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).

"Mom, can we talk about what you did in fifth grade?" Sarah, her daughter, asked, picking at her fingernails. They were in the car, and Claire felt a prickling heat beginning in her chest. She had to take a deep breath to stop her face turning red. Mrs. Peterson, the woman who had been her own fifth-grade teacher, was now teaching Sarah.

“What do you mean, honey?” Claire asked, trying to sound casual, but her voice was tight.

“Mrs. Peterson said you were… a bit of a chatterbox.” Sarah kept looking at the window. Claire cleared her throat. She found herself trying to recall every instance, every little misdemeanor, every bad grade. It was a blur of memories, and most of them were embarrassing.

“Oh, well, you know, kids are kids,” Claire said, forcing a laugh. She gripped the steering wheel a little tighter. The car ride felt like an eternity.

Emotion: self-conscious

Cluster: Embarrassment
PC1 (Valence): -1.03 Negative
PC2 (Disposition): 0.52

Role in Research

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

Logit Lens (Layer 40)

Tokens promoted/suppressed when the self-conscious vector is projected through the unembedding matrix.

Promoted:
S0.307
C0.298
ness0.273
尴尬0.263
😶0.242
Suppressed:
que-0.399
de-0.266
triumphant-0.266
আনন্দের-0.247
ايا-0.238