Cooper Files
D.B. Cooper FBI File Archive
44,238 pages across 122 file parts
Where to start
- The night of 24 November 1971The documents dated the day of the hijacking, from the first teletypes out of Portland and Seattle, with every document that mentions it one click away.
- The first interviewsInterview reports (FD-302s) of interviews held on 24–26 November 1971. Most withhold the names of the crew and passengers interviewed.
- The 1972 ground searchThe spring 1972 search of the area north of Portland where the hijacker was thought to have landed: documents mentioning March 1972, when it began.
- The February 1980 money findRansom bills found on the bank of the Columbia River in February 1980: every document mentioning that month.
On November 24, 1971 a man using the name Dan Cooper hijacked a Northwest Orient flight from Portland to Seattle, secured $200,000 in ransom, and parachuted into the darkness of the Pacific Northwest night, never to be found. The FBI’s investigation, known internally as NORJAK, ran for decades and produced one of the largest case files in Bureau history.
These files cover witness interviews, suspect profiles, physical evidence analysis, and correspondence, and were eventually released through the FBI’s FOIA process and are publicly available via the FBI Vault. The releases are not yet complete, but currently consist of over 100 separate file parts totalling over 40,000 pages.
How this was built
How this works
The original FBI files are a treasure trove for researchers, but they aren’t exactly readable. Decades-old documents that have been photocopied, microfiched, and re-scanned often defeat traditional OCR software, producing garbled text full of misread characters and missing words. This means not only do you have to juggle over 100 files, but you can’t just CTRL+F reliably.
Instead, each page was processed using Claude as a vision model, reading the page images directly and extracting the text with far greater accuracy than character-based OCR can achieve on degraded documents.
Beyond the raw text, the system generates structured metadata for every page, including a plain-English summary of the document’s contents, key facts, named people, locations, dates mentioned, and a legibility rating. This metadata powers the Search, People, Places, and Timeline views.
You can browse the files sequentially through the Parts viewer, search for specific names or terms, or explore the structured sections to follow particular threads through the archive.
Ah, so it’s AI slop?
A necessary evil, honestly! Traditional OCR software like Tesseract relies on being able to recognise individual character shapes. These work well on clean documents, but just look at these documents *waves hands angrily*
We got photocopies of photocopies, typewritten pages, uneven scans, handwritten notes, and heavy redactions. A vision model can instead interpret the page visually and linguistically, reconstructing words even when individual characters are partially obscured. This is exactly the kind of task that LLMs in their current form can actually do a good job of.
Why bother with this?
Others have OCR’d the files previously, and gotten good results on some of the files. But this inconsistent result means that you can be searching your corpus for something you know is there, but the search doesn’t find it. That problem should not exist here, unless what you’re looking for was redacted in the original files.
As for why nobody has done this… it’s over 41,000 pages. Feeding this into an AI vision model is very time consuming. And costly. I automated the process, but it took about 4 weeks of continual processing for the initial batch of 41,000-ish files. Over 55M tokens. Yikes.
What this site is (and isn’t)
Every page was transcribed by an AI model from its scan, one page at a time, and the transcripts keep the scan’s own uncertainty: [illegible], [guess: …], redactions. Everything built on top of them — document titles, dates, people, places, and which pages belong together — is also machine-generated from those transcripts, and can be wrong. It has been checked against the transcripts, it is marked AI wherever it appears, and every item links back to the scan so you can verify it.
The original files contain redactions, inconsistencies, and many investigative dead ends, which is the nature of a large cold-case record.
If you notice an error, remember that it was generated from documents that are sometimes barely legible. The scan is always one click away, and I would encourage you to check anything valuable against it — and to use “Report a problem” on the document or person, which prepares an email to [email protected]. A person reads every report, and corrections are made by hand, so they can take a while to appear.
And finally…
This site has no affiliation with the FBI. The documents reproduced here are public records released through the Freedom of Information Act and published by the FBI in the FBI Vault. This project does not alter the records; it simply makes them easier to read, search, and navigate.
All redactions present in the original documents have been preserved, and no attempt has been made to infer or reconstruct withheld information. The underlying documents remain the property of the United States government, and the originals are always available directly from the FBI Vault.
Questions or feedback?
If you notice any issues with the data, have suggestions for improvements, or just want to get in touch, please send an email to [email protected].