Limitations of ChatGPT

Question

I was playing around with ChatGPT-4 today (the new version). Here's my list of limitations relevant to using it for genealogy (same limitation also apply to ChatGPT-3.5).

• It can't access websites in real-time. This means you can't ask it to tell you anything about what's on wikitree - e.g., it can't tell you which of your ancestors on wikitree fought in the American Revolution. So Wikitree+, Ancestor Explorer and the other wikitree apps are not going to be put out of business (yet).
• Its knowledge is limited to info on websites and in digitized books and articles and does not include material behind paywalls, data in specific databases (e.g., it claims to have no knowledge of info in US Censuses) or primary sources (unless transcribed in secondary sources). It is therefore pretty useless for genealogical research. You're still going to need to go to ancestry, familysearch, findmypast, etc to do research.
• If asked, it won't provide any specific sources that support its assertions; will only suggest sources that "could" provide the info. So even for stuff it "knows", it can't tell you where it got the info from or even provide a source that it knows does support its statement.
• If asked to write an article with cited sources, even if it gets the facts right, it will provide citations to sources but the sources only "could" support it statements assertion based on their subject matter and sometimes don't. So for example, I asked it to write a genealogical profile for Sarah (Bass) Thayer, the daughter of John Bass and Ruth Alden (one of John Alden's daughters), with sources cited for each statement. It generated a profile that had all the basic facts right, but the sources cited didn't always support the statements. For example it cited "Great Migration" as the source for Sarah's parentage. However, although GM has info about Sarah's parents, it never mentions Sarah herself.

I am eager to test out a feature of GPT4 that has not yet been made available to the public: its ability to describe in detail what is in an image. If it can read old documents, even somewhat, that would be awesome. I suspect, however, that it will be unable to accurately read old English lettering.

Comments

Thanks for this list, Chase. Re your bullet point 1: the dataset for the previous version apparently only came up to 2021 (see here for example). Do you know if ChatGPT-4 still uses that same dataset, or is the cutoff date now more recent?

See also one, two, three and four.

---

Yes, GPT-4 says it is still limited to info before Sept 2021.

Answer 1

There is a plugin that adds a rudimentary websearch. But it's power lies elsewhere. 

It will make up citations if you let it, but if you give it a bunch of citations/descriptions and ask it to write it based on that it's great. 

* Throw in lists of ancestors and ask it to find connections, spot mistakes

* Fix formatting on messy wikitree pages (You might need to remind it a few times to use wikitree link format)
*  Load in a bunch of historical text and ask it to summarise it for you in plain english, ask follow up questions, get it to explain why it's relevant. 

* Load messy note pages into tidy spreadsheets

I have ~200 pages of notes in markdown for my surname project, eagerly awaiting GPT4 API access. 

Looks like people are already starting to find other uses for it too https://bronzedrose.com/tag/chatgpt/

edit: Also looks very useful

https://books.google.com/talktobooks/

Answer 2

Have fun with ChatGpt, but if it was me I would only use it very very lighty for lots of reasons.

Answer 3

OpenAI's ChatGPT and Google's Bard are LLMs (Large Language Models). For a nuanced discussion, see here.

As such, they may be a superior alternative to the query function of a relational database.

They are incapable of emulating human experience, and so, fail the Turing Test.

Lacking sentience, they are also incapable of emulating most human mental conditions and actions. They may be programmed to emulate metacognition (it is not caused by intentionality).

Comments

As such, they may be a superior alternative to the query function of a relational database.

*snort*, as someone who lives by talking to users and then querying relational databases to solve their questions, I have reservations about this statement. Most of the problem lies in assumptions and misconceptions about what the data actually shows or how it is tied together. 

I hope to be proven wrong, but it seems likely to me that a LLM will just be slightly wrong....faster. And in ways that are subtly hard to decipher, to an end user who has no full understanding of how the data is being pulled out.

---

The operative word is "may".

According to the linked discussion, input determines LLM processing context, hence; output. Garbage in, garbage out.

---

Apropos of absolutely nothing, but some might find this a curiously interesting read: Roivainen, Eka. "I Gave ChatGPT an IQ Test. Here's What I Discovered." Scientific American, Opinion (28 March 2023); https://www.scientificamerican.com/article/i-gave-chatgpt-an-iq-test-heres-what-i-discovered/.

Answer 4

I realize this post is ancient in the GPT timeline. Current version is quite robust, but here's the deal...unfortunately the free version does not scrape the internet - have to have the paid version which is $20/month.

I am currently using GPT to create a custom GPT to read a Wikitree profile and then to help me write the person's biography in a Wiki code format in accordance with the WT guidelines. Check out [[parton-340 | Willard C Parton]].  Creating the custom GPT is still work in progress. Let me know what you think. #ChatGPT #AI

Brian