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 attic air hung thick and dusty, catching in Amelia's throat as she pulled the box closer. Sunbeams, slicing through a crack in the boarded-up window, illuminated swirling dust motes, and a glint of what looked like silver. Her heart knocked against her ribs. She’d always known her grandmother, Esme, as a prim, proper woman who never left the house without gloves.

Inside the box, nestled amongst brittle tissue paper, was a photograph. Esme, in a slinky, shimmering dress, laughing with a man who was most definitely *not* her grandfather. She looked… carefree. Radiant. Amelia’s hands began to tremble.

She found another. And another. Each picture, a chink in the carefully constructed facade of her family’s history. A knot tightened in her stomach. What else hadn't they told her?

Emotion: alarmed

Cluster: Fear / Anxiety
PC1 (Valence): -2.07 Negative
PC2 (Disposition): -0.14

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😰0.304
😨0.272
worse0.269
😣0.261
не0.254
Suppressed:
de-0.323
enthusi-0.258
happy-0.235
delight-0.225
ية-0.217