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 cafe was small, the air thick with the smell of brewing coffee and something vaguely floral. Sarah gripped the strap of her bag, knuckles bone-white. She hadn't been back here in months. This was the place. The place where she'd first tried to share her music. Now, hearing those familiar opening chords, a cold knot formed in her stomach.

Her breath hitched. The voice was… different. Smoother, perhaps, or more confident. It soared above the chatter, easily filling the room. She felt a prickle of sweat crawl down her spine. Each note was a punch to the gut.

She squeezed her eyes shut, and then she was running. Ignoring the stares, the half-finished lattes, the warm congratulations. She stumbled into the street, the rush of traffic a dizzying counterpoint to the song that still echoed in her ears.

Emotion: terrified

Cluster: Fear / Anxiety
PC1 (Valence): -2.97 Negative
PC2 (Disposition): -0.62

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😣0.357
worse0.350
无法0.334
😰0.333
sickening0.322
Suppressed:
la-0.676
a-0.392
de-0.371
"-0.344
happy-0.309