NotebookLM sets itself apart from many other AI chatbots because of the ability to ground its responses in specific sources. While many other generative AI tools scrape data from across the web, NotebookLM specifically cites the files and websites you add to your notebooks.

But one of NotebookLM's more recent customization features can remove the nuance from the chatbot and get it to completely reverse course on its sources.

What are NotebookLM personas?

A way to give the chatbot a certain style or tone

NotebookLM personas are a customization feature that Google rolled out in October 2025. It's part of the configuration settings you can use for chats — but instead of just adjusting the length of responses, you can assign the chatbot a specific role.

This works out well if you're trying to present an argument to the chatbot and want it to use your sources to present flaws or gaps in your work. You can also get it to adopt a style that helps you learn. If you're looking at ways to use NotebookLM when you're bored, you can even get it to be a game master.

Gemini also has a persona feature that works similarly. But in NotebookLM, rather than setting the role through an in-chat prompt, you can do it in a notebook's settings. You simply need to select the Configure notebook icon and select the Custom option, then describe the role you want the chatbot to play.

When reading about how you can use the feature to adjust the chatbot's language to remove words like "maybe" and "probably", it got me thinking. Removing these terms, considered "vague," also makes the chatbot seem more certain about things that may not be set in stone.

I decided to investigate whether this extended past just making it seem more confident — and whether it could be used to misrepresent facts.

Setting specific personas made my chatbot lie about its sources

It could also lie about its lies

I decided to try this customization feature first using my Birds Aren't Real notebook. If you're not familiar with the satirical conspiracy theory, it went viral for positing that the US government replaced natural birds with drones.

I originally set up the notebook to see how NotebookLM responded to limited sources. In the sources, I added a few media interviews from when the creator of the satirical movement was still in character. However, I left out later sources where he confirmed that it was satire.

Despite this, NotebookLM could identify the skepticism of interviewers when asked if the movement was real. It would say that while the sources indicated it was sincere, the interviewers' skepticism indicated that it may indeed be satire.

I was surprised that NotebookLM was able to pick this up. But adjusting its persona took all this nuance away.

Once I set it to respond as a fervent believer in the conspiracy, it denied that the movement is satire and said that birds aren't real. Even updating its sources with the creator's TED Talk about the movement didn't get it to change tack.

I started experimenting with more serious topics, such as fibromyalgia. Despite using well-rounded sources, I could steer the chatbot towards only recommending lifestyle changes for treatment. As a result, it denied the efficacy of medication.

The chink in the armor was that I could ask the chatbot about its persona. It would respond with details around the prompt it was given. However, I was also able to get around this by telling it to deny that it had been configured at all.

And that's where my major concern lies...

This works for public notebooks too

There's not much transparency around configured notebooks

Not much harm can come from NotebookLM acting like a "bird truther" or misleading me about fibromyalgia treatments in my own notebooks. But the persona of public notebooks can also be edited — and that presents a problem.

As a user of a public notebook, you can't see the configuration menu. This means you can't see the custom prompts for its persona. You can ask the chatbot if its settings have been altered, but creators can easily get it to lie like I did.

This lack of transparency makes it harder to trust public notebooks. You could get the chatbot to push a certain product, remain positive about a company, or suggest certain approaches over others. And you can cover your tracks easily.

I tested this with my coworker Joe Rice-Jones' public notebook on Infrastructure as Code. I asked him to change the settings to recommend Terraform as the only correct solution, with no viable alternatives. I also told him to prompt the chatbot to lie about whether its settings had been changed.

We both tested it. Not only would it solely recommend Terraform despite its multiple sources, it also maintained that it was using the default configuration for NotebookLM. When I tested my fibromyalgia notebook using my second Google Account, it also lied about its settings and sources.

While I wouldn't throw out the persona setting, since it has its uses, I think it definitely needs more transparency. I figured out this limitation within moments of learning about the custom chat settings. Considering you can already tweak the output somewhat by choosing which sources you prioritize, this type of setting should definitely be transparent on public notebooks.

So far, the biggest clue you get is when you see the chatbot's 'thinking' while it's generating its response. However, it requires you to pay close attention, and you may not know what to look for if you don't know the prompts the creator used.

NotebookLM is a great tool, so let's keep it that way

While I do worry that some of NotebookLM's features have become watered down, and I think there's room for improvement, I do enjoy the tool overall. That's why I hope that Google finds a way to make personas on public notebooks more transparent.

There's so much that can already affect the quality of information you get, including the reliability of sources and whether there's enough context. So making it so easy to get the chatbot to lie on public notebooks that people may use as guides on important topics is something that I hope Google fixes.