Lessons Learned from Google NotebookLM

One of the most common complaints about artificial intelligence tools is that they “make things up.” Genealogists are especially sensitive to this issue because we spend so much time evaluating evidence and separating fact from speculation.

While recently experimenting with Google NotebookLM, I discovered that the tool wasn’t exactly inventing information. Instead, it was making assumptions based on the data I provided.

The source material was a narrative report generated from my RootsMagic file about my second great-grandfather, Albert Hutchinson. I used the report to create audio overviews, infographics, and videos. While the results were impressive, they also revealed some important limitations.

Assumption #1: Albert Was a Farm Laborer

In one audio overview, NotebookLM described fourteen-year-old Albert Hutchinson as working as a farm laborer in the household of Ephraim Finch in 1850.

At first glance, that seems like a reasonable conclusion. The report stated that Albert was living in the household of Ephraim Finch, a farmer. Since Albert was a teenage boy living on a farm, the AI inferred that he was likely helping with farm work.

The problem is that the report did not contain the full story.

Additional research suggests that Cynthia Finch, Ephraim’s wife, was Albert’s aunt. That family relationship helps explain why Albert was living in the household and provides important context that was missing from the narrative report.

Was Albert helping on the farm? Quite possibly.

Do I have evidence proving he was a farm laborer? No.

NotebookLM took a logical step beyond the available evidence and presented the conclusion as fact.

Assumption #2: The Mystery of the Thirteenth Child

A second error appeared in an infographic generated from an NGSQ-style modified register report.

The infographic stated that Albert Hutchinson had thirteen children.

That immediately caught my attention because Albert actually had twelve children.

After examining the report, I realized what had happened. In a modified register report, the principal person is assigned the number “1” and the children are numbered beginning with “2.”

Before uploading the report, I had removed the number “1” from Albert’s entry because it looked unnecessary.

Unfortunately, NotebookLM interpreted the numbering sequence literally. Since the first child was numbered “2,” the AI assumed there must have been a missing child numbered “1.” As a result, it counted Albert’s twelve documented children plus an imaginary “Child #1,” arriving at a total of thirteen children.

The error wasn’t caused by bad genealogy. It was caused by the AI trying to make sense of a numbering system it didn’t understand.

What I Learned

These experiences taught me that AI tools are only as good as the information and context we provide.

When using genealogy reports with AI tools such as NotebookLM, it helps to remember that the software does not understand our research process. It only sees the words and numbers that appear in the document.

As genealogists, we often know facts that are not explicitly stated in a report. We understand family relationships, recognize standard report formats, and know when evidence is incomplete. AI tools do not possess that background knowledge unless we provide it.

Moving Forward

As I continue using AI to create biographies, infographics, audio overviews, and videos, I plan to make a few adjustments.

First, I will add explanatory notes to key events when additional context is important. For example, I can note that Cynthia Finch was Albert’s aunt rather than relying on the AI to infer why he was living in her household.

Second, I will explain report formats and numbering systems before uploading the material. A simple note that the principal person is numbered “1” and children begin with “2” may prevent the AI from inventing an extra child.

Finally, I will continue reviewing every AI-generated product with the same critical eye I use when evaluating genealogical records. AI can be a powerful assistant, but it still requires a genealogist to verify the results.

In this case, Google NotebookLM reminded me of an important truth: artificial intelligence doesn’t always make things up. Sometimes it simply makes assumptions—and those assumptions can be surprisingly easy to miss.