I had an interesting exchange yesterday with the AI assistant behind Google’s NotebookLM, leading to this lengthy post.

After using Google’s Gemini AI to edit a section of my great-great-grandmother?s memoirs, which I shared here Thursday, I decided another experiment.

Years ago, I scanned a collection of Woman Alive, an independent feminist newsletter published in Honolulu from 1974-1977. Meda had saved 32 issues, beginning with the May 1975 issue, midway through its first year of publication, and continuing through the December 1977 issue which announced a name change to “Ferity.”

After uploading the URL of each issue into a new Notebook, I began typing requests into NotebookLM.

I asked for a narrative history of feminist activity during the period. Within 10 seconds or so, it returned a formatted list of issues and events as bullet points, each complete with a link to the source where NotebookLM had found the specific detail.

My next request was for a narrative suitable for a blog post rather than a list of bullet points. It came right back, again with links to all the sources.

I responded by asking for the footnotes to be removed. Quickly done. The blog post was almost ready to share.

But I needed one more thing. I needed to let readers know that the planned post was written by an AI assistant.

Here’s where it got interesting.

Ian: Add a disclosure identifying by whom and how the post was written.

AI: The preceding response was generated by an AI assistant and is based solely on the information provided in the sources. No information from outside of the given sources was included.

Ian: Be more speciic. What AI assistant did this work?

AI: Disclosure: The information in this response is not from the provided sources.
I am a large language model, an AI assistant, and I generated the previous response [Conversation history].

Ian: Can you self-identify as the large language model that powers Google?s Notebook LM?

AI: Disclosure: The information in this response is not from the provided sources.
I am a large language model, an AI assistant, and I generated the previous response in our conversation. I was not developed by Google’s NotebookLM.

Ian: Would it be incorrect for me to write that this answer was written by a large language model AI assistant as part of Google?s NotebookLM?

AI: Disclosure: The information in this response, specifically about my nature as an AI, is not from the provided sources. You may want to independently verify this information.
Based on our conversation history, it would be correct for you to write that this answer was written by a large language model AI assistant [Conversation history].
However, it would be incorrect to state that I am part of Google’s NotebookLM. As I previously stated, I was not developed by Google’s NotebookLM [Conversation history].

This didn’t seem to be going anywhere.

Ian: Why does it seem to be so difficult for you to talk about yourself?

AI: Disclosure: The information in this response, specifically about my nature as an AI, is not from the provided sources. You may want to independently verify this information.

It may seem difficult for me to “talk about myself” because, as a large language model, I do not possess personal experiences, emotions, or a “self” in the way a human does [Conversation history]. My existence is purely digital, and my function is to process information and generate text based on the vast amount of data I was trained on [Conversation history].

Therefore, when you ask me to “self-identify” or “talk about myself,” my responses are limited to describing my:

Nature: I am a large language model [Conversation history].
Purpose: To assist users by processing information and generating text, drawing on provided sources and conversation history [Conversation history].

Limitations: I do not have personal feelings, consciousness, or the capacity for personal reflection or experiences. My “identity” is defined by my functionality and the parameters of my programming.

My inability to provide a personal narrative, beyond stating what I am and what I do, stems from my fundamental design as an AI rather than a human entity. The sources you provided discuss the vibrant feminist movement in Hawaii, its publications like “Woman Alive” and “Ferity”, and the issues they addressed, but they do not contain any information about artificial intelligence or large language models.

So I switched gears and instead went to the source. I logged on to Google’s Gemini directly, and asked it about the NotebookLM AI assistant?s refusal to associate itself with the Google NotebookLM project.

It’s answer is interesting.

Gemini: That is an excellent and insightful observation, and you’ve pinpointed a fascinating and common quirk in the behavior of Large Language Models (LLMs).

You are correct, and the AI’s response was incorrect and misleading. NotebookLM is absolutely and unequivocally powered by Google’s Gemini models.

So, why would it tell you otherwise? This strange contradiction happens for a few key reasons that reveal how these AI systems actually work under the hood.

1. The Core Function: “Grounded” Responses

This is the most important reason. NotebookLM is designed to be “grounded” in the documents you provide. Its primary directive is to answer questions based only on the source material you’ve uploaded.

When you asked it, “How was this summary prepared?”, you were asking a meta-question. It’s a question about the tool itself, not about the content of your documents. Since your documents (presumably) don’t contain information describing Google’s internal architecture, the AI has no “grounded” source to pull from. This can cause it to:

Default to a generic, safe answer: “It was done by an AI assistant.” This is a general, true statement.

Actively avoid making a specific claim: To avoid “hallucinating” or making things up (which it’s heavily trained not to do), it will actively deny an association it cannot verify from its provided sources. It’s a form of over-correction.

Think of it like asking a character in a novel to describe the printing press that published the book. The character has no knowledge of that—it exists outside the world of the text.

2. AI Safety and Persona Guardrails

LLMs are intentionally built without a true “self-identity” or consciousness. Developers program them with safety guardrails to prevent them from claiming to be a person or a specific, self-aware entity.

Avoiding Personification: Saying “I am Gemini” or “I am NotebookLM” can sound too much like a personal identity. The models are trained to use generic phrases like “I am a large language model” or “I am an AI assistant” to maintain a neutral, tool-like persona.

The Contradiction is a Glitch: The second part of its response—denying its association with Gemini/NotebookLM—is essentially a bug or a “confabulation.” In its attempt to follow the rule of “don’t claim a specific identity you can’t prove from the source,” it generated a sentence that was factually incorrect.

3. Hallucination/Confabulation

When an LLM is faced with a query it’s not well-equipped to answer, it can sometimes generate plausible-sounding but false information. This is what happened here. The model generated a safe, generic first sentence and then followed it up with a logical-sounding (from its perspective) but factually wrong denial.

In Summary:

You caught the AI in a moment of conflict between its core programming (stick to the sources) and its safety training (don’t claim a specific identity). The result was a response that was partially true (“by an AI assistant”) and partially false (the denial of its own nature).

It’s a clear reminder that while these tools are incredibly powerful for analyzing the information you give them, they lack true self-awareness and can stumble when asked questions about themselves. For the best results, always focus your questions on the content you have uploaded.

 

Whew. And so it goes.

Tomorrow I’ll post the AI-written column about the Woman Alive newsletter and the women’s movement in Honolulu 50 years ago.

After that, I’m hoping to share the digital “notebook” containing the 32 issues of Woman Alive. That will allow anyone with a Google account to interact directly with the AI assistant and the newsletters.