Posts
AI's rapid advancement is caused by three factors: raw compute increases, algorithmic efficiency improvements, and "unhobbling" processes. In the current super-accelerated decade, we enjoy a 10,000x AI performance boost every four years, making it likely that AI will exceed top human experts by 2027. There are good arguments that AI progress past 2027 could be slower, but also that it could be even faster.
At some point, genealogists will have to face the fact that AI can accurately do much of the work genealogists do. An easy example, not considering document transcription and translation, would be the ability to search huge databases from a single prompt. In this regard, record hints and matches are quickly becoming more accurate. It will be only a short time before AI will be used to search nearly all the huge online genealogical record collections. For example, FamilySearch is spending huge resources to extend AI-supported Full-text search.
Meanwhile, the serious, advanced genealogists will comfort themselves with the seemingly huge archival collections that have yet to be digitized. One of the most extreme examples is the United States National Archives where only a very small percentage of the documents have been digitized.
One of the main commonly expressed arguments against the utility of AI is the issue of the inappropriately named "hallucinations." One way of responding to this background concern is to ask, compared to what? AI is essentially a computer program. Top commercial models (like Google's Gemini Flash variants and standard GPT models) have dropped below a 1% hallucination rate on basic apples-to-apples summarization tests.
See Graffius, S. M. (2026, January 7). Are AI Hallucinations Getting Better or Worse? We Analyzed the Data. ScottGraffius.com. https://doi.org/10.13140/RG.2.2.33179.53285.
One way to control the hallucination issue is to use your chat program to self-critique or correct any task you have asked it to do. For example, if I ask Gemini to transcribe and/or translate a document, I use a specially designed prompt to initiate the process and then I start a new chat and give Gemini the prompt I used and the results and ask it to review and correct the output. You may be surprised at the response.
Now, if AI continues to get better and expand in its capabilities, how much more of what genealogists do could be done more accurately and much faster than even the most expert level genealogist?
Here is a list of some of the genealogical tasks that would be enhanced or replaced by using AI.
Searching for valuable historic and genealogically useful information from online and paper-based archives and websites. Granted there is a current "paywall" issue. Without permission, most genealogy database sites are protected by account authentication layers, paywalls, and robots.txt/web-scraping blockers. But once we do the job of finding a record on a protected website, an AI assistant can transcribe, translate, and analyze the document. Currently with context-sensitive AI, anything you can see on your computer can be seen by the AI assistant and analyzed despite a paywall. This only means that the paywall sites need to catch up to the new technology if they want to do so.
Google's Gemini 3 generation operates with a standard input context window of 1 million tokens—expandable up to 2 million tokens in advanced developer and enterprise environments—and supports up to 64,000 output tokens. Across its natively multimodal architecture, a 1-million-token input translates to roughly 700,000 words, approximately 1,500 single-spaced pages of text, 1 hour of video, or 11 hours of audio in a single prompt.
See Google DeepMind, Gemini 3 Pro Model Card (Mountain View, CA: Google, 2025; updated May 2026), 2–3, Google DeepMind Media.
How long would it take you to search through this number of pages?
From my own experience I have added 120 original documents that include deeds, probate files, town records and other similar documents and using Gemini Notebook, Gemini generated a spreadsheet showing all the people named in all the documents with citations to the documents and showing the relationship, if present, of all the people named. The long part of this analysis was locating and uploading the documents. Gemini did the analysis in a matter of a few minutes.
How long would it take you to do the same analysis?
Using FamilySearch's Full-text search, I was able to find deeds and probate documents about my ancestors that I had never and would not have even spent the time to find.
What new-to-you documents do you think a full-text search might reveal?
As AI matures and becomes more sophisticated and reliable, the tasks it will be able to do will continue to expand. Those genealogists that are still working on paper or even those who do not care to learn about AI may one day wake up to the fact that any work they do is already done by an online website such as FamilySearch.
See Trent Toone, “What Latter-day Saints Should Know About Temple and Family History,” Church News, March 6, 2026, https://www.thechurchnews.com/leaders/2026/03/06/elder-mark-bragg-rootstech-2026-what-latter-day-saints-should-know-temple-family-history/.
Can a genealogist remain relevant? The ones that are relevant today will be the ones who are relevant tomorrow. These relevant genealogists are the ones that are anxious to adopt whatever advances in technology that will help them do their work better, faster, and with more accurate content. AI is already doing this for a large number of genealogists.
Additional Note: I use Google Gemini for my examples for the simple reason that Google has optimized Gemini for deep research and synthesis, which aligns directly with what I do as a genealogist.
No comments:
Post a Comment