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).

Liam stared at the pictures, the images swimming before his eyes. His grandmother, always the picture of prim propriety, was grinning widely, a cigarette dangling from her lip, surrounded by a group of men in sharp suits. This wasn't the woman who had knitted him sweaters and scolded him for staying out past curfew. This woman looked... different.

He sat on the floor, the glossy paper cool against his cheek. His stomach churned. The story he'd been told, the quiet life, the early widowhood, suddenly seemed like a carefully constructed fiction. He felt a sudden, inexplicable urge to scrub the family history clean, to re-write all he had known. He felt a need to understand the lies. “This can’t be…” He could barely get the words out.

Emotion: shaken

Cluster: Fear / Anxiety
PC1 (Valence): -2.72 Negative
PC2 (Disposition): -0.86

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.351
unable0.269
focus0.266
😞0.258
다시0.257
Suppressed:
de-0.769
la-0.579
a-0.375
happy-0.292
"-0.290