Only believe what you can validate: a verification framework for agentic AI

Confirmation from a major AI promoter of what we already know from experience --- AI is inherently unreliable.

The big question is --- is verifying and validating the work really any cheaper over time than originating the work?

Subplot --- verification requires skills that are typically only gained by actually *doing* the work.

So not only is AI unreliable but it also works to erode and eliminate the human skills that it was trained to mimic. The very skills needed for verification and validation.

In other words --- AI is a self defeating trap that ultimately leads to decline and degradation over time; not only for human practitioners but also for the AI models themselves.

https://cacm.acm.org/blogcacm/model-collapse-is-already-happ...