How we use AI, and where people decide
Machines read the pages and draft the description. People check what matters, and every field in the record says which of the two it came from.
We use machine learning in three places: reading text, drafting description and powering search. Each saves time that small institutions rarely have, and each makes mistakes that a reader cannot see unless someone says where the machine was involved. This page sets out what the machines do, what people check, how you can tell the difference in the records we deliver, and what you can refuse. It describes how we work. It is not a summary of any law, and no professional body we know of requires all of it.
What machines do
- Text recognition. Optical character recognition (OCR) for print and typescript, and handwritten text recognition (HTR) for manuscripts. The output is text with the position of each word on the image. See text recognition.
- Suggested description. Models read the recognized text, and sometimes the image, and suggest names, places, dates and subjects, matched where possible against the vocabularies you choose, such as Library of Congress Subject Headings or your own name list. See AI cataloging.
- Summaries. A short first draft describing what an item is about, for a cataloger to correct or replace.
- Embeddings for meaning-based search. A model turns each passage and record into a list of numbers, placed so that passages with similar meaning sit close together. A search for “strike” can then find “the smelter men walked out” even though the word never appears. Embeddings add nothing to your records; they are part of the search index. The example collection shows the idea.
Machines do not decide rights statements, access restrictions, what is sensitive or what is published. People decide those.
What people do
- Transcribe sample pages exactly, so recognition can be measured against them.
- Correct text to the level you choose for each batch: full, targeted or none.
- Accept, edit or reject every suggested field before it is marked reviewed.
- Write description where the machine has nothing useful to offer.
- Make every rights, restriction and sensitivity decision with you.
Reviewers can be our catalogers, your staff or your volunteers working from a review sheet we set up; the agreement says who reviews what.
How machine output is labeled
Every descriptive field carries one of four labels, with the reviewer and the date where a person was involved.
| Label | What it means |
|---|---|
| Machine-suggested | Produced by a model. No person has checked it. |
| Reviewed | Suggested by a model, then checked and accepted by a person. |
| Edited | Suggested by a model, then changed by a person. |
| Human-entered | Written by a person with no machine suggestion. |
The labels travel with the records. In CSV they sit in a column beside each field; in Dublin Core, MARC 21 or EAD they go in a note, or in a field your system provides for it. On a search site we build, unreviewed fields show a visible label. Page text is labeled the same way: each page says whether it is uncorrected, corrected for names, places, dates and numbers, or fully corrected, with the error rate measured for its batch.
Why this matters: in the Library of Congress’s Exploring Computational Description experiments, catalogers reviewed machine-generated subjects and names, and the Library concluded that “no current AI tools are currently good enough to run fully automatically” for catalog records. Identifiers, authors and titles were extracted at around 85 percent accuracy; subject headings, genre and dates at or under 50 percent. The Library’s own AI principles stress transparency and keep “human involvement and expertise” at the center. Labeling each field is our way of acting on that. It is a choice we make, not a rule we are citing. A 2026 document prepared for the Society of American Archivists’ Council recommends forming an AI task force to develop best-practice guidance, including for description and metadata; we will review our practice against whatever it publishes.
How we measure accuracy
- Text. Character error rate and word error rate, measured on human transcriptions of sample pages, separately for print, typescript and each hand, and measured again after correction. Transkribus, in its own guidance, describes 0.5 to 2 percent as good for print and 2 to 8 percent for handwriting, depending on how many hands a model was trained on. We report what we measure on your pages rather than quoting ranges. The OCR accuracy calculator shows the arithmetic.
- Fields. For each kind of field, we count how many suggestions were accepted, edited and rejected. That tells you where machine help saves time and where a cataloger may as well start from scratch.
- Summaries. Read against the item for errors of fact. A summary is never treated as a transcription.
- Changes. If we change an engine or model partway through a project, we measure again before relying on it.
Sensitive material
Collections hold information about people who may still be living: medical, school, court, adoption and financial records, and anything a donor restricted. We identify these during the assessment and agree how each is handled. The options include leaving the material out of machine processing altogether, processing it only on systems we control, keeping restricted fields out of every public export, and marking restrictions in the record.
Originals also use language that is offensive today. Transcriptions keep the original words, because they are evidence. Description is a different matter: machine suggestions can repeat those terms or add new ones, so reviewers apply your description policy, and, if you ask, no subject term is published until a person has reviewed it.
Culturally sensitive collections
Some material belongs to communities with their own rules about who may see, hear or describe it. The Society of American Archivists endorsed the Protocols for Native American Archival Materials as an external standard in 2018, and Local Contexts Traditional Knowledge Labels let communities state their own conditions for sharing, including restrictions on sacred, ceremonial, seasonal or gendered material. For collections like these, we ask for the written agreement of the community or the holding institution before any machine processing, we follow their decisions on description and publication, and we stop if asked. Records can carry TK Labels and restricted fields on request. See community archives.
Your material and model training
Your images, text and records are not used to train any third party’s general-purpose model without your written permission. If a recognition model is trained for your collection, for example on one writer’s hand, the agreement says who may use that model afterward. The pilot report names every engine, model and service used on your material, and says whether anything left systems we control.
What you can opt out of
You can say no to any machine step, for the whole collection or for one series:
- suggested names, places, dates or subjects, so catalogers describe by hand;
- machine summaries;
- meaning-based search, leaving keyword search only;
- any outside or hosted service;
- training a custom model on your material;
- all machine processing of material you mark sensitive.
Opting out usually means more hours of human work, and we will tell you how many before you decide. Your choices are written into the agreement.
Professional guidance we work from
The Library of Congress’s AI Planning Framework moves through three phases, Understand, Experiment and Implement, and asks institutions to weigh data, models and people at each. A pilot box is a small, measured experiment on your own material before anything is implemented at scale. The Reading Room’s cataloging and metadata section covers these questions in more depth.
Send one box first.
A pilot box shows you the images, the recognized text and the catalog records before you commit a whole collection. Tell us what is in the box and what you want people to find.
Studio and contacts
24 South Weber Street, Suite 225Colorado Springs, CO 80903(719) 244-8450sales@dastok.com