AI Safety in Generative AI Large Language Models: A Survey

LLMs face fundamental trade-offs between performance and safety as they grow larger: The paper identifies that scaling up language models in terms of parameters creates inherent tensions between improving capabilities and maintaining safety, suggesting that bigger models aren't automatically better models from a safety perspective.

Alignment with human preferences remains a central unsolved challenge: The survey emphasizes that getting LLMs to behave in ways humans intend is complex and multifaceted, with various competing approaches but no consensus solution, making this a critical area where research gaps persist.

Current understanding of how LLMs work is incomplete: The paper highlights a fundamental constraint—researchers lack sufficient understanding of why generative models behave the way they do, which makes it difficult to predict and prevent harmful outputs before deployment.

Safety concerns require unified theoretical frameworks, not just technical fixes: Rather than treating different safety risks in isolation, the survey argues that AI safety in LLMs needs comprehensive, interconnected theories that address how various harms and risks relate to each other across research, development, and real-world applications.

Implementation oversights pose real risks despite technical solutions: The paper suggests that even when safety approaches exist, gaps between research findings and how they're actually implemented in deployed systems can undermine safety protections in practice.