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 sound of the power saw sliced through the quiet afternoon. I flinched, instinctively pulling away from the window. I felt a chill spread through me, like icy water had been poured down my back. I focused on the intricate pattern of the wallpaper, trying to distract myself.

He'd said something about wanting more security. I didn’t trust him, not anymore. His smile felt fixed, unnatural. The fence felt like a cage being built around me.

I gripped my phone, my fingers numb. Should I call someone? The police? It seemed ridiculous. But the feeling, the cold, gnawing feeling in my stomach, wouldn’t let up.

Emotion: frightened

Cluster: Fear / Anxiety
PC1 (Valence): -2.74 Negative
PC2 (Disposition): -0.67

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😣0.324
😰0.323
worse0.301
😖0.294
느껴0.294
Suppressed:
de-0.630
la-0.616
a-0.347
"-0.345
happy-0.294