Yet another milestone has been reached. ChatGPT (and Claude) can now conduct archival research in a digitized archive.
A month and a half ago, I tried to get ChatGPT to investigate a topic by employing digitized materials in the French National Library (BnF; https://gallica.bnf.fr/). The results were quite impressive; however, in examining the process that ChatGPT had gone through, I realized that it had not actually “read” digitized materials.
Instead, it had found relevant information in the metadata on the BnF site, and then had combined that with information that it had ingested as part of its training.
Then a couple of weeks ago I saw that Canadian historian Mark Humphries published an essay on Substack where he noted that Claude CoWork with Fable 5 can now conduct research on its own. (https://generativehistory.substack.com/p/no-more-muddling-through)
Humphries subscribes to the $200-a-month version and therefore gets access to more high-powered models than I do at $20 a month. Nonetheless, whatever appears at that level today, will show up at my level before too long, so I have been waiting for that moment.
Then last week OpenAI released GPT 6 Astra, and it has been blowing people’s minds with, among other things, its ability to create good-quality 3D models from images.
I went to check to see if I have Astra (not yet), but then I realized that a month and a half ago I had tried to get ChatGPT to conduct archival research from the Chat function, rather than from ChatGPT Work, its agent-based tool for complex, multi-step workplace tasks.
So, I started experimenting (still with GPT 5.6), and yes, it can conduct research.
A month and a half ago, it could look at metadata, and then go through the information it was trained on. Now, however, it can actually “read” documents in the BnF that it was not trained on, and acquire new knowledge, rather than having to rely on whatever was baked into the model, or what I manually uploaded to it.
It now works something like this:
trained model + ability to locate sources + ability to read sources + reasoning over those sources
That is a significant change. It means that the decisive question isn’t “Does ChatGPT know this?” It is “Can ChatGPT get into the right archive, recognize the relevant evidence, read enough of it, and construct an argument from it?”
And the answer to the second question is, “Yes! It can!!”
Buckle up, everyone! Not only has the train left the station, but it is now accelerating.
