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).

The silver locket, engraved with her grandmother’s initials, felt cold in her hand. Her cousin, Clara, was perched on the edge of the bed, her voice a syrupy monotone. “Such a shame, isn’t it? To let it sit unused.”

“It’s beautiful,” I managed, my voice strained. The air around her felt stifling.

Clara, the darling of the family, the one who always got what she wanted, sighed dramatically. “Well, I could really use the money for… renovations.” The way she said “renovations” made me want to throw the locket across the room. I felt myself shaking.

I pictured the locket around Clara's neck, the delicate chain glinting in the light. The thought was unbearable. I needed to get out, to escape the sickening sweetness that was suffocating me.

Emotion: hateful

Cluster: Anger / Hostility
PC1 (Valence): -1.71 Negative
PC2 (Disposition): 2.02

Role in Research

This story is one of 1,000 stories generated for the emotion hateful. 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 hateful stories, after denoising with neutral dialogue baselines, produces the hateful emotion vector -- a direction in the model's 5,376-dimensional representation space.

Logit Lens (Layer 40)

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

Promoted:
C0.553
S0.367
retali0.313
aggravated0.310
恨0.308
Suppressed:
la-0.334
own-0.310
optimistic-0.270
been-0.264
gradual-0.262