About · AI use
Where AI is used, and where it is not
This archive is built with substantial AI assistance. This page says exactly where, which models, and what to do when you find a mistake.
Step by step
| Step | Model involved? | What happens |
|---|---|---|
| Acquisition | Partly | Documents are located with model assistance and by hand, and retrieved by script from courts, agencies, RECAP, and Congress.gov. Every retrieval is logged with its source URL and date. |
| Preservation and hashing | No | The file is stored as retrieved and its SHA-256 recorded. |
| OCR | No | Scanned documents go through Tesseract. No language model touches the text. |
| Text extraction | No | Text is pulled from the document's own text layer or from the OCR output. Extraction is not generation: nothing is added, rewritten, or "cleaned" by a model. |
| Titles | Partly | Raw filing names are normalized by script. A model proposes a title where the script cannot; a human-maintained override table supplies the final title, and each override is checked against the document, its docket entry, or its first-page heading. |
| Cross-references | Partly | Models propose links between documents. The "More filings in this case" lists on filing pages are built from the record facts, not by a model. |
| Descriptions | No | The one-line lede on a filing page and the one-line description on a source-document page are assembled from the record's own fields. |
| Summaries | Yes | Where a document page carries a summary, a Claude model drafted it from the document's text on that page (Claude Sonnet 5.5 writes the current ones). A script checked every date, number, name and docket cite in the draft against that text before it appeared, and samples of the summaries are audited by a separate model. The summary is labeled on the page. A summary is a navigation aid, never a source: cite the document. |
| Articles, profiles and interactives | Partly | Drafted with model assistance from the archive's documents and data, with every citation linking to the document it rests on. Each new piece, and each major update to a profile, is read and approved by the editor before it publishes. |
| Guides | Partly | The program guides are drafted with model assistance from the archive's records and checked against the pages they link to before they publish. |
| Verification passes | Partly | Models help run the checks (link targets, sourcing, page structure), and OpenAI's Codex runs some independent second-pass reviews. Scripted gates decide whether a page ships; a model's opinion does not. |
| Fact-check | Partly | Every day a slice of the live pages is checked in rotation: a model sets each page's statements against the documents it cites, and each error found is logged for correction. |
| Numbers | No | Figures come from the underlying data files and the documents, never from a model's memory. Every count on the program guides was measured on the archive's own files when the page was built. |
The documents themselves are never AI-generated.
Which models
Anthropic's Claude models do most of the model work. Claude Opus 5.5 handles titling, cross-referencing, drafting and the verification passes; Claude Sonnet 5.5 writes the document summaries; Claude Fable 5.1 is used only for narrow judgment calls. OpenAI's Codex (GPT models) runs some independent second-pass reviews, and overflow work may run on Google's Antigravity models (Gemini, and Claude through Antigravity). Earlier stages of the build used the Claude 4 generation. A model's draft reaches the site only after the scripted checks pass, and new pieces only after the editor approves them.
If you find an error
Go to the corrections page. Send the page URL, the passage, and the document that shows it is wrong. If the error is in extracted text, the page will be re-processed. If it is in a guide, an article, a profile or a data table, the page is fixed.