For seminaries, theological schools and religious institutions
The edition you chose,quoted word for word.
A simulation built from an archive your faculty assembles: the translation your courses use, the commentators you read, the writings of the subject itself. When it quotes, our system copies the words out of the document.
30 minutes · with a dean, a course lead or your library director · nothing signed
- 0
- quotes the AI made up, across 1,323 in our internal benchmark
- 6
- religious schools working with Humy today
- 1,000
- pages per record, handwriting included
A rule that says check every AI quotation has handed someone a checking job with no end to it.
One Roman Catholic seminary writes it into its academic regulations: AI output “can be inaccurate, misleading, or entirely fabricated,” and the student must “carefully review your AI-generated content before employing it in any fashion.” Saint Meinrad Seminary, AI policy (one school’s policy only; archived 14 Mar 2026)
You supply the canon, and it has no second library.
- 01
Scripture in the translation you teach
The edition you load is the edition it quotes. Where courses work from more than one, publish a separate version for each.
- 02
The commentators you rely on
Patristic, rabbinic, scholastic, confessional or modern, in the editions your faculty considers sound.
- 03
The writings and records of the subject
Letters, sermons and treatises; the acts and canons of a council; the records a site has left. No summary stands in for the text.
- 04
Nothing beyond that
A text your faculty did not supply is a text it cannot quote, and it is instructed to say when the archive is silent.
“Late have I loved you, beauty so old and so new: late have I loved you.”
The AI never types a quotation. It marks the passage; our system copies the words out of your document and discards anything that is not an exact match. A commentary is handled the same way as canonical text, which is where interpretive risk actually sits.
- Per simulation
- Up to 50 source records.
- Per record
- Up to 1,000 pages, so a bound volume goes in whole.
- Scanned pages
- Pages that are pictures, not text, are transcribed one at a time by Gemini 3.1 Pro, historical handwriting included.
- Held per source
- Creator, date, document type, holding repository, accession number, and whether the account is first-hand or later.
A collection assembled by a vendor has to pick a tradition. The sector does not have one.
Counted across the 281 member schools of the Association of Theological Schools. These are the sector’s numbers, not Humy’s customers. ATS Annual Data Tables, 2024-25 (Table 1.3; all 281 member schools, US and Canada)
The 281 ATS member schools, by affiliation
- 160
- Protestant denominational schools
- 65
- Non, inter or multidenominational schools
- 49
- Roman Catholic schools
- 5
- Orthodox schools
- 1
- Eastern Rite Catholic school
- 1
- Jewish school
Nothing reaches a classuntil one named person releases it.
The same regulation puts that with the person teaching the course: generative AI may be used in coursework only with the express permission of the instructor of record. A figure sits in draft while somebody works on it. No committee has to convene.
- 01
Assemble the sources
Editions, commentaries and the school’s own papers go in as records. Scans that are pictures are transcribed before anything can read them.
Librarian or faculty - 02
Set the scope, per version
Reading level, language, and what the figure is for. Questions of doctrine or pastoral direction go back to a member of faculty.
Course lead - 03
Test it while it is unpublished
Put the questions you are worried about to it before anyone else can reach it. Change the sources, or leave it unpublished.
Instructor of record - 04
Publish, then read the transcripts
Conversations are logged in full, searchable and exportable, so a tutor can see how a question was actually handled.
After release
On invented quotations
A paraphrase inside quotation marksis the failure nobody catches.
Chapter and verse supplied with complete confidence, for wording the passage does not contain. It is the first objection a dean raises, and it should be. Ours is not permitted to write a quotation at all.
The checking duty your policy already carries does not go away. It changes size. What is left to check is the wording around the quote, because the quoted words were copied out of a document you supplied.
We publish the rate at which we still get that wrong. We have not found a supplier in this category that publishes one of its own.
- 0
- quotes that did not match the document
- 0
- highlights on the wrong passage
- 0
- quotes from outside the figure’s own sources
Humy internal benchmark, August 2026, one configuration. Across the 1,323 quotes in that run, one citation opened nothing, and we report it rather than round it away.
What that does not cover: the sentences around a quote are still written by the AI. Your documents cut made-up facts from 22.6% of answers to 12.2%, about one answer in eight still getting a date or a reference wrong. The quote itself is always real, and the document stays one click away. Read the benchmark.
Scoped for a graduate school, not sold off a price list.
- 01
One archive, several courses
Sources and figures are shared inside an institution, so an edition your library has loaded and checked once can carry a Greek exegesis course and a church history course, each publishing its own version.
- 02
What it costs
There is no per-seat list price for a seminary. Scope depends on how many courses are involved and how much of your material has to be transcribed first.
- 03
What it does not do yet
A figure does not remember a student between sessions. Every conversation starts clean. Memory across sessions is coming, and it is not here today.
What we can and cannot claim
Six religious schools work with Humy, and the Private University of Education of the Diocese of Linz, in Austria, took ten licences for a one-year research collaboration. That is a small body of experience and not a long track record. Ask on the call and we will put you in touch with someone using it.
Name a text, and we will build the figure.
Name one text in the edition you teach from and say what you want built on it. You can put your own questions to it before anybody decides anything.
Questions a dean tends to ask.
Source-grounded AI simulations for study: a historical figure, a council or controversy, or a site of significance. Each is built from an archive your faculty selects: scripture in the translation you teach, the commentators and confessional documents you rely on, and the writings and records of the subject itself. It answers from that archive with the passage attached. Students can question Augustine, Aquinas, Maimonides, or a founder of your own tradition, and check every citation against the text.
No. The builder takes a name, a short description and the documents, so the steps that produce a historical figure produce a council, a controversy or a site of significance just as well. A class can put questions to the Council of Chalcedon and get answers out of its acts and the documents that argued over them, or to a site the school studies and get answers out of the records about it. Nothing about the quotation changes: the words between the quotation marks are still copied out of a document you supplied.
The one you supply. Translation is an institutional decision and it stays one here: the edition you load is the edition it quotes, and it cannot quote a text that is not in the archive. Where a tradition works from several editions, you publish separate versions of the same figure for different courses.
The instructor of record, matching the policy quoted further up this page. The workspace is a curation step: one person assembles the sources, sets the figure's instructions and reading level, puts the questions they are worried about to it while it is still unpublished, and publishes it when they are satisfied. No committee has to convene, and the product contains no doctrinal review board.
The model never writes a quotation. It marks which passage of a document it means, and our system copies those words straight out of the source, so the AI never types a quote itself; anything that is not a character-for-character match is thrown away. In our internal benchmark of Humy 2.6, all 1,323 quotes matched their documents word for word, every highlight landed on the right passage, and none came from outside the figure's own sources. The wording around the quote is a separate question, answered next.
That is still written by the AI. In the same benchmark, giving the figure your documents cut made-up facts from 22.6% of answers to 12.2%, so about one answer in eight still got a date or a reference wrong in the text around the quote. The quote itself is always real. The sentence framing it needs the checking a student would give any secondary source, which is why the source document stays one click away.
You set what a figure is for. Instructions are configured per version, so a figure can be scoped to explain what a text says and how a named commentator read it, and to refer a student to faculty for pastoral or doctrinal direction. Conversations are logged in full, so a tutor can see how a question was handled.
No. A figure is a way to work a text and the commentary written about it. Questions of vocation, conscience and pastoral direction belong with your faculty and formation staff, and a figure can be instructed to send a student back to them.
Yes. Up to 50 source records per figure and up to 1,000 pages per record, including scanned volumes and handwritten material. Pages that are pictures, not text, are transcribed one page at a time by Gemini 3.1 Pro, historical handwriting included, so older editions and archival papers can be used, not only what already exists in modern digital form.
Yes, and it is a short list. Six religious schools use Humy, and the Private University of Education of the Diocese of Linz in Austria took 10 licences for a one-year research collaboration. We are glad to put a prospective institution in touch as part of a scoping conversation.
Sources and figures are shared inside an institution, so an edition your library has assembled and checked once can carry several courses, each publishing its own version at its own reading level. An arrangement spanning separate institutions is scoped with us directly, and each institution still decides for itself what its own students see.
Conversations are never used to train AI models. We hold signed student data privacy agreements covering 16 US states, 15 of them through The Education Cooperative and California separately, and we align with FERPA. Data is stored in the US or in Sweden on Microsoft Azure, and the institution can search and export the full conversation logs. The signed agreements are published at humy.ai/security.
Something not covered here? Write to hello@humy.ai.