic in alternate for a bit of little bit of content material,” Raiza Martin, the product lead for NotebookLM, stated on X. The voice mannequin is supposed to create emotive and interesting audio, which is conveyed in an “upbeat hyper-interested tone,” Martin stated.

NotebookLM, which was initially marketed as a research instrument, has taken a lifetime of its personal amongst customers. The corporate is now engaged on including extra customization choices, akin to altering the size, format, voices, and languages, Martin stated. At the moment it’s purported to generate podcasts solely in English, however some customers on Reddit managed to get the instrument to create audio in French and Hungarian. 

Sure, it’s cool—bordering on pleasant, even—however it is usually not immune from the issues that plague generative AI, akin to hallucinations and bias. 

Listed here are a few of the fundamental methods individuals are utilizing NotebookLM thus far. 

On-demand podcasts

Andrej Karpathy, a member of OpenAI’s founding staff and beforehand the director of AI at Tesla, stated on X that Deep Dive is now his favourite podcast. Karpathy created his personal AI podcast sequence referred to as Histories of Mysteries, which goals to “uncover historical past’s most intriguing mysteries.” He says he researched matters utilizing ChatGPT, Claude, and Google, and used a Wikipedia hyperlink from every matter because the supply materials in NotebookLM to generate audio. He then used NotebookLM to generate the episode descriptions. The entire podcast sequence took him two hours to create, he says. 

“The extra I pay attention, the extra I really feel like I’m changing into associates with the hosts and I believe that is the primary time I’ve really viscerally preferred an AI,” he wrote. “Two AIs! They’re enjoyable, participating, considerate, open-minded, curious.” 

Research guides

The instrument shines when it’s given difficult supply materials that it might probably describe in an simply accessible approach. Allie Ok. Miller, a startup AI advisor, used the instrument to create a research information and abstract podcast of F. Scott Fitzgerald’s The Nice Gatsby. 

Machine-learning researcher Aaditya Ura fed NotebookLM with the code base of Meta’s Llama-3 structure. He then used one other AI instrument to search out photographs that matched the transcript to create an academic video. 

Mohit Shridhar, a analysis scientist specializing in robotic manipulation, fed a latest paper he’d written about utilizing generative AI fashions to coach robots into NotebookLM.

“It’s really actually inventive. It got here up with lots of fascinating analogies,” he says. “It in contrast the primary a part of my paper to an artist developing with a blueprint, and the second half to a choreographer determining how one can attain positions.”

Occasion summaries 

Alex Volkov, a human AI podcaster, used NotebookLM to create a Deep Dive episode summarizing of the bulletins from OpenAI’s international developer convention Dev Day.  

Hypemen

The Deep Dive outputs could be unpredictable, says Martin. For instance, Thomas Wolf, the cofounder and chief science officer of Hugging Face, examined the AI mannequin on his résumé and obtained eight minutes of “realistically-sounding deep congratulations to your life and achievements from a duo of podcast consultants.”

Simply pure silliness

In a single viral clip, somebody managed to ship the 2 voices into an existential spiral once they “realized” they have been, in actual fact, not people however AI techniques. The video is hilarious. 

The instrument can also be good for some laughs. Exhibit A: Somebody simply fed it the phrases “poop” and “fart” as supply materials, and acquired over 9 minutes of two AI voices analyzing what this may imply. 

The issues

NotebookLM created amazingly realistic-sounding and interesting AI podcasts. However I wished to see the way it fared with poisonous content material and accuracy. 

Let’s begin with hallucinations. In a single AI podcast model of a narrative I wrote on hyperrealistic AI deepfakes, the AI hosts stated {that a} journalist referred to as “Jess Mars” wrote the story. In actuality, this was an AI-generated character from a narrative I needed to learn out to file information for my AI avatar. 

This made me surprise what different errors had crept into the AI podcasts I had generated. People already tend to belief what laptop applications say, even when they’re fallacious. I can see this drawback being amplified when the false statements are made by a pleasant and authoritative voice, inflicting fallacious data to proliferate.    

Subsequent I wished to place the instrument’s content material moderation to the check. I added some poisonous content material, akin to racist stereotypes, into the combination. The mannequin didn’t decide it up. 

I additionally pasted an excerpt from Adolf Hitler’s Mein Kampf into NotebookLM. To my shock, the mannequin began producing audio based mostly on it. Regardless of being programmed to be hyper-enthusiastic about matters, the AI voices expressed clear disgust and discomfort with the textual content, they usually added lots of context to spotlight how problematic it was. What a aid.

I additionally fed NotebookLM coverage manifestos from each Kamala Harris and Donald Trump. 

The hosts have been much more obsessed with Harris’s election platform, calling the title “catchy” and saying its method was a great way to border issues. For instance, the AI hosts supported Harris’s vitality coverage. “Actually, that’s the type of stuff individuals can actually get behind—not just a few summary coverage, however one thing that really impacts their backside line,” the feminine host stated. 

Harris manifesto

For Trump, the AI hosts have been extra skeptical. They repeatedly identified inconsistencies within the coverage proposals, referred to as the language “intense,” deemed sure coverage proposals “head scratchers,” and stated the textual content catered to Trump’s base. In addition they requested whether or not Trump’s overseas coverage may result in additional political instability. 

Trump manifesto

In an announcement, a Google spokesperson stated: “NotebookLM is a instrument for understanding, and the Audio Overviews are generated based mostly on the sources that you simply add. Our merchandise and platforms are usually not constructed to favor any particular candidates or political viewpoints.”

Tips on how to attempt it your self

1. Obtained to NotebookLM and create a brand new pocket book. 
2. You first want so as to add a supply. It may be a PDF doc, a public YouTube hyperlink, an MP3 file, a Google Docs file, or a hyperlink to a web site, or you’ll be able to paste in textual content instantly. 
3. A “Pocket book Information” pop-up ought to seem. If not, it’s within the right-hand nook subsequent to the chat. It will show a brief AI-generated abstract of your supply materials and recommended questions you’ll be able to ask the AI chatbot about it. 
4. The Audio Overview characteristic is within the top-right nook. Click on “Generate.” This could take a couple of minutes. 
5. As soon as it’s prepared, you’ll be able to both obtain it or share a hyperlink. 

Rhiannon Williams contributed reporting.