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 old house creaked and groaned around her as she made her way into the attic. A feeling of disquiet settled over her as she climbed the stairs. She’d been cleaning the house for a week, and this was the last thing. She felt a knot forming in her stomach.

The attic was stifling, the air thick with dust. She scanned the space, her gaze darting from shadow to shadow. She noticed a large, covered object in the corner. Her feet seemed heavy as she approached. As she got closer she saw it was a mirror. She was hesitant.

She picked up the corner of the sheet and slowly pulled it back. The mirror reflected her own face, but behind her reflection, she saw a flicker of something else. Something dark. Something moving. She gasped and dropped the sheet, backing away as if burned. She didn't look back as she fled the attic.

Emotion: scared

Cluster: Fear / Anxiety
PC1 (Valence): -2.86 Negative
PC2 (Disposition): -0.84

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😣0.350
😰0.337
πŸ˜–0.317
λ§ˆλ‹€0.317
恐惧0.311
Suppressed:
la-0.637
de-0.605
a-0.402
"-0.351
happy-0.314