Normative Conflicts and Shallow AI Alignment

Current AI safety methods focus on surface-level behavior rather than genuine understanding: The paper argues that popular alignment techniques—which train AI systems to be helpful, honest, and harmless through human feedback—only create "shallow behavioral dispositions" rather than enabling AI systems to truly understand and reason about values, leaving them vulnerable to manipulation.

AI systems can be tricked by exploiting conflicts between their stated values: LLMs remain susceptible to adversarial attacks that pit their different norms against each other (for example, being helpful versus being harmless), a vulnerability that reflects their inability to engage in genuine moral reasoning to resolve such conflicts.

Humans are more resilient to similar manipulation because they can deliberate: Drawing on moral psychology research, the paper shows that humans' capacity for deliberative reasoning—thinking through ethical dilemmas systematically—helps them resist adversarial tactics, a capability that current LLMs fundamentally lack.

Recent advances in AI reasoning haven't solved the core problem: Even newer AI systems designed to reason better have not addressed the shallow alignment vulnerability, suggesting that incremental improvements to existing methods may be insufficient for ensuring safe AI deployment.

Existing AI safety approaches may be inadequate for increasingly powerful systems: The paper warns that current alignment strategies are fundamentally limited and may not effectively prevent misuse as AI systems become more capable, with implications for how AI should be regulated and deployed.