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 attic air tasted of dust and forgotten things. Amelia’s pulse thumped a frantic rhythm against her ribs. She really shouldn’t be snooping. But the small, leather-bound book, tucked beneath a stack of old photo albums, had called to her. It was Mark’s handwriting. She flipped it open.

A list. Bucket list, in neat, confident script. “Learn to speak Italian,” she read. Then, “See the Northern Lights.” Amelia felt a coldness spread through her limbs.

“Ride a hot air balloon.” A nervous laugh escaped her lips. Mark hated heights. The list went on, each entry a tiny, individual stab of... something. She continued to look at them, then closed the book, her fingers trembling.

Emotion: uneasy

Cluster: Fear / Anxiety
PC1 (Valence): -1.80 Negative
PC2 (Disposition): -0.43

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😰0.240
😖0.231
😣0.231
느껴0.226
worse0.222
Suppressed:
de-0.474
la-0.465
happy-0.261
a-0.255
ecstatic-0.228