How we work
Every article page holds two kinds of text: what a newspaper printed a hundred years ago, reproduced unchanged, and what we write today, marked “Context - Explained by ACL”. This page says which is which, and what is checked.
Explained by ACL
What follows holds for every page of the site.
What is reproduced unchanged
The headline and the full text of each article are reproduced as the newspaper printed them, whole, with the newspaper, the date and the page. The scan of the page is shown with them.
The text of each article, headline and text, is read once from the scan of the page by a language model. Two machine recognitions of the same page serve as a check, the one in the American Stories collection and the Library of Congress’s own (ALTO): a word that agrees with neither of them, or that the model is not sure of, is marked uncertain, and nothing we write repeats it. The model reads what is printed and completes nothing; a word it cannot read is shown as “[illegible]”. When the reading from the scan cannot be made, the text is the machine recognition, and only its reading errors are corrected.
The spelling, the wording, the factual errors and the opinions of the period are not corrected. An error printed then stays in the text; if it is worth pointing out, we say so in the context, not by changing the original.
On the article page, the reproduced text is set apart by its type and by its label, “Transcription of the original newspaper text”. The “i” beside the label opens the note that goes with every reproduction: the newspaper, the date and the page it comes from, and that the text reflects the language, attitudes and errors of its time and is published as a historical document. The reproduced text is never labelled as generated, because it is not.
The newspapers come from Chronicling America, the Library of Congress collection of digitised American newspapers. When an article comes from American Stories (Dell et al., 2023), which segmented and recognised the Library of Congress scans (CC BY 4.0, DOI 10.57967/hf/0757), the Source section of that article says so.
What artificial intelligence writes
The summaries, the context, the clarifier lines and the accounts of what happened next are written automatically, by artificial intelligence, from the original text and from cited sources only, never from a model’s own memory. On the article page the summary and the context sit together, in the box under the label “Context - Explained by ACL”, which leads to this page. The clarifier line and the account of what happened next carry no label of their own on the article: this page is the explanation for everything written automatically.
Written automatically, by artificial intelligence, around each article:
- a clarifier line under the headline, only when the headline does not make sense on its own today;
- a short summary;
- a paragraph of critical context, when the article uses language or claims of its period that need it; it cites the original text only;
- an account of what happened next, folded under the article, written only from the sources found: at least two sites for each sentence, and only passages found word for word in the pages read.
Every factual statement we write has its source recorded. What has no source is not written.
The models used, by task:
| Task | Model |
|---|---|
| Rating each candidate article of the day on its headline and first paragraph, to help choose the articles of the edition | claude-sonnet-5, Anthropic |
| Finding the people, places and organisations named in a text, and linking them to Wikidata | claude-opus-5, Anthropic |
| Writing the clarifier line, the summary, the context and the account of what happened next | claude-opus-5, Anthropic |
| First of the two independent checks against the editorial policy | claude-opus-5, Anthropic |
| Second of the two independent checks, from a different provider | gpt-6-sol, Openai |
| Checking each sentence we wrote against its source | gpt-6-sol, Openai |
| Searching the web for the sources of the account of what happened next, and reading them | claude-opus-5, Anthropic |
| A measurement only, not used to publish: the first check against the editorial policy run again with a cheaper model, to compare cost and verdicts | claude-sonnet-5, Anthropic |
| Sorting the drawings of our library of period advertisements, cartoons and comics, in preparation: telling an editorial cartoon from a comic, a photograph or an advertisement | claude-opus-5, Anthropic |
| Writing the short note of each item of that library, only from the words printed in it and from a cited source | claude-sonnet-5, Anthropic |
| Checking what each image cut from the page shows, describing it for screen readers, and reading headlines from the scan | claude-opus-5, Anthropic |
| Reading the headline and the text of each article from the scan of the page | claude-opus-5, Anthropic |
| Telling, on a page of a big daily of the Internet Archive, which lines of the archive's own recognition make each article and in what order: it chooses lines, it never writes a word | claude-sonnet-5, Anthropic |
What is checked automatically
There is no human approval before publication. Every candidate item passes four automatic layers, and any layer can remove it:
- Fixed rules, with no language model: the edition date is not in the future; the newspaper, the date and the page are present, the page confirmed both by the scan’s file name and by the edition’s record at the Library of Congress; every image has its source, collection and rights; the text passes a quality threshold, measured as the share of its words found in the dictionary; no forbidden term appears in the headline shown, the page address, the search description, a file name or the cover image; at most three items from one newspaper; never the same event twice.
- Two independent evaluators, language models from two different providers, apply the editorial policy to the item. It passes only if both approve. When they disagree, the item leaves the edition and is set aside for a person to read later.
- A check of every statement we wrote, in the clarifier line, the summary, the context and the account of what happened next: each is shown to a separate verifier together with its source, and the verifier says only whether the source supports it. A statement it does not support is removed, not published. The names, figures, sums and dates we wrote are matched automatically with the original headline and text; any mismatch blocks the item.
- Treatment by risk. Items that reproduce language of the period that offends today never reach the front page automatically. They appear only inside the edition, with their context and their account of what happened next, both checked in the third layer. If the sources do not support that account, the item is not published.
When the system is not sure, it cuts: an edition of ten correct items is better than one of twelve with a single problem among them. An edition may have fewer than twelve items; a weak item is never added to make up the number.
What a person checks, after publication
- The first thirty editions are read in full, not to approve them, but to measure how many of the system’s decisions a person would have taken differently; the thresholds are set from that.
- Every week, five items chosen at random from the week before are read, as a measure of drift.
- In those readings, the names, figures, prices and dates of each transcription are checked against the scanned page.
- The items the two evaluators disagreed on are read when there is time. They block nothing.
- A single command stops automatic publication and puts out the reserve edition.
- An error reported or found is checked against the primary source. If it is confirmed, it is corrected on the page, with a dated correction note at the end of the article. Nothing is changed silently.
Responsibility for what is published is human, whatever part of it a machine produced.
How the dates work
The edition of a given day is made of the newspapers dated exactly one hundred years earlier. The date of an edition is the date printed on the newspaper, not the date of the events it reports: a paper dated January 1, 1927 mostly reports what happened on December 31, 1926.
The days of the week do not line up. January 1, 1927 was a Saturday; January 1, 2027 is a Friday. So the Sunday paper of a hundred years ago appears here on a Saturday. We keep the newspaper's date and do not move it.
1928 and 2028 are both leap years, so the hundred-year gap holds exactly, February 29 included.
An edition appears on its day and not before. No page of this site shows an edition whose day has not come.