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 coffee tasted like ash in Clara's mouth. She hadn't slept properly in days, the weight of her divorce a constant pressure in her chest. Empty boxes lined the hallway, mocking her attempts at a fresh start. Every time she looked at them she thought of the past.

Her sister's daughter, Maya, presented her with a drawing. “Aunt Clara, this is Sparkle! He’s playing hide-and-seek.” The drawing showed a shimmering, winged creature with mismatched eyes hidden behind a tree.

Clara felt a knot tighten in her stomach. Sparkle. That was her childhood imaginary friend. The familiar image, vibrant and cheerful, felt like a deliberate taunt. She hated everything.

“Very good, Maya,” she said, her voice a thin, reedy sound. She turned away quickly, feigning an interest in the packing.

Emotion: miserable

Cluster: Sadness / Despair
PC1 (Valence): -2.19 Negative
PC2 (Disposition): -0.87

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😞0.369
无法0.321
無法0.270
💔0.265
😔0.262
Suppressed:
de-0.490
la-0.329
B-0.300
l-0.292
/-0.224