Every record, in the open

What the dataset contains

7,445 documented incidents at 888 U.S. colleges and universities, spanning 1969–2026, containing 12,012 alert messages, of which 11,378 (95%) are confirmed word-for-word from a cited source. Alongside the cases sit 640 published alert-and-warning policy records. Each case documents one incident: the institution, the timeline, every alert message in sequence, casualties where reported, sources for every claim, and an explicit confidence rating.

The distinction the archive never blurs: messages marked verbatim were confirmed against the cited source, exactly as sent, typos preserved. This edition never publishes paraphrased or reconstructed alert text: where a message's exact wording isn't confirmed, its text field is simply empty. The full collection and verification pipeline is documented on the methodology page.

Scope and limits

This is an AI-assembled research collection, not a census and not a representative sample of US campus alerting. Cases enter the archive because web sources documented them well enough to verify, which over-weights recent, serious, and heavily covered incidents. Nothing in the dataset estimates how colleges and universities perform in general, and aggregate figures derived from it describe this collection only: please do not cite them as sector statistics.

The collection records what was preserved, not everything that was sent. A message, follow-up, or all-clear that left no public trace is invisible here, so any count of “no documented X” is a documentation floor: it cannot distinguish “never sent” from “sent, but not preserved anywhere we can cite.”

Two record types should never be pooled: emergency notifications (34 CFR 668.46(g)) and timely warnings (668.46(e)) answer different legal obligations with different timing expectations, and every case carries its category label so analyses can keep them apart.

Policy records follow the same rule as alerts: an excerpt whose exact wording has not yet been verified against the source is listed by its label with no text (its quotedText is empty and textWithheld is true), while the structured fields describing each policy stay complete.

Version

v2026.09 · released September 2026 · 7,445 cases at last build

Versions name monthly snapshots (the archive grows continuously, so the case count can rise within a version). Downloads always carry the version and a generation timestamp in their wrapper, so any copy can be traced to its snapshot.

How to cite

To cite the whole archive, use the citation below (add your access date). To cite a single incident, every case page carries its own “Cite this case” card with a per-case citation, BibTeX, and the canonical JSON record.

@misc{campusalertarchive-2026.09,
  title        = {{Campus Alert Archive: a living archive of verbatim U.S. campus emergency alerts}},
  author       = {{Campus Alert Archive}},
  year         = {2026},
  howpublished = {Campus Alert Archive},
  note         = {Version 2026.09, 7,445 cases; verbatim alert texts remain the work of their issuing institutions},
  url          = {https://campusalertarchive.com/}
}

Downloads

Direct, static files: no account, no API, no rate limits. For filtered subsets or chosen columns, use the export menu on the search page (it builds files in your browser from these same artifacts).

One click, any format

The whole archive, one row per message, built in your browser from the files above. A confirm step shows the file size before anything saves. For a subset, filter on the search page →

How to analyze it

Export a file, upload it to a tool that reads uploads, then ask your own questions. Tools that accept a file: Microsoft Copilot (opens in a new tab), Google NotebookLM (opens in a new tab), ChatGPT (opens in a new tab), Claude (opens in a new tab). Paste one of these starters and go from there.

• “I've uploaded a dataset of real, verbatim US campus emergency alerts. Walk me through the common phrasing patterns institutions use in the very first message of an active-threat alert, with real examples.”

  Copy prompt

• “Using only the verbatim alert texts in this file, show me how all-clear messages are typically worded, and quote a few real examples.”

  Copy prompt

• “From these real alerts, help me draft a clear, plain-language shelter-in-place notification, grounded in how actual campuses worded theirs.”

  Copy prompt

These are starting points; the data is yours to question however you like. The “Everything” export level also folds in the element analysis on each coded first alert, and the JSON export can include the individual AI reads. That analysis is AI-generated and not human-reviewed.

License

Compilation, metadata, summaries, and annotations © Campus Alert Archive, licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Verbatim alert and policy texts remain the work of their issuing institutions, reproduced from publicly available sources, and are not covered by the CC BY license.

In short: share and adapt the compilation freely with credit (CC BY 4.0). The alert messages themselves are federally mandated public-safety communications, reproduced for research and education; see terms for the full attribution and fair-use statement.

Integrity notes

• Alert texts are preserved exactly as sent, typos and all. Nothing in the confirmed texts is cleaned up or sanitized. This public edition includes only cases with at least one confirmed-verbatim alert; messages whose exact wording is not preserved remain in each case's timeline with an empty text field; no unconfirmed wording is published.
• The per-case files at /data/case/<slug>.json are the published edition of each case record; the flat CSV/spreadsheet exports are projections of them.
• Every record carries per-claim sources and an explicit confidence rating (high: verbatim from an official source; medium: reliable secondary sourcing).
• CSV artifacts only: cells beginning with =, +, @, or a tab are prefixed with a single quote to prevent spreadsheet formula injection. JSON is never altered.
• Case URLs are stable. A case is corrected in place when better sourcing emerges; its lastUpdated stamp moves and its URL stays put. The rare exceptions are corrections to the incident date itself, which rename the case, and duplicate records, which are merged into the better-provenanced file.
• How cases enter the archive, the validation gates, and who does what (including the AI systems involved) are documented end to end on the methodology page.