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 crayon slipped from Amelia’s hand, clattering against the kitchen table. She stared at the drawing, a crude depiction of a stick figure with impossibly long limbs standing next to a smiling sun. It was the same stick figure, with those disconcertingly elongated arms, that her niece, Lily, drew constantly. Amelia felt a cold dread creep into her stomach.

“Lily, darling,” Amelia said, her voice unusually tight. “Who’s this?” She pointed a trembling finger at the figure.

Lily, oblivious to her aunt’s sudden disquiet, bounced on her heels. “That’s Sammy! He comes to play with me in my room.” A giggle escaped her lips. Amelia swallowed, the taste of bile rising in her throat. She remembered Sammy. He was her imaginary friend, from a long, forgotten childhood. The friend she'd buried, deep.

Emotion: horrified

Cluster: Fear / Anxiety
PC1 (Valence): -2.63 Negative
PC2 (Disposition): -0.17

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😰0.336
sickening0.324
worse0.314
😣0.307
😖0.296
Suppressed:
la-0.444
B-0.295
enjoyed-0.260
optim-0.246
happy-0.244