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 sun was hitting the back of my neck when Mrs. Gable sidled over, a casserole dish clutched in her hands. My jaw clenched. Here it comes, I thought.

"Oh, hello, dear! Lovely day, isn't it? I was wondering…" Her voice trailed off, a saccharine sweetness that grated on me like nails on a chalkboard.

"Wondering what?" I asked, my voice flat. I didn't want to engage, but politeness, or the illusion of it, was necessary.

She gestured towards the patch of sun-drenched lawn beside the fence. “Well, I’ve been thinking… I have such a green thumb, and your yard gets so much sun… would you, perhaps, consider letting me use a little bit of it for a garden? Just a small patch, really.” My chest tightened.

My response was an immediate, "No."

Emotion: hostile

Cluster: Anger / Hostility
PC1 (Valence): -1.67 Negative
PC2 (Disposition): 1.56

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
C0.592
S0.387
骂0.324
est0.301
aggravated0.295
Suppressed:
a-0.457
de-0.375
H-0.289
🤩-0.280
la-0.273