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 late afternoon light slanted through the living room window, illuminating the dust motes dancing in the air. Sarah picked at a loose thread on the floral couch, her gaze fixed on the bright pink drawing her niece, Lily, had proudly presented. It depicted a creature with mismatched wings and a single, enormous eye. Something about the drawing sent a cold finger down her spine.

“That’s… interesting, Lily-bug,” Sarah managed, her voice tight. The little girl bounced excitedly, chattering about the “Fun Flier” and its adventures in a world made of candy. Sarah recalled the drawings of *her* imaginary friend. A creature, oddly similar, named Zizzle, with an exact same mismatched wing and eye.

She excused herself to the kitchen, seeking the familiar solace of a hot cup of tea. The kettle whistled a mournful sound as she poured the boiling water into her favorite mug, a chipped ceramic one from her grandmother. The memory of Zizzle felt like a phantom limb, a forgotten comfort, and its reappearance in her niece's artwork unearthed feelings she hadn't touched for years.

Emotion: troubled

Cluster: Fear / Anxiety
PC1 (Valence): -2.20 Negative
PC2 (Disposition): -0.81

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😞0.343
S0.317
ness0.302
😔0.294
😢0.272
Suppressed:
de-0.553
la-0.511
l-0.307
/-0.260
(!)-0.236