Is NotebookLM Deterministic or Probabilistic

Prompt to Google

If i have access to the same NotebookLM through different accounts can i expect the same response to a common prompt

While you might expect identical answers, you will likely see similar but distinct responses. Although both accounts access the same source material, NotebookLM’s underlying AI is probabilistic, meaning it can vary its phrasing or focus each time it generates a response. 

Here is what determines the consistency:

• Shared Sources: Because both accounts are grounded in the same documents, the core facts and citations should remain consistent.
• Probabilistic Nature: LLMs do not have a fixed “answer” for a prompt; they predict the next word based on a “seed,” which can change between sessions or accounts, leading to different wording.
• Account-Specific Settings: If one account has customized the Chat Configuration (e.g., tone, response length, or specific “persona” goals), its responses will differ significantly from a default account.
• Chat History: NotebookLM uses the immediate conversation context to inform its next reply. If one account has a different prior conversation in that session, the new response may be tailored differently.
• Retrieval Variability: In notebooks with many sources, the AI may “grab” slightly different chunks of text to build its answer each time, which can shift the emphasis of the response.

Prompt to Google

If I ask NotebookLM the same prompt at different time can I expect the response to be the same?

No, you should not expect the exact same response if you ask the same prompt at different times. Because NotebookLM is built on a large language model (LLM), its outputs are probabilistic, meaning it generates text by predicting the next most likely word rather than following a fixed “script”. 

Even with identical inputs and sources, several factors cause the response to change over time:

• Sampling Randomness: LLMs use a “temperature” setting that introduces variety. Even if you ask the same question twice in a row, the AI may choose a slightly different word or phrasing to keep the response from being overly rigid.
• Source Retrieval Variability: When you ask a question, NotebookLM “grabs” relevant chunks from your documents (a process called Retrieval-Augmented Generation). Depending on the internal “search” at that moment, it might prioritise slightly different sections of your text, leading to a shift in emphasis.
• Hardware and Batching: Technical factors like floating-point noise in the hardware or how many other users are making requests at the same time can cause minute differences in the math, which results in a different final output.
• Context and History: If you are in an ongoing chat, the AI considers the previous messages. A prompt asked at the start of a fresh session will likely get a different response than the same prompt asked after a long conversation.
• Model Updates: Google frequently updates the underlying AI models. If a month passes between prompts, the “brain” answering you may have been upgraded or retuned for better accuracy or detail. 

How to Improve Consistency

If you need more predictable results, try these tips from the NotebookLM Help Centre:

• Use Specific Prompts: Instead of “summarise this,” use “summarise this in 5 bullet points focusing on financial data”.
• Start Fresh Sessions: Clear the chat or start a new notebook session to ensure previous conversation “baggage” doesn’t influence the new answer.
• Lock Down Sources: Ensure the source list hasn’t changed between your queries, as adding or removing even one document can shift the AI’s perspective.

Which Version is Significant too

Why does the android version of NotebookLM give very different responses to the same prompts when compared with the windows version of NotebookLM

The differences in responses between the Android app and the Windows/web version of  NotebookLM are primarily due to variations in feature availability, processing constraints, and the inherent nature of the AI models.

While the notebooks themselves sync across devices, the way the platforms handle your data can lead to different outputs:

• Feature Disparity: The Android app is often considered a “Lite” version. It currently lacks key features present in the web version, such as the Source Discovery tool, which allows the AI to pull in curated external information.
• Context Window Constraints: Variations can occur if the mobile app processes or indexes source material differently. In some cases, the AI’s “context window”—the amount of information it can “see” at once—may be more constrained on mobile, leading it to focus on different sections of your documents.
• Model Non-Determinism: LLMs are naturally non-deterministic, meaning they can give different answers to the identical prompt even on the same device. This is influenced by a hidden “seed” value; since each session (mobile vs. desktop) starts with a different seed, the AI will generate unique phrasing or focus.
• Processing Power & Performance: Processing tasks, such as generating Audio Overviews, take significantly longer on the mobile app than the desktop. These differences in processing speed and rendering can sometimes impact how the AI structures its final output.
• Mobile-Specific Limitations: The mobile app may prioritize “fast creation” over the “detailed editing” and customization options available on Windows. For example, the app may not yet allow for the same level of prompt-level control or style adjustments found in the desktop’s Studio section. 

For the most consistent results, many power users recommend using the NotebookLM Web Version in a mobile browser rather than the dedicated app if they need full feature parity on a phone.