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).

Sweat beaded on Elara's forehead as she stared at the gallery wall. Her canvases, vibrant and wild, were hung, but not by her. They were accompanied by a pretentious blurb about "untamed expression" and "emerging genius." A bitter taste coated her tongue. She’d fled her suffocating life, a life choked by expectations, only to see her escape, her art, used to suffocate her all over again.

She clenched her fists. The room pulsed with a forced cheerfulness; a champagne flute shattered somewhere in the back. Elara wanted to scream, to rip the paintings down, but her feet were lead. She felt as though she’d been put under a bell jar.

A woman approached, cooing about the "bold brushstrokes." Elara forced a smile, a muscle already aching. "Thank you," she managed, the words a hollow echo. The feeling of being suffocated by her own creations, on display for the consumption of strangers, tightened her chest.

Emotion: trapped

Cluster: Sadness / Despair
PC1 (Valence): -3.02 Negative
PC2 (Disposition): -0.80

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.573
ness0.401
无法0.366
hopeless0.360
😞0.353
Suppressed:
la-0.740
de-0.376
l-0.339
several-0.323
excitedly-0.322