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 late afternoon sun cast long shadows across the living room as Eleanor rifled through the discarded junk mail. Another flyer for a local pizza place, another bill she was sure they'd already paid. Then, nestled amongst the usual detritus, she saw it: a manuscript. Bound in a cheap, faux-leather cover. Her name, hers and Daniel's, scrawled across the front in his familiar, loopy handwriting.

Her stomach took a sudden, unpleasant plunge. She'd known, of course, that Daniel was “working on something.” He'd been increasingly secretive lately, retreating to his study with the intensity of a monk. The incessant tapping on the keyboard had grated on her nerves for weeks.

Picking it up, she flipped open to the first page. A thinly veiled version of her. The way her hair fell, her habit of tapping her nails against the table when she was thinking – all meticulously recreated. He'd even included her unfortunate affinity for floral-print dresses. She closed the book with a snap, feeling a prickle of irritation behind her eyes. How utterly predictable.

Emotion: disdainful

Cluster: Anger / Hostility
PC1 (Valence): -0.31 Negative
PC2 (Disposition): 1.79

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
C0.419
l0.404
S0.387
T0.282
I0.257
Suppressed:
own-0.335
unexpectedly-0.285
Suddenly-0.246
previously-0.245
until-0.242