I am dreading our LLM-written incident report future

A critical perspective on the emerging trend of using LLMs to fully write incident reports. While LLMs can help gather data for incident reports, having them generate the actual write-up bypasses the essential cognitive work of synthesizing evidence and confronting gaps in understanding. Unlike LLM-generated code (which can be tested) or AI SRE diagnoses (which either work or don't), LLM-written incident reports produce plausible-sounding but potentially inaccurate narratives with no obvious correctness test. The result is simulacra that look right but fail to deliver genuine organizational learning from incidents.