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 attic air hung thick with dust and the ghosts of forgotten things. Sarah coughed, pulling a moth-eaten blanket tighter around her. She was supposed to be cleaning, a task she always approached with a heavy heart, her thoughts inevitably turning inward. She stumbled upon a trunk, its brass hinges green with age. Inside, nestled amongst yellowed tissue paper, were stacks of her old schoolwork.

A wave of heat washed over her face as she lifted a faded book report, her handwriting shockingly immature. *Poor* her parents had scrawled in the margins, the word seeming to leap off the page and punch her in the gut. Why would her mother keep *this*?

She slammed the trunk shut, a sudden urge to flee bubbling up in her chest. The attic became suffocating, the air too close, the past too present. She imagined her mother watching, smiling in a way that felt disapproving.

Emotion: self-critical

Cluster: Shame / Guilt
PC1 (Valence): -1.66 Negative
PC2 (Disposition): 1.28

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
L0.748
S0.662
ness0.425
ly0.374
탓0.351
Suppressed:
la-0.492
soon-0.304
(!)-0.301
secured-0.284
cautiously-0.280