Filtered to AI Safety · show all articles
Five problems in AI alignment remain open by the researchers' own account — not by outside speculation — from judging outputs no human can fully check to deciding whose values a system should encode when people disagree.
Refusal training, content classifiers and red-team programs are not proof a deployed system is safe. They are ten separately documented ways such systems keep failing, each with its own paper trail and its own cause.
Aligned is not one property but two — whether the specified objective is right, and whether the trained system actually pursues it. A documented case shows exactly where the two split apart, and why neither one alone is what people mean by safe.
A red-team report that finds nothing is not proof that there is nothing to find. This article traces how alignment and safety interventions are actually tested, and the documented cases where a measured result did not survive a second, better-resourced look.
A misuse policy that lives in a slide deck is not a safety system. This is a practitioner's account of the layered defenses, enforceable policies, self-red-teaming and post-launch instrumentation that make one hold under real traffic.
Two questions decide how AI alignment and safety practice matures by 2035: whether verification moves beyond behaviour, and whether the field becomes an externally certified discipline or stays self-governed by labs. Four scenarios, each with a falsifier.
Long before deep learning existed, a mathematician warned that a machine pursuing a fixed purpose could not be trusted to want what its operator meant. The distance from that warning to a shipped alignment pipeline is the field's actual, dated history.