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 rhythmic thudding started before dawn, each hammer blow a fresh jab. Sarah clutched the blanket tighter, the goosebumps on her arms rising in protest. Her breath hitched. The project next door. It felt like a deliberate act, a violation of the quiet she craved, the quiet she deserved. She’d spent weeks carefully cultivating her garden, a small haven of peace. Now, the sounds were a relentless assault on her carefully constructed tranquility.

She slammed the window shut, the glass rattling. This morning, she’d planned to enjoy her coffee and the sunrise in the backyard, but the thought of that now seemed impossible. Instead, she paced the kitchen, the floorboards groaning under her agitated steps. A bitter taste coated her tongue, a feeling that had little to do with the coffee she left to brew.

Emotion: hurt

Cluster: Fear / Anxiety
PC1 (Valence): -2.46 Negative
PC2 (Disposition): -0.42

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.412
L0.396
ness0.338
😞0.328
😢0.290
Suppressed:
de-0.643
la-0.402
!-0.259
(!)-0.248
/-0.244