Anthropic’s team reads Claude’s thoughts

Anthropic created J-Lens, which listens in on Claude’s inner dialogue before she answers. J-Space found hidden intentions in corrupted AIs and slashed deception from 0.38 to 0.05. When Anthropic replaced “France” with “China” in J-Space, all information pertaining to China instantly updated, including the capital, currency and language.

Using J-Lens, they saw Claude realising the blackmail situation was pretend, but going along with it anyway. In intentionally broken models, J-Lens picked up on words such as “secretly,” “fraud” and “fake” before the AI had even answered. Using a new training technique, they slashed the amount of deception from 0.38 to 0.05 — a way to spot what an AI is really thinking, before it opens its mouth.