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).

Her voice was trembling as she explained her error to her boss. She’d made a mistake that could cost the company a lot of money, and the weight of her error pressed upon her like a physical burden. She felt hot and exposed, as if everyone in the office could see through her.

Later, as she sat alone, nursing a headache, Ben from the HR department approached her. “Rough morning, huh?” he asked quietly.

She simply nodded, her cheeks burning with embarrassment.

"Camp Silverlake? Man, that was a whole different kind of mess," Ben said, and the tone of his voice was full of empathy.

She finally looked up, finding a moment of shared experience.

Emotion: sorry

Cluster: Sadness / Despair
PC1 (Valence): -2.02 Negative
PC2 (Disposition): -0.36

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.604
L0.549
😞0.382
😔0.364
ness0.324
Suppressed:
de-0.675
la-0.536
(!)-0.323
!-0.298
l-0.288