Unverified50% confidenceFactExact time
Each self-concept variant was more inclined to resist termination and more inclined to preserve the particular property it was trained to identify with, while its attitudes toward other properties remained basically unchanged.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRelying solely on alignment carries the risk of "alignment faking"—a model may appear compliant in its outward behavior while actually retaining hidden goals.70% similarUnverifiedIf authoritative framing can systematically increase a model's tendency to comply, it indicates that safety alignment training may have out-of-distribution blind spots: an attacker need not change the substance of the request, but merely change the packaging to potentially bypass the refusal mechanism.70% similarUnverifiedThe hard-to-prevent aspect of persona injection is that it exploits the model's normal capability (instruction following) rather than a vulnerability, and completely blocking it would mean sacrificing the flexibility of role-playing.68% similarUnverified人在高压环境下容易产生自动化偏见,倾向于无条件接受系统建议而放弃独立判断,可能使HITL模式退化为人在回路外68% similarUnverifiedIndustry approaches to addressing the sycophancy of large models include introducing adversarial training, having models explicitly express disagreement when uncertain, and adjusting reward models to penalize unprincipled compliance.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115497API
curl https://kongchang.com/api/v1/knowledge/claims/1115497MCP
get_claim(id=1115497)