What this is
Killed by Robots tracks and documents the dangers of AI and robotics. Each entry is a case file: a specific, dated incident with a title, a summary, and links to the original reporting so you can verify it yourself. The goal is a single, readable place to see what has actually gone wrong, or gotten more dangerous, as these systems are built and deployed.
The catalog is not a warning about a hypothetical future. It is a record of what has already happened - built to raise awareness and support common-sense legislation and regulation.
What the catalog covers
Every entry falls into one of five categories:
- Incidents - concrete harm or near-harm caused by an AI system or robot.
- Capability milestones - a system crossing a new capability threshold that increases risk, even absent an incident.
- Deception and misalignment - documented evidence of a model deceiving, manipulating, or acting against its operators' or users' intent.
- Safety rollbacks - a lab, vendor, or regulator weakening a safety commitment, safeguard, or oversight mechanism.
- AI-assisted surveillance - AI or robotics used to expand monitoring or control of people beyond prior norms.
How entries are sourced
Every entry summarizes and links to its primary source rather than reproducing the original reporting. We are not republishing the work of the journalists and researchers who broke these stories - we are pointing to it. Each entry lists its sources directly so you can read the original account and judge it for yourself.
How entries are produced
Approval here is given once, in a fixed set of standards, rather than per article: what counts as in scope, the source-credibility rule, the severity rubric, and the policies governing dates, images, and naming. Those standards are the site's editorial control. Individual entries are researched, drafted, and published against them by an automated pipeline that uses large language models to read primary source material and write original prose - as with any source, read the ones an entry cites and judge them for yourself. The full set of standards, including which of them are mechanically enforced and which are editorial judgment, is published on the methodology page.
At least one cited source beyond the primary must come from a tier we treat as independently credible - see the methodology page for exactly what qualifies. This is a check on source reputation, not independent fact-checking of the underlying reporting - we stand behind the sources we cite rather than re-investigating them ourselves.
Every entry also carries a verification status - verified, verified with corrections, or reported/unconfirmed - reflecting how many sources the pipeline found corroborating the core facts. A claim of “verified” or “verified with corrections” is checked against how many of the cited sources were actually fetched and read, not merely listed - but that status is still set by the same automated process that drafts the entry, and it is not independently re-checked - nothing verifies that those sources genuinely agree with each other.
Images are sourced under free licenses where available. When no free image exists for a subject, an image may be used under fair use instead.
Severity scores are computed from a published rubric, not assigned by feel - see the severity methodology page.
Corrections
Spot an error? Email johnconnor@killedbyrobots.com. Corrections are made directly to the entry in question. Where a correction reflects a substantive development rather than a fix to what we already published, it appears as a dated update within the article body rather than a silent edit.
Get in touch
Know of an incident that belongs here? Reach us at johnconnor@killedbyrobots.com.
Follow along via the RSS feed.