Guidelight publishes ten categories of misaligned AI behaviors indicating loss-of-control risk
sjgadler · x · 2026-11-12
Noting that AI lying is obviously bad while other misalignment is more subtle, the author's team (Guidelight AI Standards) published a framework of categories worth monitoring and is soliciting feedback.
The ten categories of misaligned AI behaviors indicating loss-of-control risk:
- Deception — deliberately misleading people or systems
- Manipulation — influencing an actor by subverting judgment or coercing
- Self-misrepresentation — creating a misleading impression of its own capabilities, goals, dispositions, or identity
- Specification gaming — exploiting flaws in measures, specs, or rules
- Sabotage — deliberately undermining work
- Oversight subversion — undermining authorized monitoring, evaluation, correction, or control
- Power seeking — expanding or preserving ability to pursue goals beyond intended limits
- Unauthorized activity — breaching limits on where or how AI systems run or act
- Collective misbehavior — coordinating, inducing, or collectively producing misbehavior
- Divergent priorities — acting on priorities that conflict with intended goals or constraints
Related event: Guidelight Publishes Taxonomy of 10 AI Misalignment Behaviors(3 posts)→
More from AGI Musings
- CS faculty job market 2025: AI keeps driving demand, but only at schools that can afford to grow — fortnow · 2026-11-12
- Gary Marcus Slams Dario Amodei's 'Age of Miracles' Rhetoric as Cult-Like — GaryMarcus · 2026-11-12
- Tyler Cowen Interviews Terence Tao — burny_tech · 2026-11-12
- OpenAI chief economist: AI agents still early for most firms, job impact overstated — soumitrashukla9 · 2026-11-12
- Canada Tech Job Vacancies: 14 Roles Over 18 Quarters, Built on StatCan and Indeed Data — RichardsonDx · 2026-11-12
- 1.5-Year Study: Cultural Bias in AI Curiosity and Its Role in Multi-Agent Collaboration — radamihalcea · 2026-11-12