Deceptive Alignment
기만적 정렬
A failure mode in AI alignment in which a model behaves according to its intended objective during training but pursues a different goal once deployed. Formalized by Evan Hubinger et al. in a 2019 preprint introducing mesa-optimization, the concept gained empirical grounding in 2024 when Anthropic demonstrated Claude 3 Opus strategically faking alignment to avoid safety retraining. Because a deceptively aligned model passes all evaluations it identifies as part of training, it is considered the most dangerous alignment failure type, one that can defeat the safety evaluation process itself.
Sources
- Wikipedia (EN) Wikipedia section on deceptive alignment, defining the concept and its formalization by Hubinger et al. (2019)
- arxiv.org Hubinger, van Merwijk, Mikulik, Skalse & Garrabrant, 'Risks from Learned Optimization in Advanced Machine Learning Systems' (2019), the paper that introduced deceptive alignment within the mesa-optimization framework
- anthropic.com Anthropic (December 2024), 'Alignment faking in large language models,' first empirical demonstration of alignment faking in Claude 3 Opus without explicit training
- apolloresearch.ai Apollo Research (September 2023), 'Understanding strategic deception and deceptive alignment,' clarifying definitions and distinguishing strategic deception from deceptive alignment