How We Use AI

Many of the texts and translations in this library were produced with the help of AI. That may seem obvious, but it should not be left implicit. This page says openly how AI is used here, what it can and cannot guarantee, and how its errors are caught and fixed.

ScholaThomistica is a collaborative platform, and I do not ask contributors how they produce their transcriptions or translations; I review the quality of what they submit. What follows describes how I work. Since I digitize most of the texts in the library, it covers most of what you will read here.

At the center of this platform are the texts themselves, insofar as they hand on Catholic truth and tradition. They are here for the sake of contemplation, which perfects the human intellect and lifts the mind to God.

Technology, AI included, is used in the service of that end. It is a means of making these texts available to be read, not a replacement for reading them.

AI is used, first, to make these texts available at all. Turning an old printed book into an accurate digital text used to take hundreds of hours of expert labor. With AI agents it can now take a dozen hours or fewer, at a fraction of the cost.

That cost is precisely why so many of these works were never digitized. Expert scholarly labor is expensive, and there was never enough of it for the wider Scholastic tradition. Now it is feasible, and so the texts are here.

I have written about the method in How I Digitize Old Theological Books with AI. The details have moved on since then, but the principles are the same.

Through a great deal of trial and error, we have found processes that produce text of quite high quality. Each work is checked against independent readings of the same page, and disagreements are settled by looking at the scan itself.

But we cannot guarantee complete accuracy. No machine-learning method can. That is why we provide facsimiles: for a growing number of works you can open the scan of the printed page beside the Latin transcription and compare them directly. We are working to extend this to every work.

Because this is a collaborative platform, readers who are already reading a text can propose a fix when they spot a mistake. An administrator reviews the correction, and once it is accepted the text is updated for everyone.

These corrections, paired with their scans, are also a kind of data labeling. In the future we may use them to train our own transcription models, so that each fix improves not only the passage it touches but the system that produces the next texts. We have not trained any models yet; this is a direction, not something already in place.

So the text is never finished. Errors creep in, but there are mechanisms to catch them and to fix them, and the library improves continually.

The goal of the platform is the flourishing of the person, and especially of the intellect, through contemplation. Learning Latin serves that goal, which is why the original language is at the center here. Every work opens in its original language: Plato in Greek, St. Thomas in Latin.

AI translations are treated as provisional. In my opinion, current large language models translate Latin into English quite competently, and their translations are good enough to serve as working translations in many cases. Personally, I would take a translation of the Summa by a frontier model such as Claude Opus over the English Dominican Fathers’ translation: it reads the Latin better and renders it better. That is my own judgment, not a claim the platform makes.

Some scholars have generously contributed their own translations, and those can be offered as scholarly translations. But we cannot wait for scholars to translate every text, or we would wait a very long time. So our practice is to provide provisional AI translations for everything we can. Having the Latin means we can translate into any language: English, Spanish, German, or others.

Like the Latin, the translations can be improved by readers, and some already have been.

I would not recommend AI translations for scholarly use. Do not quote them in a paper unless you know Latin and can check them against the original yourself.

Instead, I encourage you to learn Latin and to use this platform to do it. Read the original alongside the translation, and use the translation as an aid that brings you to the point where you can read the Latin on your own.

Accuracy and scholarly authority are different things. AI and reader corrections can make a text accurate; authority comes from a scholar who takes responsibility for it. I have written more about this distinction in AI Can Translate the Books. Who Will Verify Them?

We use AI, openly, in the service of the human person. We have mechanisms to prevent errors, mechanisms to fix the ones that get through, and a library that keeps improving as people read it.