For most handwritten collections the practical plan is both: a machine makes a first draft of every page, volunteers correct what matters most, staff check a sample, and every page records who or what produced its text. Volunteers alone remain the better choice for small collections, hard hands, community material and projects where engagement is itself the goal. Machine transcription alone suits large volumes that need to be searchable soon, as long as the text is labeled as uncorrected and nobody quotes it as a transcription.
This article compares the three models using published experience from US programs: the Library of Congress’s By the People and the FixIt+ correction platform for public broadcasting, the National Archives’ Citizen Archivist program, the Library of Virginia’s Making History: Transcribe, and a controlled experiment on the Zooniverse platform. It covers what volunteers contribute beyond text, how machine drafts change their work, how to review and label the result, and how to choose for your own collection. How accurate machine transcription of old handwriting is, by method and condition, is covered in our companion article, Can AI read old handwriting?
A note on interest. Dastok is a Colorado Springs firm that digitizes historical documents, photographs and museum collections and runs text recognition for print and handwriting, so machine transcription is part of what we sell. The evidence below comes from the institutions and studies cited, not from our own projects.
Three ways to transcribe a collection
The three models differ in who makes the first reading of each page and who checks it. In the volunteer model, people read every page from the image. In the machine model, software reads every page and people check only a sample. In the hybrid model, software makes a draft and people correct it, fully or in targeted fields.
| Volunteers only | Machine only | Machine draft, volunteers correct | |
|---|---|---|---|
| First reading | A volunteer, from the image | An OCR or handwritten text recognition (HTR) engine, or a language model | The machine |
| Who checks | Other volunteers or staff | Staff, on a random sample | Volunteers, then staff on a sample |
| Speed to searchable text | Slowest; depends on volunteer hours | Fastest | Fast draft, slower final text |
| Text quality | High when reviewed; uncertain readings marked | Varies widely with the hand and the engine | Depends on how thoroughly the draft is checked |
| Engagement and local knowledge | Highest | None | Moderate, if tasks stay meaningful |
| Staff time | Recruitment, training, moderation, review | Setup, measurement, sampling | All of the above, plus managing two sources of text |
| Main risk | A program that stalls | Fluent errors taken as fact | Volunteers accepting the draft without reading the image |
A fourth pattern is worth considering: publish uncorrected machine text for search right away, labeled as such, and run a volunteer program on the most used series over time. That is a sequencing choice rather than a separate model, and it often suits a small archive best.
Decide first what the text is for
The right model depends on what the text has to do. Text that only needs to make pages findable can tolerate errors, as long as it is labeled. Text that will be read in place of the handwriting has to be right. By the People notes that once volunteer transcriptions are integrated into loc.gov they become “fully searchable and readable by accessibility technologies,” which means a screen reader user may meet your transcription as the only version of the page they can read. Sorensen’s list of reasons for accurate transcripts at the American Archive of Public Broadcasting includes access for Deaf and hard of hearing users, support for educators who rely on text, and helping researchers “pinpoint specific moments, topics, and voices.” Write down the uses for each series before choosing a model, because they set how much checking the text needs.
What volunteers give beyond text
Volunteers produce transcriptions, but programs that run them report other results that matter as much: more people know the collection, names and places get read by people who know them, and the text is checked by readers who care about it.
Scale and reach
The Library of Congress’s By the People program, launched in 2018, announced in March 2026 that its volunteers had completed one million transcriptions by the end of 2025. In 2025 it welcomed its 50,000th registered volunteer, and since 2018 volunteers have taken part in more than 50 transcription campaigns. The Library of Virginia’s Making History: Transcribe has run for more than ten years on the FromThePage platform, according to a 2025 post by Jessi Bennett.
Search that finds names in the text
Transcriptions make handwritten pages searchable by the words inside them. In a November 2022 analysis, By the People staff compared loc.gov search data before and after volunteer transcriptions were added to three collections:
| Collection | Period after transcriptions | Change in search terms leading to items | Change in visits from search |
|---|---|---|---|
| Branch Rickey papers | One year | +86% | +93% |
| Carrie Chapman Catt papers | Six months | +47% | +43% |
| Rosa Parks papers | Not stated in the post | +3% | +23% |
The authors suggest the Rosa Parks figure was modest because that collection is already among the Library’s most visited. They also found that many search terms leading to the collections were names “found only in the transcriptions, not in the titles or other metadata,” such as correspondents in Catt’s network of suffragists. That is the case for transcription in one sentence: catalog records name the creator, and transcriptions name everyone else.
Motivation, learning and local knowledge
Volunteers come for reasons a machine cannot replace. In By the People’s 2021 volunteer survey, the top motivations were making collections more accessible and discoverable, giving back to an organization that serves the public good, and interest in the topics and people in the campaigns. Over 90 percent of respondents agreed or strongly agreed with each of those three statements. Respondents also described improving skills such as handwriting, research and Spanish, and feeling part of a learning community.
Local knowledge is the hardest contribution to replace. Meghan Sorensen of the American Archive of Public Broadcasting, interviewed by the Library of Congress in April 2026, put it this way: “Local lingo, dialects, and unique names often get lost in the mix, and it’s best to rely on folks who know them well.” She also described volunteers correcting transcripts of interviews with their own relatives and learning family history in the process.
Readings that admit uncertainty
Human transcribers mark what they cannot read. FromThePage, the transcription platform the Library of Virginia’s program runs on, argues in an October 2025 blog post that people are “willing to write things like ‘[illegible]’ or ‘[unclear: possibly Jones?]’ when they aren’t certain,” while language models tend to “fill in words that make sense.” The post is a platform operator’s view, not a study, but the distinction it draws is one every program should design for.
Example (fictional). The Aster County Historical Society runs a small volunteer evening on Box 3, the Okerlund family papers. A volunteer who grew up near Wren Gulch reads a faint name in Ida Okerlund’s July 1895 letter as “Halvorsen,” the town’s photographer, because she knows the studio’s cabinet cards. A machine draft of the same line reads “Halverson,” which matches nothing else in the box. The volunteer’s reading links the letter to two photographs. No error rate would have shown the difference.
What machines give, and where they fall short
Machine transcription gives speed and scale: a whole collection can be searchable in days rather than years. Its weaknesses are errors that look like good text, uneven accuracy across hands and sources, and no sense of when it is guessing.
The American Archive of Public Broadcasting, a collaboration between GBH and the Library of Congress, holds more than 100,000 digitized programs and uses speech-to-text to make them searchable. Its statement on content, metadata and transcripts is frank: such transcripts “often yield inaccuracies, in particular for lower quality audio, programs that contain a variety of different types of non-speech sounds, and those featuring speakers with regional or cultural accents or dialects.” Old handwriting raises the same problems in visual form: faded ink, unusual hands, local names.
For handwriting, published tests show how much results vary. In one study of 18th and 19th century English handwriting, Humphries and colleagues (2024, a preprint) report character error rates of 5.7 to 7 percent for general-purpose language models without training, and as low as 1.8 percent when a language model corrected a draft. That is useful for search and still short of a text you can quote. Our companion article on AI handwriting recognition reports these and other studies with their conditions.
Machines also fail in ways volunteers rarely do. FromThePage’s post describes language models silently regularizing spelling and adding plausible words, and notes that “spot-checking LLM text can be fiendishly difficult.” The Library of Virginia found a related problem with a different kind of software: a volunteer using speech-to-text dictation produced “anime” for “enemy,” “corpse” for “corps,” and a capitalized “Poles” that turned telephone poles into Polish people in a Second World War separation notice. Staff first took the changes for a transcriber’s odd choices; only a sentence about a veteran’s civilian job showed that dictation software was at work.
When machine-only is the right call
Machine transcription without volunteer correction is a sound choice in three situations. The first is printed or typed material in good condition, where recognition is often good enough for search and volunteer hours are better spent on handwriting. The second is a large backlog that nobody can search at all today: uncorrected text, clearly labeled, makes the pages findable while correction is planned. The third is material that volunteers should not see, such as closed records, where staff may use a machine draft internally and check what they need.
In each case, three conditions keep it honest. Label the text as uncorrected wherever it appears. Measure its error rate on a sample and publish the figure with the text. And keep the page image beside the text, so anyone who finds a page by search reads the original before quoting it.
How workflow design changes quality and speed
How volunteers see each other’s work matters as much as who they are. The best published evidence comes from a controlled A/B experiment on the Zooniverse platform, and it bears directly on what happens when volunteers see a machine draft.
The Anti-Slavery Manuscripts experiment
Blickhan and colleagues (2019), in the Journal of Data Mining and Digital Humanities, tested two workflows on Anti-Slavery Manuscripts, a Zooniverse project for letters held by the Boston Public Library, from January to September 2018. Identical sets of 2,173 letters went to two groups. In the individual workflow, several volunteers transcribed each line without seeing the others’ work, and software combined their readings. In the collaborative workflow, volunteers could see earlier transcriptions of a line and either accept one or enter their own.
The results favored collaboration on every measure the team reported:
- Quality. Compared with reference transcriptions of 19 sample pages, prepared by a content expert from the Massachusetts Historical Society, the collaborative results were closer to the expert text on all but five pages, a statistically significant difference.
- Output. With about the same number of volunteers, the collaborative group produced 25 percent more completed transcriptions.
- Speed. The collaborative workflow finished in October 2018, about ten months after launch; on that date the individual workflow was not yet half done. Zooniverse then moved the whole project to the collaborative method.
What the experiment says about drafts
One finding deserves attention before anyone hands volunteers a machine draft. When collaborative volunteers opened the list of earlier readings for a line, about 80 percent of the time they selected one and submitted it without editing. In that project the earlier readings came from other volunteers, who were usually right, so agreement saved time without lowering quality. A machine draft is different. A language model’s errors read fluently, and a volunteer who is inclined to agree will pass them through.
The authors also reported the cost side of an older project. Citing Transcribe Bentham’s 2012 results, they note that the time research staff spent moderating and editing volunteer work could, at the time, have produced more text if spent transcribing, and that “the need for editing considerably reduced the cost-effectiveness of crowdsourced transcription.” Volunteers are free; checking their work is not.
How machine drafts change volunteer roles
A machine draft turns volunteers from transcribers into correctors and reviewers. That can make a program faster, but only if the tasks are redesigned so that people still read the image rather than the draft.
Status you can see, line by line
FixIt+, the open-source correction platform GBH maintains for the American Archive of Public Broadcasting, shows the idea clearly. In the Library of Congress interview, a screenshot caption explains that gray lines are AI-generated text “that have not yet been checked” and green lines “have been reviewed and edited as needed.” Sorensen calls it a “human-in-the-loop” workflow in which “volunteers improve transcripts generated by computers instead of relying solely on the program’s output.” For handwriting, the same principle works at the line or field level: every unit of text shows whether a person has checked it.
Tasks that keep people looking at the page
- Narrow the task. “Check every name, place, date and number on this page against the image” can be done reliably. “Make this page perfect” drifts.
- Keep the original spelling. Tell correctors not to modernize spelling or expand abbreviations unless the project’s rules say so. FromThePage describes language models that “silently normalize irregular words,” and the Library of Virginia found abbreviations such as “co.” written out in full by dictation software.
- Mark uncertainty. Give a simple convention, such as [illegible] or [unclear: Halvorsen?], and accept it as a correct answer.
- Hide the draft for hard pages. For the worst hands, ask volunteers to transcribe from the image first, then compare with the draft.
- Seed known errors. Plant a few pages with deliberate draft errors in each batch and see whether correctors catch them. The catch rate is a direct measure of whether people are reading the image.
- Tier the work. The National Archives’ Citizen Archivist missions include a call for “expert transcribers” for records with tricky elements. Route the hardest pages and all final review to experienced volunteers or staff.
Keep the engagement in the work
A program that asks people only to approve machine lines risks losing what brought them. Volunteers in the By the People survey valued getting close to historical people and learning to read old hands. Keep some pages for full transcription, especially letters and diaries, and give correctors context: a short note on the writer, the place, the period, and the people who appear often.
Give volunteers a place to talk, too. In the Zooniverse experiment, volunteers in the individual workflow accounted for 62.8 percent of visits to the project’s discussion boards from the page header. The authors suggest that working alone can feel isolating, so people look for confirmation, while collaborative volunteers got feedback from seeing each other’s readings. Correctors working on machine drafts get no such feedback from the machine, so a forum, a regular online session or a named reviewer who answers questions does that job.
Writing conventions that people and machines can share
Every transcription project needs written conventions, and a hybrid project needs them more, because the machine’s habits and the volunteers’ habits must end up in the same text. Decide the rules before the first draft is made, and give volunteers a one-page version with examples from your own collection.
The Zooniverse team’s experience shows why the details matter. In Anti-Slavery Manuscripts, volunteers could mark insertions and deletions with tags such as [insertion] and [deletion], and a missing space after a closing tag was enough to scramble the combined text of a line. Blickhan and colleagues use it to stress “the importance of reviewing results before accepting crowd-generated transcriptions as complete or correct.”
| Question | A common choice | Why it matters with machine drafts |
|---|---|---|
| Spelling | Keep the original, including errors | Language models tend to modernize; the rule tells correctors to undo it |
| Abbreviations | Transcribe as written; expand only in a separate field or note | Expansions hide what the writer actually wrote |
| Line breaks | Keep them, or mark them | Line-level checking and IIIF display need them |
| Illegible or uncertain text | [illegible], or [unclear: reading?] | Gives machines and people the same way to say “not sure” |
| Struck-through and inserted words | A short tag for each, with spacing rules | Drafts often drop or merge them |
| Names, places, dates | Transcribe as written; add standard forms in the catalog record | Keeps the text faithful and the search consistent |
| Printed parts of forms | Transcribe or skip, decided once | Prevents half the pages doing it one way |
Write the conventions into the instruction or settings used for the machine as well, where the tool allows it, and test that the draft follows them on a few pages. The example collection shows labeled page text for a fictional box, which can serve as a model for your own guide.
Should volunteers use AI tools on their own?
Volunteers already ask. Set a written policy: either forbid personal AI tools in transcription, or allow them only with disclosure, and run any machine drafting centrally so it is labeled. Undisclosed AI output in a volunteer transcription is the worst outcome, because it carries a human label it has not earned.
FromThePage’s post makes the case against personal use from the operator’s side: institutions “have staff who can automate sending all of their documents to AI engines,” so pasting images into a chatbot one at a time wastes volunteer effort, and the result looks like careful human work when it may not be. The Library of Virginia’s experience shows that even ordinary dictation software changes text in ways reviewers did not expect. Its staff say they review every transcription before ingesting it into their digital collections, which is the safeguard that caught the errors.
A short policy is enough. For example: “Please transcribe from the image yourself. Do not paste images into chatbots or other AI services, and do not use dictation software unless you have told the project team. If a page is too hard, mark it and move on; experienced reviewers will take it.” Pair the rule with a reason volunteers can accept, such as protecting donors’ privacy and keeping the text trustworthy for researchers, and with a route for hard pages, so nobody feels pushed toward a shortcut.
There are also reasons outside accuracy. Donor agreements, privacy and community protocols may forbid sending images to outside services at all. Material held under community control, including tribal archives, should not go to any external model without the community’s agreement; our page on community archives describes working under such agreements.
Community and tribal collections
For material that belongs to a community, the question of who transcribes is part of the community’s decision, not only the archive’s. Community members may be the right transcribers because they know the names, places and language, and some material may not be transcribed or shared at all. Ask before planning any of the three models, keep such material off public platforms and outside services unless the community agrees, and record its decisions with the text.
Recording who or what produced each transcription
Every page of text should carry a record of how it was made and checked, so a researcher can tell volunteer readings, staff corrections and machine drafts apart years later. The National Archives shows one way to do it in a public catalog.
The National Archives’ Lifecycle Data Requirements Guide defines an attribution type called AI / Machine Generated Label. It “attribute[s] contributions created through the use of artificial intelligence and machine learning tools by NARA or a NARA partner,” and its stated purpose includes “transparency for AI / machine generated content.” The label can combine with other attributions: “tags and transcriptions could be AI / Machine Generated and donated by a NARA Partner.” The same catalog accepts contributions from volunteer Citizen Archivists.
For a smaller institution, a few fields per page or batch are enough:
| Field | Example values |
|---|---|
| Source of first reading | Volunteer; staff; machine |
| Engine or model, with version and date | The engine’s name and version, or the model and the instruction used |
| Correction level | Uncorrected; corrected for names, places, dates and numbers; fully corrected |
| Corrected and reviewed by | Volunteer or staff identifier, with dates |
| Measured error rate | Character error rate on a sample, with the date measured |
Decide in advance what happens when a person edits a machine line. A simple rule works: the line keeps its machine origin in the record, its status changes to corrected, and the person and date are added. Never let an edited line lose the fact that a machine drafted it, and never let an unedited line pass as checked because someone opened the page.
Keep the raw machine output alongside the corrected text, so later corrections can be compared and a better engine can be tested against the same pages. The companion article on labeling AI-generated metadata covers the same principle for catalog fields, and our AI policy sets out the labels we use on page text.
Example (fictional). The record for page ACHS-3-012, Ida Okerlund’s letter of April 12, 1893, reads: first reading, machine (HTR engine, version and run date recorded); corrected for names, places and dates by volunteer V-07 on March 3; reviewed by staff on March 10; batch character error rate after correction 1.9 percent on a ten-page sample. A researcher who quotes the letter can see exactly what kind of text she is quoting.
Choosing a model for each part of the collection
Choose by series, not for the whole collection. The kind of record, the hands, the volume, the sensitivity of the content and the use of the text each push the decision one way or another.
| Collection trait | Leans toward | Why |
|---|---|---|
| Printed or typed pages in good condition | Machine only, with sampling | OCR on clean print is often good enough for search; volunteer time is better spent elsewhere |
| Many pages in a few consistent hands | Machine draft, volunteers correct | A model trained or tested on those hands gives a strong draft |
| Many hands, poor condition, heavy abbreviation | Volunteers, or drafts hidden until after a first reading | Drafts on hard pages invite agreement with wrong readings |
| Names and places central to use (registers, ledgers, letters) | Hybrid with targeted correction | Names, places and numbers are where errors cost most |
| Small collection with strong community interest | Volunteers | Engagement and local knowledge are the main outcomes |
| Restricted, sensitive or community-controlled material | Staff or community members only, no outside services | Access rules come before efficiency |
| Audio and video | Speech-to-text draft, volunteers correct | The FixIt+ model; accents and local names need people |
| Text that will be quoted, published or used as data | Full correction and review, whatever the first reading | An uncorrected error becomes a citation |
Example (fictional). For Box 3, the Aster County Historical Society splits the work. The 41 letters (96 pages) get a machine draft and volunteer correction of names, places and dates, with two experienced volunteers reviewing. The Wren Gulch Mercantile day book (212 pages of cramped columns) goes to full transcription by volunteers working row by row, because tests showed the machine dropping rows. The clippings and the telegram (nine items) are printed or typed, so OCR alone serves, labeled uncorrected. Suppose the society later finds a closed file of county relief applications in the same accession: it stays off every platform and out of any outside service until the board decides its access terms.
Questions to settle before you choose
- Which series matter most to your users, and how will their text be used?
- Do any series carry access restrictions, donor terms or community protocols?
- How many volunteer hours a month can you count on, and for how long?
- Who on staff can train volunteers, answer questions and review batches?
- Have you measured a machine draft on your own pages, or only read about results elsewhere?
- Where will the final text live, and can that system show its source and status?
Running the program: platforms, staff time and costs
Whichever model you choose, plan for staff time, a platform that records status and exports cleanly, and a way to measure quality. Volunteer programs are not free, and machine transcription is not finished when the engine stops.
Platforms
Transcription platforms differ in task design, review and export. The ones that appear in this article illustrate the range. Zooniverse describes itself as a platform for “people-powered research” whose projects “combine contributions from multiple individual volunteers,” and offers a Project Builder for new projects. FromThePage hosts programs such as the Library of Virginia’s. The Library of Congress runs By the People on its own platform, Concordia, and GBH maintains FixIt+ as an open-source project. We do not rank them; ask each the same questions:
- Can it import machine drafts and show which lines or fields a person has checked?
- Does it support review by a second volunteer or staff member before text is final?
- Can volunteers mark uncertain readings in a consistent way?
- Which formats does it export (plain text, page-level files, structured data), and do they keep the status and contributor information?
- Can it open your images through IIIF, or do you upload copies?
- What are its terms on data retention and on any machine processing it performs?
Staff time
Volunteer programs need recruitment, training sessions, written conventions, answers to questions, moderation and final review. Established programs invest in the training side: the Library of Virginia offers monthly volunteer training, online or in person, and By the People runs virtual orientation sessions for new volunteers. A small program can do the same with an hour a month and a shared document of answered questions. The lag between transcription and publication can be long even at large programs: in November 2022, By the People reported over 580,000 completed transcriptions, of which over 146,000 had been integrated back into their collections on loc.gov. Budget the integration step, not only the transcription.
Machine transcription needs setup, a test on representative pages, measurement of error rates, sampling during production and labeling. The OCR and handwriting accuracy calculator turns error counts from a sample into an estimate of correction hours, which is the number that decides most budgets.
Measuring quality during the program
Quality needs to be measured during production, not only at the start. Three methods work at any scale:
- Reference pages. Keep a set of pages with careful staff transcriptions, slip them into batches, and compute the character error rate of whatever comes back, machine or volunteer.
- Agreement as a signal. Where several people read the same line, agreement shows where to look. In Anti-Slavery Manuscripts a line counted as complete when on average three or more transcribers agreed on its words, and the authors suggest “focusing effort on lines for which the consensus score is below average” to speed up review.
- Random samples. Check a random sample of every finished batch against the images, sized to the batch, and record the result with the batch.
Example (fictional). Suppose the Aster County Historical Society times its own volunteers on ten letter pages from Box 3. Transcribing from scratch takes them about 30 minutes a page; checking names, places and dates in a machine draft takes about 10. For the 96 letter pages, that is roughly 48 volunteer hours against 16, plus staff review of a 10 percent sample in either case. These numbers are invented for illustration. The point is the method: time your own volunteers on your own pages, with both approaches, before you plan the program around either.
A framework for the decision
The Library of Congress’s LC Labs AI Planning Framework moves through three phases, “Understand,” “Experiment” and “Implement,” and looks at the data, the models and the people involved. Its emphasis on people fits this decision well: the volunteers who will correct the drafts, the communities represented in the records, and the researchers who will rely on the text. Our page for historical societies describes how a small organization can scope this work.
What to do next
A short sequence will answer the question for your own collection:
- List the series you want transcribed, with page counts, kinds of hand, sensitivity and how the text will be used.
- Pick 20 representative pages per series, including the hardest, and make a careful reference transcription.
- Run a machine draft on the same pages and measure its error rate against the reference.
- Time a few volunteers transcribing from scratch and correcting the draft, and check what each method misses.
- Choose a model per series, write the correction conventions, and set a policy on personal AI tools.
- Record the source, engine, correction level and reviewer for every page, and keep the raw machine output.
- Measure again on a random sample during production, and whenever the engine or model changes.
If you would rather have the machine side measured on your own pages before deciding how to involve volunteers, a pilot box runs a small batch of your real material through recognition and reports the error rates, so your volunteers can start where their knowledge counts most.