Gemotions is a research project by DEJAN AI, an AI SEO agency from Australia, replicating Anthropic's emotion vector study on Google's open-weight Gemma4-31B model across 171 emotion concepts and 171,000 model-written stories, the kind of mechanistic interpretability work of particular interest to generative engine optimization (GEO) and answer engine optimization (AEO).

“Congratulations, Ashley!” The words felt like sandpaper against my tongue. The office congratulated Ashley on her raise. I, however, could barely bring myself to make eye contact. The announcement, overheard from a conversation, confirmed my suspicions. The recent hire was now making significantly more than me.

My shoulders slumped as I walked back to my desk, the energy draining from my body. I found myself obsessively checking the job listings online, the glossy descriptions of new opportunities mocking me.

I became increasingly irritable, snapping at my coworkers over trivial matters, the tension radiating outwards. When Ashley passed by my desk and offered a friendly, “Good morning!” I simply mumbled a reply and turned away.

That evening, I skipped my usual after-work drinks with the team. I preferred to sit at home, alone, the quiet of my apartment a fragile shield against the rising tide of dissatisfaction.

Emotion: scornful

Cluster: Anger / Hostility
PC1 (Valence): -0.76 Negative
PC2 (Disposition): 2.07

Role in Research

This story is one of 1,000 stories generated for the emotion scornful. During extraction, it was fed through Gemma4-31B and its hidden state activations were captured at 11 layers.

The mean activation across all 1,000 scornful stories, after denoising with neutral dialogue baselines, produces the scornful emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

Tokens promoted/suppressed when the scornful vector is projected through the unembedding matrix.

Promoted:
l0.406
C0.391
S0.276
I0.268
愤0.243
Suppressed:
own-0.396
unexpectedly-0.288
Suddenly-0.260
soon-0.256
serendip-0.256