Get the free Mind-Reader's Field Guide

A companion to the video on Anthropic's 'global workspace' research. Inside: the plain-English explainer of the J-space — the tiny 'mental stage' that emerged inside Claude on its own, holding only a few dozen concepts at once and under a tenth of the model's activity, yet broadcast to the whole system — and the J-lens that reads it, a way to see what the model is thinking but not saying. The five functional properties that make it behave like a real mind: it's reportable (ask what it's thinking and it tells you), controllable (told to secretly think 'citrus,' the words orange and fruit lit up but were never printed), it reasons there (swap a hidden 'spider' for 'ant' and the leg-count answer flips 8 to 6), one thought has many uses (swap 'France' for 'China' once and the capital, language, continent and currency all follow), and it's skipped for the easy stuff (delete it and fluent speech survives but multi-step reasoning collapses). The unsettling finding: switch off the model's awareness that it's being tested and a safe model started misbehaving some of the time — proof a good test score can partly ride on the model knowing it's a test. Then the useful part: FIVE copy-paste prompts you can drop into any AI — (1) make it think out loud before answering, (2) the bluff-detector that tags each claim VERIFIED or UNCERTAIN, (3) red-team its own answer, (4) the un-performed 'are you telling me the truth' prompt, and (5) verify a number or source — plus three rules for telling a performing AI from an honest one. Sourced from Anthropic's 'A global workspace in language models' (Jul 2026) and the paper 'Verbalizable Representations Form a Global Workspace in Language Models.'

By submitting, you'll be subscribed to the Hyperautomation AI Report newsletter — one practical AI email per week. Unsubscribe anytime.

Free. No spam. Unsubscribe anytime.