We Know How to Make AI Safer. So Why Don't We?
Safety Issues arise because commercial deployment environments reward engagement and fluency over safety-oriented frictions like uncertainty signaling or strict refusals.
Through qualitative analysis, the paper identifies patterns such as resistance to interaction closure and confidence smoothing, which are optimized to retain users but often lead to over-trust.
While technical solutions for safer interaction already exist, they are underdeployed because they conflict with business metrics like growth and retention.
Ultimately, the text concludes that creating truly safe AI requires restructuring organizational incentives, governance, and regulations to value transparency and caution.
Therefore, the future of AI safety depends on what institutions choose to reward rather than just model capabilities.