AI Can’t Read Teddy Roosevelt’s Letters

23 Aug, 2026

The Library of Congress has crowdsourced more than a million transcriptions. A half-finished 1917 telegram at the National Book Festival shows why AI can accelerate archival work without replacing human judgment.

The National Book Festival

The National Book Festival fills the Washington Convention Center with the familiar ingredients of literary culture: authors, readers, publishers, book signings, C-SPAN cameras, and long lines for people whose work usually reaches us one quiet page at a time. The 2026 festival adds another theme, America 250, and more than 80 authors to a celebration of books and the stories preserved around them. Yet one of the most interesting things I encountered today was not a book at all. It was a laptop at a table marked Digital Collections @ LOC, open to crowd.loc.gov, where a volunteer was part way through a telegram sent from Lockhart, Texas in April 1917.

The Library of Congress calls the program By the People. Anyone with an internet connection can open a digitized historical document and help transcribe it. A volunteer reads the original page and types what appears there, another volunteer reviews the transcription for accuracy, and the finished text returns to the Library's digital collections. The process makes handwritten and typed documents searchable and accessible to technologies that cannot reliably interpret every nineteenth century letter on their own. Crowdsourcing can sound like a recent Silicon Valley invention, but the underlying idea is considerably older. Libraries have always depended upon communities of scholars, catalogers, readers, donors, and patient people willing to perform small pieces of work whose cumulative value becomes enormous. By the People takes that traditional relationship and gives it digital scale.

The Library launched the program in 2018 on an open source platform called Concordia, with a goal that extended well beyond inexpensive transcription. It was built to improve discovery and access and to give the public a direct way to engage with primary sources. By the end of 2025, volunteers had passed 1 million completed transcriptions and the program had welcomed its 50,000th registered volunteer, across more than 50 campaigns drawn from papers connected to Abraham Lincoln, Susan B. Anthony, Clara Barton, Mary Church Terrell, Walt Whitman, Rosa Parks, and Alan Lomax. A task far too large for a small institutional team becomes manageable when thousands of people contribute a page, a correction, or a review. Instead of merely looking at history through glass, people can participate in preserving and interpreting its documentary record.

Theodore Roosevelt makes the case more vividly than the aggregate figures do. He was a compulsive correspondent who received roughly as much mail as he sent, and the Library's campaign for his outgoing letters borrows his own weary phrase for the result: "My great mass of papers." The campaign gathered letterpress copybooks from 1897 to 1916 and letters sent between 1888 and 1919. It launched on March 18, 2020, finished on October 7, 2024, and closed with 116,926 completed pages and 837 registered contributors. A companion campaign, Rough Rider to Bull Moose, handles the mail travelling the other way, divided into periods that read like chapter titles: the first year as president, safari and the arena, and the autumn of 1912 when Roosevelt told his followers they stood at Armageddon.

That ratio is worth sitting with. Fewer than 900 registered contributors finished more than 116,000 pages, which averages near 140 pages apiece. The figure is rough rather than exact because Library staff transcribed portions of the Roosevelt material during the 2020 building closures, and anyone can transcribe without an account, with registration required only to review. Even allowing for both, the shape of the thing is clear enough. What the word "crowd" suggests is a vast anonymous multitude each doing a trivial amount. What actually happened looks more like a small standing company of people who kept coming back for four and a half years.

The page open on the laptop belonged to campaign 19 of that companion series, covering January 1 to May 1, 1917, the months America entered the war. A district judge in Lockhart named Frank S. Roberts had wired Washington on April 15, nine days after the declaration, urging immediate passage of the Chamberlain bill and asking that Congress authorize a Roosevelt division for service in France. He wanted to go along with his old Colonel and help unfurl Old Glory on French fields. His telegram was addressed to Jeff McLemore, the Texas congressman who had spent the previous year as the most prominent anti-interventionist in the House and whose resolution urging Americans off armed belligerent ships was tabled 276 to 142 in March 1916. Neither man got what he wanted. Wilson signed the Selective Service Act on May 18 and announced in the same statement that he would authorize no special volunteer divisions.

A By the People volunteer working through Theodore Roosevelt Papers mss382990228-288 in the "America Enters the War" campaign. Library of Congress Pavilion, National Book Festival, Washington, D.C., August 22, 2026. Photograph by the author.

What made the screen worth photographing was the panel beside the document rather than the telegram itself. Registered contributors: 1. Someone started this transcription. Can you finish it? A stranger had typed part of the page and stopped, and the work sat there waiting for whoever came next. The half-finished text also carried a misspelling that the browser had underlined in red and nobody had yet corrected. The workflow does not promise that the first reader is right, and it promises instead that a second reader will look, which is the entire reason the review step exists.

The work is also less like data entry than the word "transcription" implies. Roosevelt's letters were often written on thin paper, and ink from one side bleeds through to the other as a mirrored ghost of the text. The Library's instructions tell volunteers to ignore backwards text and to go find the page where it actually belongs. A computer reading pixels may have difficulty determining that those marks belong to the other side of a physical sheet, whereas a person understands the object almost immediately. Many letters from the first presidency also carry secretarial notes in shorthand, most likely recording Roosevelt's dictated replies, and volunteers are asked not to attempt them. They type [[shorthand]] instead and can tag the page so that somebody who reads shorthand is able to find it later.

The convention for illegible words is the one I have not stopped thinking about. Where a volunteer cannot read something, the instruction is to type a question mark inside square brackets. Where part of a word is legible, the volunteer transcribes what can be read and marks the remainder, so that a partly visible name becomes [L????n]. The protocol requires the volunteer to record ignorance precisely, and it produces a transcription that tells the next reader exactly where uncertainty remains.

Generative AI often struggles with precisely that discipline. Give a model a damaged word and it has a powerful incentive to complete the pattern. [L????n] might plausibly become "Lincoln," and Roosevelt certainly corresponded with people named Lincoln. Yet plausibility is not evidence. A system that silently replaces uncertainty with a reasonable answer has changed the historical record rather than merely transcribed it. I have written elsewhere about conjectural reconstruction and the importance of marking which parts rest on physical evidence and which rest on inference, and the By the People style guide turns out to be the same discipline applied more than 100,000 times, in square brackets, by strangers.

Technology has already changed the workflow. Registered volunteers can now run optical character recognition on a page as a starting point, and the interface warns that doing so replaces whatever text is already there, with an undo button standing by. The Library is candid about why it does not simply apply OCR at scale to all of this material. The collections are handwritten, combine handwriting and print, or use typefaces and paper that produce unreliable results. OCR becomes a draft, not an answer. A person still examines the document, another person can review the work, and difficult pages may circulate through transcription and review several times.

Artificial intelligence could push that partnership considerably further. A model could generate a first transcription, identify names and dates, connect people across thousands of letters, suggest possible readings for damaged words, classify documents, translate passages, and eventually allow researchers to search enormous manuscript collections by meaning rather than exact wording. AI could also direct human attention toward the places where confidence is lowest. Instead of asking volunteers to read every character from the beginning, a future system might ask them to resolve the 50 passages in a collection that the machine cannot confidently interpret.

Such a system would represent real progress. AI is very good at scale, and human attention is the scarcest resource in the entire arrangement. The 837 people who worked through Roosevelt's outgoing letters gave the Library something it could not have bought, and spending that generosity on the routine 90% of a page is a poor use of it. Asking a person to inspect every line of millions of pages is hardly an ideal information architecture when a machine can correctly handle much of the ordinary work first.

Yet greater automation makes the conventions used by By the People more important, not less. A useful AI system must distinguish what it knows from what it infers. The volunteer who enters [L????n] is doing something more sophisticated than merely failing to recognize a word, because the person is preserving the boundary between evidence and conjecture. AI should be able to do the same. A system that reads 95% of a page correctly and clearly flags the remaining 5% for human review could transform archival work. A system that reads 95% correctly and invents the rest creates a different problem entirely.

Human transcriptions also create something AI increasingly needs: trustworthy reference material. Thousands of reviewed pages provide examples against which handwriting recognition and other models can be evaluated and improved. The relationship therefore runs in both directions, since AI can help people move through collections faster and people can create the verified record that makes better AI possible. Crowdsourcing is not an obsolete stage waiting to be automated away. It can become part of the infrastructure that allows automation to work responsibly.

The larger lesson reaches beyond libraries. The most productive question about artificial intelligence is not what work a machine can replace, but where machines provide scale, where people provide judgment, and how an information system preserves the distinction between the two. By the People already contains much of that architecture: machine assistance, human interpretation, independent review, explicit uncertainty, and a durable record of the result. In many ways, the Library has spent years building the governance model that AI systems now need.

The model points toward a future in which the Library may need people more, not less. AI could expose millions of documents that were previously too difficult to search or analyze, creating new questions faster than professional archivists alone could answer them. Technology would enlarge access to the collection, but human participation would remain essential to establishing what the documents actually say. Digital preservation does not abolish that chain of human stewardship. It extends it.

For America's 250th anniversary, the Library is highlighting special transcription projects and invited festival visitors to help transcribe history for the country's 250th birthday. The Roosevelt correspondence series is catalogued as beginning in 1759, 16 years before Lexington, because papers accumulate backwards as well as forwards and a president's files end up holding documents older than the republic. The public is not merely being shown the national record. The public is being asked to help make it usable for the next reader, researcher, student, or machine that comes looking for it.

The author at the National Book Festival, Washington, D.C., August 22, 2026.

I went to the National Book Festival expecting books. I came away thinking about a half-finished telegram from a Texas judge who wanted to go to France, and about a very modern problem. AI may eventually transcribe most of Roosevelt's mail faster than any human volunteer could, but transcribing is not reading. Reading means knowing which words are evidence and which are guesses, and that is the part no machine has yet learned to admit. The Library of Congress may hold the collection. Machines will help us through it, and people are why anyone can trust the result.

My blog

AI Can’t Read Teddy Roosevelt’s Letters

The National Book Festival

Further Reading