Dictionary entry Trending
AI control
also AI control research or control protocol
Definition of AI control
-
: an approach to AI safety that seeks to prevent harmful outcomes through external constraints, monitoring, security mechanisms, and limited trusted oversight even when a capable model is treated as potentially misaligned or intentionally subversive
- The AI control protocol routed suspicious code to a trusted reviewer instead of assuming that the stronger model would always act honestly.
- Researchers tested the control system against an attack policy designed to hide a backdoor from both the monitor and the human auditor.
Two meanings, one word
In everyday English
The ability to direct, restrict, or stop how a system behaves.
In AI
An approach to AI safety that seeks to prevent harmful outcomes through external constraints, monitoring, security mechanisms, and limited trusted oversight even when a…
Origin & history
AI control became a distinct research approach with the 2023 paper AI Control: Improving Safety Despite Intentional Subversion, which tested safety protocols under the assumption that a powerful untrusted model might actively try to defeat them. The field later expanded through ControlArena, work by the UK AI Security Institute and Redwood Research, and safety cases examining whether monitoring and safeguards remain effective against subversive behavior.
Editorial references
Sources & further reading
- AI Control: Improving Safety Despite Intentional Subversion↗ arXiv · 2023-12-12
- Intro to Control↗ ControlArena / UK AI Security Institute
- Our first phase of embedded assessments↗ OpenAI · 2026-11-13
Test yourself
Which of these is the meaning of AI control?
Cite this entry
"AI control." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/ai-control/. Accessed 16 Nov. 2026.