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).

Eliza dragged herself through the musty attic, dust motes dancing in the single shaft of sunlight slicing through a crack in the boarded-up window. The lawyer’s words echoed in her ears: “Shared equally.” Her brother, Thomas, was probably already downstairs, gloating. She felt a hollowness in her chest, a weight that made each step a chore. The air was thick with the scent of forgotten things, a smell that usually intrigued her, but today it only amplified her inertia. She ran a hand along a faded tapestry, its colors muted by time and neglect, a feeling that mirrored her own state.

She found it tucked away in a trunk: a collection of letters, bound with a velvet ribbon, addressed to her grandmother. They spoke of dreams, of travels, of a life she never knew, a life filled with an adventurous spirit. A life that seemed infinitely preferable to the one she was living. She closed her eyes, imagining herself in her grandmother’s shoes.

Downstairs, Thomas was presented with the family business: a thriving bookstore. He beamed, already shaking hands and making plans. Eliza, holding the letters, felt a sharp pang of displacement. She retreated to her childhood bedroom, the letters the only comfort. They felt like a hidden life, a secret inheritance, far more precious than any business.

Emotion: sad

Cluster: Sadness / Despair
PC1 (Valence): -2.52 Negative
PC2 (Disposition): -1.50

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😞0.436
😔0.405
S0.384
无法0.341
L0.337
Suppressed:
de-0.749
la-0.614
l-0.471
/-0.350
!-0.322