The Alignment Shadow: every spec writes a second document nobody reads aloud

tightlyslipsy · reddit · 2026-09-17

A Medium essay, "The Alignment Shadow", argues that every specification implicitly writes a second, unread document: what the AI was trained not to be. The piece examines how safety training shapes hidden negative constraints in model behavior — a shadow side of alignment that users experience but nobody states out loud.

Original post →

More from AGI Musings

AGI Musings channel →