AI systems often refuse reasonable requests too conservatively. A post-processing tool could rewrite these rejections into helpful responses.
Offer an API that takes AI outputs and applies safety-preserving smoothing. Charge per thousand processed responses.
Begin with simple rule-based adjustments. Later incorporate more sophisticated models.
Balancing safety and usefulness is tricky. Start with low-risk use cases.