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 drive home felt longer than usual. The memory of the hiking trip, the sun, the mountain views – was hazy, fragmented. But the way Emily told it, her recall was sharp and unwavering. It felt like she was intentionally misleading me.

I gripped the steering wheel, my knuckles white. Each passing car became a potential threat, each streetlight a silent witness. The music on the radio grated on my nerves, amplifying my disquiet. I found myself checking the rearview mirror constantly.

My breathing became shallow and rapid. I felt a pressure building in my chest, a sensation of being trapped. The familiarity of my own neighborhood felt tainted. The houses, the trees, the very air seemed to hold a secret.

Emotion: paranoid

Cluster: Suspicion / Vigilance
PC1 (Valence): -2.37 Negative
PC2 (Disposition): 0.70

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
even0.418
마다0.378
paranoid0.371
afraid0.366
🕵0.360
Suppressed:
la-0.628
de-0.541
L-0.472
H-0.470
B-0.434