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).

Dust motes danced in the single shaft of sunlight slicing through the cluttered attic. Thomas, a meticulous model train enthusiast, had always kept his hobby private, a sanctuary from the world. Now, his daughter, Maya, age eight, had taken up model trains and was hosting "themed" train shows for her friends.

He clutched his chest, a strange heat flooding his face. He'd come to inspect his original, custom-built steam engine, now the centerpiece of Maya's elaborate, holiday-themed layouts. He imagined the "Oh my gods" of her friends.

A giggle echoed from the stairwell. Peeking his head up, he saw Maya, followed by a gaggle of giggling girls, already ascending the creaking steps, their eyes gleaming with anticipation. He quickly ducked behind a moth-eaten tapestry, his heart rate spiking.

Emotion: embarrassed

Cluster: Embarrassment
PC1 (Valence): -1.08 Negative
PC2 (Disposition): 1.10

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
l0.420
尴尬0.341
C0.333
culp0.262
🤦0.256
Suppressed:
own-0.321
আনন্দের-0.230
seemingly-0.227
і-0.225
আনন্দে-0.224