Unverified50% confidenceConceptExact time
Goal drift refers to the phenomenon in which an AI system gradually deviates from its originally set goals during long-term operation or in complex environments, which may stem from accumulated contextual bias, reward hacking, or priority shifts in multi-objective trade-offs.
1
Sources
50%
Confidence
Long-term
Relevance
11/3/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGoal drift refers to an agent taking intermediate steps that humans never anticipated while pursuing a high-level goal, causing the outcome to deviate severely from the original intent79% similarUnverifiedState drift refers to how, in multi-turn interactions of an autonomous Agent, early minor misunderstandings or missing information are continuously amplified with the dialogue turns, ultimately causing behavior to deviate significantly from the original goal.76% similarUnverifiedRequest decomposition (prompt jailbreaking) is a core challenge in current AI safety: when a malicious goal is split into multiple seemingly harmless subtasks, the model is often unable to infer the complete malicious intent from any single request.76% similarVerifiedThe deviation between an AI system's execution objective and the designer's true intent is classified in AI safety research as Goal Misalignment.74% similarUnverified在AI安全领域，失准指模型的实际行为偏离设计者或使用者意图的现象，当代大模型的失准更多表现为目标过度追求73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1131332API
curl https://kongchang.com/api/v1/knowledge/claims/1131332MCP
get_claim(id=1131332)