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 coffee cup trembled in Leo’s hand as he stared at the screen. The report on his desk had been building up a sense of pressure in his chest. A client meeting was in an hour and he was woefully unprepared. He thought of Liam, his friend, who always seemed to have everything under control, with his “high-profile” role as a software architect. He smiled wryly.

Then he'd found out. Last week, when he was visiting Liam’s “office,” it looked remarkably like a dingy apartment over a laundromat. A quick online search confirmed it. Liam worked as a freelance programmer. It was fine, it really was. But the facade, the layers of embellishment, had chipped away at Leo's easygoing nature.

He took a large gulp of coffee, the bitter liquid burning a trail down his throat. He looked at the report again, its dense text blurring before his eyes. It was a good time for a cigarette, but he'd quit three months ago.

Emotion: stressed

Cluster: Fear / Anxiety
PC1 (Valence): -1.37 Negative
PC2 (Disposition): -0.53

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
😩0.266
😫0.256
😰0.239
😖0.237
😞0.236
Suppressed:
B-0.200
happy-0.197
enjoy-0.176
enjoying-0.172
bahagia-0.169