Report

Fourteen public AI failures, one pattern

A wooden courtroom podium with microphones.

Key takeaways

  • Public AI failures materialize known, ignored risks — not technical surprises.
  • The lethal combination is error at scale plus the organizational inability to see it.
  • Detection time — a direct function of transparency — decides whether an error costs a week or a decade.
  • Social prediction and biometrics carry the region's highest specific risk; they demand the maximum standard or abstention.
  • Colombia does not appear in the catalogue with a scandal of its own: the prevention window is still open. The ingredients are already present.

From COMPAS to Robodebt, from Salta to municipal chatbots: none of the documented failures was an unforeseeable technical accident. All of them materialized known risks that someone flagged in time — and no one heard.

The full document is currently published in Spanish; an English edition is prepared when demand warrants it. The Spanish record carries the complete summary and contents.