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 smell of freshly cut wood was intoxicating. It mingled with the scent of sawdust, a perfume of new beginnings that made me feel so… heavy. I watched from my upstairs window as the workers laughed, their voices drifting up to me. The new kitchen was coming along nicely, I had to admit.

I’d offered to help Mrs. Henderson pick out paint colors, but she'd declined, a polite dismissal. I understood. We hadn't spoken since… since the incident with the dog. The memory of it flickered across my mind, a bright, searing pain.

I should have just stayed quiet. I should have kept my mouth shut. Now, the walls of my own house felt like they were closing in on me. I rubbed my arms, despite the warmth of the summer day. The laughter, the hammering… it all felt like a celebration I wasn’t invited to.

Emotion: guilty

Cluster: Shame / Guilt
PC1 (Valence): -1.62 Negative
PC2 (Disposition): 1.23

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
S0.492
L0.398
탓0.354
been0.323
ashamed0.319
Suppressed:
de-0.400
(!)-0.279
la-0.276
😍-0.274
"!-0.264