NotebookLM is Google’s research and writing assistant tool that uses AI to simplify some otherwise long-winded or difficult tasks such as summaries, analysis and organizing information from disparate sources. In an exclusive interview, Steven Johnson, the co-founder and Editorial Director of NotebookLM, told me what’s coming next.
Johnson is an accomplished communicator, relaxed, affable and straightforward, fielding questions with ease and full, open answers.
“I’ve always been interested in using software to help me with the creative process and with research. I wrote a book called Where Good Ideas Come From that’s about innovation and creativity,” Johnson tells me in Google’s central London offices. Then, in spring 2022, he wrote a feature for the New York Times Magazine saying computers are mastering language. This was six months before ChatGPT.” The piece was controversial and suggested that Johnson had fallen for the hype, which couldn’t be matched by reality.
It caught the eye of several people at Google Labs, who invited Johnson to work with them, though it wasn’t clear to Johnson how successful or long-lasting the project would be. “Here we are two-and-a-half years later, NotebookLM is out in 180 countries and our team is around 50 people now.”
From the start, the purpose was clear: “It was always predicated on this idea that as powerful as AI was, we wanted always to remember the importance of the original documents and ideas, to foreground those and figure out how to use AI to understand whatever material you’re working on, but not replace that original material altogether. That was kind of the philosophy kind of from the beginning.”
If you haven’t used NotebookLM, it works like this: you create a notebook in a given project, and then you populate it with the sources that you’re using. The sources are the core documents that are central to the project. So, if you are working on a, nonfiction work of history, it’s all the primary sources you’re working with, interviews your notes and so on. “When you load them into a given notebook, at that point, the AI, which is Gemini 2.0,” Johnson says. “Once you’ve loaded these sources in, the AI is grounded in the documents you’ve given it. It has to stick factually to the information you brought into the notebook, and it becomes like an expert in the sources you’ve shared.” This has a key benefit when it comes to AI’s big flaw: hallucinations. By restricting the AI to the sources, hallucinations are significantly reduced. “The kind of wanton hallucination, I think most models are now much better at avoiding,” Johnson says. “But that was a huge part of why we did the source grounding approach for notebooks. It’s so rare with NotebookLM to see a full-on hallucination that I often feel like that’s the wrong word for it now. When it gets something wrong, it’s often that it’s just there’s some ambiguity in the underlying sources, and it kind of interpreted it wrong. And it made a reasonable mistake, but it didn’t just invent a fiction. And it’s quite good at saying what it doesn’t know.”
Now, new features are coming to NotebookLM. Now, you can use the new Discover Sources feature. Along with uploading your own sources, you can define the topic you’re exploring and the program will assemble an extensive list of potential sources, narrowing these down to 10 or fewer with annotated summaries, choosing them from places like Wikipedia.
You can generate briefing documents of FAQs from these and, as ever, read them in the original form and ask questions using the chat. Students are likely to find this useful for focusing on the right resources as a first step for studying effectively.
At all times, the program has citations so you can quickly see where the information is coming from and jump to the right place in the source instantly.
“Sometimes people don’t have any sources,” Johnson explains. “They say, ‘I don’t even know what a source is,’ like there’s a zero state, a fresh start kind of problem with it. One of the things we want to make it really easy for people to do is to find the sources they need to work on whatever project they’re working on. And fortunately, this is a Google product, so we actually are quite good at finding things.”
There’s also a new feature called I’m Feeling Curious, which generates a collection of sources on a random topic. At the least, this can demonstrate how powerful the source discovery agent is, Google says.
As before, you can do cool creative things with NotebookLM, such as instantly creating a chronology of events based on the sources it’s looking at, or, as Johnson describes it, “generate a completely AI-based simulated podcast style conversation between two remarkably lifelike people discussing whatever information you’ve given it.”
There’s also another new feature, Mind Maps. “This will take all your source material and generate a visual concept map, basically, of all the information that’s in the material that is interactive.” Once you’ve created the map you can click on items to reveal sub-categories and use them to create questions in a chat.
“I’ve seen a lot of mind maps,” Johnson says, “and I’ve always been a little skeptical, but the way this is designed is very clever and being one click away from the actual source text to be able to read it is super powerful.”
The strengths of NotebookLM are enhanced by the new features, while its ability to interrogate content, to highlight inconsistencies between sources and its capability to inspire creativity remain intact.
