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 was thick with the scent of dust and forgotten things. Clara hummed as she carefully sorted through a box overflowing with yellowed photographs. Sunlight streamed through the single grimy window, illuminating dancing motes in the air. She’d been doing this for hours, the gentle rhythm of the task a comfort.

A photograph surfaced, its edges curled and cracked. It depicted her grandmother, but not the stern, stoic woman Clara had known. This one smiled, a wide, uninhibited grin, her arm looped around a man who wasn't her grandfather. Her stomach lurched, but she took a deep breath, smoothing the photo with a trembling finger.

“Well, this is… unexpected,” Clara murmured, the words barely audible. She tucked the photo away, continuing her meticulous search. A slow, steady exhale escaped her lips.

Emotion: safe

Cluster: Calm / Serenity
PC1 (Valence): 1.86 Positive
PC2 (Disposition): -2.55

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
soon0.402
affectionately0.357
disfrut0.329
необходи0.325
🥰0.317
Suppressed:
S-0.474
C-0.416
L-0.408
worse-0.379
even-0.370