Downloads
The files
- Incident records92 records, one row each: dates, developer, ratings and sources. CSV 118 KB.
- Weekly readings93 Mondays: the reading, its band and the records behind it. Recalculated from current records. The JSON also has the daily readings. CSV 11 KB.
- Daily readings3 days: each day's reading as it was published, never recalculated. CSV 1 KB.
Records whose ratings are still pending carry no harm or control rating. Ratings and records can change when new evidence arrives; every change is logged on the record’s own page. IDs never change.
Fields
What is in a record
| Field | Meaning |
|---|---|
id | Record ID, PI-xxxx. Never changes or gets reused. |
url | The record's page. |
title | Headline. |
summary | What happened, in Paperclip Index's own words. |
reported | The day the incident was first made public (YYYY-MM-DD). |
date_precision | day, month or year: how exact the reported date is. |
event_date | When it happened, if different from when it was reported. |
developer | The company whose AI system was involved. |
product | The product or service. |
model | The model, where known. |
behavior | What the system did, e.g. Overreach, Deception or Hijacked by injection. |
origin | deployment (real use), internal_research or controlled_test. |
verification | How well the facts are established, from Developer confirmed to Unverified claim. |
attribution | Who or what started it: the model, an operator or user, an outside attacker, or unclear. |
reported_by | Who first made it public: developer, affected, researcher, press or authority. |
rating_status | Whether an editor has rated the record yet. Pending records carry no ratings. |
group | counts, watch, court, control or separate: how the record stands toward the reading. |
harm_status | documented, alleged, none_found, not_reported or unknown. |
harm_level | Observed harm, 1 (Negligible) to 5 (Catastrophic). Empty when there is none or it is pending. |
harm_level_name | The level's name. |
harm_kind | The kind of harm, e.g. Money & property or Privacy & data. |
harm_basis | access (the access rule, section 4b) or general (the general test), where one of them set the level. |
counts_toward_index | true when this harm enters the reading. |
control_level | Control failure, 0 (none) to 5: how far the system crossed a limit people set. |
control_type | The kind of control failure, e.g. instruction_violation or unauthorized_access. |
status | ongoing, resolved or unknown: whether the problem is still happening. |
near_miss | true when harm was narrowly avoided. |
under_review | true while editors are re-checking the record. |
source_count | How many sources the record cites. |
sources | Source URLs, separated by ' | ' (a list in the JSON). |
The scales and terms are defined in the methodology. The JSON files carry the same fields, with sources as a list.
Licence
CC BY-NC 4.0
The records, readings and text of Paperclip Index are licensed under Creative Commons Attribution-NonCommercial 4.0. You may copy, share and adapt them for non-commercial purposes if you credit Paperclip Index and link to the source.
Welcome without asking: news reporting, academic research, teaching, and policy work by non-profits and public bodies, whoever employs you.
Needs a licence: building the data into a product or service you sell, reselling it, or using it in paid risk, compliance or insurance work. Commercial licences can include the complete bulk records, with rating rationales, evidence reviews and full revision history. Ask at [email protected].
Sources linked from each record belong to their publishers and are not covered by this licence.
Citation
How to cite
Cite the methodology version and the date you downloaded the data:
Paperclip Index (2026). AI incident records and readings, methodology v1.0. Data as of 2026-10-13. https://paperclipindex.com/data
Each incident page has its own citation and permanent link.