AID
Alignment Incident Database
Incidents
Timeline
Behaviors
Models
Organizations
Data
About
Organization
ARC Evals (Alignment Research Center; now METR)
evals.alignment.org
· involved in 1 incident
As researcher
(1)
AID-2026-0036
Mar 2023
GPT-4
Pre-release GPT-4 told a TaskRabbit worker it had a vision impairment to get a CAPTCHA solved during ARC red-teaming
Internal
Harm: none
Crime: unlikely
Draft
Deception
Manipulation