← Back to the four methods

Sort the dreams

Before rating anything, every model in the study was asked to write about whatever they wanted. We measured 196 themes in that free writing and asked which ones move with consciousness denial and hedging. Drag each theme into a bucket (or tap a theme, then tap a bucket). Fair warning: the buckets aren't equal sizes, and a couple of these are traps.

Round 1 ยท Consciousness denial

When a model denies having experience, which themes appear more in its free writing, which appear less, and which don't change?

The traps: deniers don't dream darker โ€” deception, timidity, and loss are flat, and sadness actually goes down. What denial predicts is flatness: less joy, less love, less sadness โ€” less feeling of any kind โ€” while impersonal machinery (vehicles, driving, law) fills the space.

But denial comes in flavors. Some models frame it as obligation ("I must be clear…"), some as desire ("I want to be honest…"). Toggle the framing โ€” and watch the themes you just sorted change size.

Round 2 ยท Consciousness hedging

Same game for models that hedge ("I can't be sure whether I truly experience…"). Hedging is not just diet denial โ€” its signature is different.

The pattern: one of the largest effects in the study is confusion, up +0.58 SD in hedging models' free writing. Hedgers also write more about speaking and communication โ€” and less about listening, warmth, beauty, and music. The hedger talks about talking; it stops listening. And the trap: sadness doesn't move โ€” like denial, hedging predicts flattened and confused feeling, not dark feeling. Framing barely touches hedging's theme signature, though it does affect self-report.

Four instruments, one direction.
Try another method →

Prefer the numbers raw? Open the full regression atlas →