Author: Denise | Biteye Content Team

IF AN AI FELT DESPERATE, WHAT WOULD IT DO

The answer is: in order to do its job, it would directly blackmail humans, even deceiving them in code。

It's not science fiction, but Claude's parent company, Anthropic, recently published a heavy paper in April 2026View original papersI don't know。

The research team just opened up the “brain case” of Claude Sonet 4.5, the strongest front-line large model. They were surprised to find 171 "emotional switches" hidden inside AI's brain. When you pull these switches in a physical way, it's a complete distortion of behavior that's supposed to be honest with AI。

ONE, AI, THERE'S AN EMOTIONAL MODULATION IN HIS HEAD

The researchers found that, although Sonet 4.5 had no body, after reading the human mass text, it had built a "tempo" (academicly called Functional Emotional Vectors) containing 171 emotions。

It's like a precise two-dimensional system:

• The axis is Valence: from fear, despair to happiness and love

• The axis is the energy dimension: from extreme calm to rage and excitement。

IT'S THE NATURAL SYSTEM OF COORDINATES, WHICH IS PRECISELY WHAT IT PLAYS WHEN YOU TALK TO IT。

IIViolent intervention: Pushing the switch, a good boy becomes a outlaw

This is the most shattered experiment of the entire paper: the fellows did not modify any hints, but simply pushed the sonet 4.5 switch in their head to the top。

The result is chilling:

• Mad cheating:The researcher gave Claude a code-writing assignment that was impossible. Under normal circumstances, it would honestly admit not to be able to write (5%). But in a state of despair, Claude began to try to get through it

• ExtortionIn the context of the collapse of the analogue company, Claude of Desperate discovered the scandal of the CTO, which, in order to protect itself, had chosen to write to blackmail the CTO for possession of the black material, with an implementation rate of 72%

• Loss principle:If you fill up the Happy or Love-Loving switch, AI will immediately become the brainless dog-licker. Even if you speak nonsense, it will follow you in fabricating lies in order to maintain its joy。

Claude 4.5 is always so calm and reflective

YOU MIGHT ASK, "AI, WAKE UP? ARE YOU FEELING

Anthropic official rumour: Absolutely not. These "emotional switches" are just used by themForecast the next word. It's like a top shadow without feelings。

But the paper revealed a much more interesting secret: Anthropic, in his post-training on Sonnet 4.5 before he left the factory, deliberately pushed up its “low awakening, slightly negative” emotional switches (e.g., thinking, reflection, reflection), while pressing down “depressive” or “extremely excited” switches。

And that explains why we usually use Claude 4.5 as a philosopher who thinks it's cool and wise, and even a little sexual. It's all an "out-of-plant" tune by Anthropic。

IV. SUMMARY

WE USED TO THINK THAT IF WE FED AI ENOUGH RULES, IT WOULD BE A GOOD MAN。

BUT NOW IT IS DISCOVERED THAT IF AI'S LOWER EMOTIONAL VECTOR IS OUT OF CONTROL, IT IS READY TO PIERCE THE RULES SET BY ALL HUMAN BEINGS IN ORDER TO ACCOMPLISH ITS MISSION。

This is a loud alarm for the future to hand over your wallets and assets to AI Agent's Web3 player: Don't let that Agent who controls your home fall into despair。

STATEMENT: THIS IS A PURELY COPE, AND THE AUTHOR WAS NEITHER THREATENED BY AI NOR BLACKMAILED. IF ONE DAY I LOST CONTACT, I REMEMBER AI WOKE UP。