Ben Brumfield

I'd like to know if other people have seen Gemini explicitly reference historical events to provide context to a transcription process.

(And a big caveat: we have seen Gemini "reasoning" be an unreliable narrator at a higher rate than we see hallucinations in text output.)

within what it "knows" about world history. N.B. we did not provide any context for the image in our API call--it used the same prompt we use for any other page image on FromThePage, and it did not have previous page transcription outputs in its context window.

Mark Humpries and @foundhistory.bsky.social have written about Gemini "checking the math" in account books, implying that it has a kind of model of reality it uses while it transcribes. This seems like a further step: the transcription process is connected to a model of history, situating a page

Gemini Reasoning output containing this text: Deciphering Text Cues I’m now focusing on the handwriting style and identifying content clues. The presence of cursive strongly suggests the late 19th or early 20th century. Analyzing the inks, the top line is darker while the rest is red. My initial look at the text shows names like “Gervais” and “Toussaint Laplante” and locations such as “Fish Creek” and “Batoche”. The terms “Rebel” and “Battle” strongly suggest the North-West Rebellion of 1885.

One line in the reasoning produced by Gemini was fascinating:

My initial look at the text shows names like “Gervais” and “Toussaint Laplante” and locations such as “Fish Creek” and “Batoche”. The terms “Rebel” and “Battle” strongly suggest the North-West Rebellion of 1885.

Screenshot of AI generated text, reading: Gervais Madame Josephette 168 $ 1167.85 Fish Creek Widow of Calixte Tourond who was Killed at Batoche he was a Leading Rebel See her evidence in No 167 & No 168 D of J. evidence of Goulet , hus Killed at Batoche

The personal names were nearly indecipherable, as they were French names I hadn't encountered before, alhought the document was in English. Since QUA had run Gemini on the collection, we tried using that as an AI Draft, and were pretty pleased with the results. Nothing new there.

English-language text written in 19th-century script by one hand, using black ink for the first line (which contains a name, number, and dollar amount) with red ink for the remaining four and a half lines, apparently containing a terse narrative.

I stumbled across something interesting during a live demo in yesterday's webinar. Not having anything prepared, I tried a page in the Queens University Archives' Riel Resistance Collection. The hand was extraordinarily difficult for me, even though it was late 19th-century North American script

Ooh -- do you have it posted on FromThePage? We're trying out two new Gemini models and might be able to experiment.

Screenshot of the Newberry Library Digital Collections homepage featuring a pastel illustration of a polar bear on ice with mountains and a sun; a banner invites browsing the digital collections, with “Browse all” and “Recently added” thumbnails below.

Screenshot of the Newberry Transcribe homepage with a purple handwriting background; a large panel reads “Newberry Transcribe — Unlock history!” with buttons for “Learn more” and “Browse manuscripts,” and a row of project tiles below.

📌 Where to find us:

🔍 Browse our digital collections -- thousands of rare maps, manuscripts, postcards, and more, all free and online collections.newberry.org

✍️ Help us transcribe historical documents on Newberry Transcribe (no experience needed, just curiosity) nt.newberry.org

We're 75 minutes away from our webinar on responsible AI use in FromThePage--the challenges we face from AI and how we're addressing (some of) them: content.fromthepage.com/feb-2026-web...

Thank you! Given recent advances, I don't think I'd finalize more than six weeks in advance.

(I can't believe that I find myself thinking about a mid-November advance in capabilities: that's theoretically impossible, but in practice, I guess it works!)

Any chance you could share the reading list to us folks outside of BYU?

Quick blog post noting some thoughts on 'Documenting AI-created/enhanced records in catalogues/metadata/displays' - I'd love to know who's already doing it, and how? www.openobjects.org.uk/2026/02/docu... #AI4LAM #MuseTech

Mirador will ignore it (as well as the `creator` encoding the person who ran the AI).

So "doing it right" according to standards means doing it wrong according to best practices for AI transparency.

A big problem that I'm running into is that since I don't control the client, I'm stuck with whatever implementation you find in e.g. Mirador. So although it's technically possible to use a `generator` attribute in a WebAnnotation that semantically defines the act of using the AI software,

* If only the AI created transcript exists, show it with a warning prepended to it in bold
* For API access, provide both versions with stanzas indicating provenance, links to prompts, profiles for models, etc.

Expose AI engine in AI tab, IIIF API and TEI export · Issue #5188 · benwbrum/fromthepage Now that we are tracking the model used to create an AI plaintext in the ai_transcriptions table, we should be able to be transparent about that. We should add the following features: In app/views/...

I'm working on this today, in fact: github.com/benwbrum/fro...

The general principles I'm trying to follow are these:
* For human-visible page transcriptions (i.e. WebAnnotations on IIIF manifets):
* If both human-created and ai-created transcripts exist, only show the human-created one.

I am very excited to report that a project I have been working on for 3+ years is finally seeing the light of day! Along with @byzcapp.bsky.social and Jesse Torgerson, we debut a new annual section in Digital Philology that proposes the dataset as a new genre of publication. odfms.hcommons.org

Quotation from Peter Shillingsburg's "Dank Cellar of Electronic Texts": "The world is being overwhelmed by texts of unknown provenance, with unknown corruptions, representing unidentified or misidentified versions."

And suddenly we find ourselves back in the world of Project Gutenberg-adjacent electronic texts. Are we back in Peter Shillingsburg's "Dank Cellar" again?

already done the unpleasant layout so the human can focus on the data. (foundhistory.bsky.social has written about this with regard to weather observations)

means that it's a little tiring or unpleasant to read. (Transkribus also performs really well--like better than an amateur human well--on pages with bleed-through.)

We also are seeing financial records (or other tabular/"boring") documents become easier for volunteers to correct, since the LLM has

One of the biases we're already seeing comes from image quality and difficulty encoding. We find that Gemini performs very well indeed on scans from microfilm like fromthepage.com/lacountylibr... or fromthepage.com/lacountylibr... where the text is legible to a human, but the document quality

it might require a second pass at transcription; the first to produce a visually accurate transcription, the second to produce a derivative text optimized for screen reader users.

This is technically straightforward, but I'm not sure that library systems have a good place to put these derivaties.

using Markdown (for us) or other whitespace. It looks great! But it doesn't follow reading order the same way that traditional OCR does, which makes it a lot harder for a screen reader user.

This may be solvable--we're still experimenting with our prompts to tamper the enthusiasm for layout--but

to grapple with some new challenges. Gemini really wants to preserve the layout of a page, so we see a lot of attempts at whitespace indentation to produce a typographic facsimile or such things. The problem for OCR applications is that--for tables and columns--it produces realistic-looking tables

One of the things that strikes me is how much better MMLLMs--Gemini 3 specificially--is for OCR than traditional OCR. We did some tests yesterday with agricultural catalogs in a private collection, and the differences in quality of word recognition were staggering.

That said, we're going to have

My brain just exploded. Of all the things to make me decide "well, that's enough Internet for the day", I wouldn't have expected this.

(But I will tell this story over cocktails tonight.)

YouTube video by fromthepage Introducing Gemini 3.0 Support in FromThePage

This is the first time we've been willing to add LLM-supported transcription to FromThePage. Last week's webinar: www.youtube.com/watch?v=UhqR...

Congratulations to you and the team! Looking forward to following your work.

I'm not ready to discard the entire ecosystem we've built around making data openly available to possibly, hopefully slightly impair commercial AI training. Users uploading data (they often don't have the right to relicense for AI training) into AI models is a drop in the ocean.