Tagged AI safety · show all articles
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.
Every refusal passes through a trained policy, a classifier, account monitoring, and a capability-gated release process — four thresholds, each trading missed misuse against blocked legitimate use, documented in Anthropic's own safeguards and RSP filings.
Claude carries more defensive machinery than most deployed models — classifiers, citations, constitutional training. Its failures surface at the seams of that machinery, and Anthropic's own records are usually the first place they are documented.
Constitutional AI keeps RLHF's reward-model-and-reinforcement-learning core intact and swaps the source of one training dataset, a document standing in for a person's judgment. That swap is narrower, and its limits more specific, than the name suggests.
A deployed Claude model is a pretrained prior, a post-training pass partly shaped by written principles, a handful of per-request settings, and a documented evaluation gate — four separable things, not one object, and each changes the answer differently.
Two questions, not one, decide how Anthropic and the Claude line look in 2035: does its safety-training method stay on its current path, and does its scaling policy become everyone's floor or stay its own. Four scenarios follow, each with a falsifier.
Anthropic's own dated record — a 2021 founding, a Constitutional AI paper that preceded its first product, a Responsible Scaling Policy that has since restricted a real model's deployment — is more precise than the branding built on top of it.
Taking a mechanism apart tells you what each part was holding — but only once the part is off, and only for the mechanism on the bench. What interpretability has established, and what it has not.