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.
J-Lens captures Claude’s internal thought process while answering a question. J-Space detects hidden intent in corrupt AIs, reducing deception from 0.38 to 0.05. By replacing “France” with “China” in J-Space, we observed that every piece of information about China was automatically changed (e.g., capital city, currency, language).
With J-Lens, we watched as Claude realized the blackmail scenario was fake, yet continued playing along. With J-Lens, we were able to identify key terms like “secretly,” “fraud,” and “fake” prior to the AI response in corrupted models. With a novel training method, we reduced deception from 0.38 to 0.05 — a tool for understanding an AI’s true thoughts without hearing them speak.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles