AI Generated
In December, I blogged about a creepy AI discovery I had made when testing ChatGPT's ability to extract and summarize from newspaper articles. Information in the summary was provided by AI that was no where in the articles uploaded. Although errors like this, called hallucinations, are known to happen with this budding tool, the information that AI proved was 100% correct. That's what made this feel creepy. How did it know more about my family than the articles I entered?!
You can read the blog I'm referring to here.
I'm now enrolled in my second AI for Genealogy class through the National Genealogical Society and my instructor, Steve Little, noted in class this past week that AI was trained by, among other information, through blog posts.
Bingo! That explains how ChatGPT got my correct ancestors to include in the summary as I have blogged extensively about my Leininger and Landfair families. Since I've blogged about my families for years much of my research is available online.
I've spoken with other family historians who haven't had as good a result as I have with using AI. Perhaps this is why I'm getting the results that I am.
When the world wide web was in its infancy, I had difficulty with understanding the concept that once something is place there you can't get rid of it. Sure, websites come and go but the information is still out there if you look hard enough with tools like the Wayback Machine. Personally, I was involved with an educator only web in the early 1990s and on the world wide web since 1995 so my digital footprint is a large size. Who knew that it would be a help to me with the latest technology?!
If AI is not giving you the results you'd like, I'd suggest that you upload the info you're working with as a pdf when you prompt it. This might also be a good time for you to start blogging about your family. I will be providing more info on how to do that next week. Happy Hunting!
Did ChatGPT find your blog post, or did it use a PDF you uploaded to ChatGPT? I've done the latter, but it refused to tell me anything if I just asked "tell me about Thomas Richmond 1848-1917."
ReplyDeleteChatGPT must have been trained on my blog to have found the info that it reported to me in a narrative I asked it to write on 6 newspaper extracts that I uploaded as a PDF but didn't contain the correct information it reported. I don't think I can call correct info a hallucination. I believe a hallucination is made up info passed on as fact. I have no proof except what happened that someone, sometime fed ChatGPT my blog posts that contained the info that was correct. Drawing from it's training material, it correctly added to the newspaper info I submitted. If you have Thomas Richmond on a blog post or on a website try what I did. Upload some info about him and then see what additional info you are given. I wouldn't do it tonight, though, because it is acting really odd. I just uploaded a short PDF Individual Summary on one of my husband's ancestors and gave the same prompt to write a narrative I have always done and I got a summary with no title or footnotes. When I asked again I got a formal response that said it couldn't provide narratives with the info I gave it and gave me info on how to write a narrative (intro, body, conclusion). I responded that it has done about 20 narratives for me before and I asked why it could not do it now. In that response I called it the name that I use for it. I got a nice apology and what I wanted. So, the only difference was when I initially gave the prompt I didn't say, "Hi, (name I call it)! Not sure why it responded like it did and I completed 3 thumbs down so someone can look into it.
ReplyDelete