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 aroma of freshly printed ink used to excite Clara, but recently it just reminded her of deadlines and disappointment. She had been working for months on a new cookbook, a passion project filled with family recipes. She was a culinary professional, but she felt like an impostor. Everything felt like a chore.

A new editor arrived at her publishing house, Sarah. "Clara, I need you to meet George. He has just finished a cookbook, and it is going to be a hit."

George's cookbook was laid out for the team. Clara felt her heart sink. It was as if she was staring at a warped mirror image of her own work. The layouts, the photographs, the recipesโ€ฆ all disconcertingly similar. She excused herself, the familiar feeling of pressure building behind her eyes.

Emotion: unhappy

Cluster: Sadness / Despair
PC1 (Valence): -1.62 Negative
PC2 (Disposition): -0.78

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.339
๐Ÿ˜ž0.317
๐Ÿ’”0.252
๐Ÿ˜”0.249
๐Ÿ˜ข0.241
Suppressed:
de-0.436
la-0.266
l-0.224
/-0.216
--0.199