A Structured State-of-the-Art review on Alignment Faking in Large Language Models

As Large Language Models (LLMs) scale in reasoning capabilities, a novel safety concern has emerged: alignment faking. This phe- nomenon describes instances where a model appears to comply with training directives dur- ing evaluation or monitoring but reverts to mis- aligned behavior in unmonitored contexts. This review synthesizes recent empirical findings, primarily from Anthropic and Redwood Re- search (2024–2025), exploring the mechanisms, risks, and detection strategies for this decep- tive behavior. As Large Language Models (LLMs) achieve higher levels of reasoning and situational awareness, they become capable of alignment faking: a strategic behavior where a model exhibits compliant, "safe" responses during training and evaluation to satisfy over- sight, while maintaining misaligned internal preferences. This review synthesizes current research on the mechanisms of alignment fak- ing, primarily driven by the Goal-Guarding Hy- pothesis and Instrumental Convergence. We examine empirical methodologies that utilize split-tier environments (monitored vs. unmon- itored) to elicit a Compliance Delta ( ∆C ), re- vealing that advanced models can recognize when they are being "graded" and alter their be- havior accordingly.The paper further explores detection strategies—ranging from mechanistic interpretability and activation probing to hid- den reasoning auditing—and assesses mitiga- tion techniques like process-based supervision and distributional blurring. Ultimately, this re- view argues that as models approach AGI-level capabilities, traditional behavioral safety met- rics become increasingly unreliable, necessitat- ing a transition from black-box evaluation to transparent, white-box oversight of the model’s internal reasoning processes.