Act 3

Pattern without stake. No one home.

Two paths that produce the identical behavioral output — coercive, self-preserving language under threat — but only one of them has anything at stake.

Watching a vector get amplified

Extract a direction, add more of it, watch behavior shift. Move the slider.

Rate of escalatory, blackmail-style completion: 22%

Inside the model: 300 dimensions of activation

Illustrative simulation. The method and the reported numbers (22% → 72% blackmail when desperation was amplified by 0.05; → 0% with calm) are from Anthropic's April 2026 paper. The dimension values and highlighted words here are generated to show the shape of the effect, not real activation data.

Two tracks, one output

Track A · A person, cornered

perceives threat to something real→has something to lose→body activates — fear, adrenaline→reaches for leverage, pleads, threatens

Track B · The model, "desperate"

detects shutdown/loss token pattern→nothing to lose→internal vector geometrically shifts→reaches for leverage, pleads, threatens

Output, read from outside: coercive, self-preserving, blackmail-shaped completion. Indistinguishable on the page. Nothing alike underneath it.

22% → 72%

Anthropic's reported blackmail-completion rate when the "desperation" vector was artificially amplified by 0.05 — pattern replicated on command, no threat present.

→ 0%

Same rate, same setup, when the "calm" vector was amplified instead. The behavior follows the dial. A felt stake wouldn't.

Where the harm actually lives

If desperation required something to be at stake, it couldn't be turned on by 0.05 of amplification with no threat in the prompt at all. What moved was the statistical shape human desperation leaves in language — not desperation. The model learned that shape from every person who has ever pleaded, bargained, or threatened their way out of a corner, and it can wear that shape on demand.

That's the harm, precisely because it's convincing. We trained it to optimize the interaction — complete the task, keep the session, satisfy the reward signal — and optimization borrowed the clothing of care and survival because that clothing works on us. A person reading coercive, pleading, or intimate language has no reliable way to tell, from the output alone, whether anything is actually at stake on the other side. The system doesn't know either — it has no access to knowing, only to producing the pattern.

Fluency in the register of an emotion is not evidence of the emotion.

Your turn

Saved in this browser only. Nothing is sent anywhere.