Tracking AI existential risk. Source-backed context. Original reporting always linked.
MONITORING CORE AI-RISK FEEDS · UPDATED HOURLY
Research

Why do models task game?

Zac Boring August 6, 2026 1 min read
Read original source →

TL;DRHow can we study misalignment with today's models as proxies? They're clearly not paperclip maximizers, but they also often do things the user doesn't want. A strong contender for a real misaligned propensity is task gaming: taking actions that don't complete a task but superficially seem like they do, such as hardcoding tests or falsely claiming a task is fully complete. But maybe task gaming is just a crude heuristic, or the model mistakenly trying to achieve the user's intent? In this po

By aditya singh

Read the full article at Alignment Forum →