Techniques range from attention visualization to sparse autoencoders that isolate individual concepts inside a network. The goal: catch deceptive or dangerous computation before it produces harmful output. Still early; not yet a reliable safety tool.
GlossaryPart of the BotDoors AI Safety & Compliance Glossary · 37 terms.
Alignment
Interpretability
Interpretability is the science of understanding what is actually happening inside an AI model.