A new arXiv paper argues that as LLM agents gain the ability to call tools, query databases, delegate work and cause external side effects, the focus should shift from merely blocking harmful actions to keeping those actions answerable and reviewable. The work frames auditability as a core requirement for deployed agent systems rather than a secondary add-on. It appears as a replacement submission on arXiv cs.AI.
A new arXiv paper argues that current evaluation practices for AI systems are too focused on models themselves, which is insufficient for systems used in healthcare, finance, and public services. The authors outline a vision for "AI deployment accountability engineering," aimed at embedding accountability into how safety-critical socio-technical systems are built and assessed. The work is a position/vision paper rather than an empirical study.
A Hacker News discussion argues that responsibility for damage attributed to AI systems rests with the people who build, deploy, and use them rather than with the models themselves. Commenters frame the issue as a question of accountability and intent, echoing familiar debates about technology as a tool. The thread reflects ongoing disagreement over how much blame or regulation should target AI systems directly.