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).

He’d always considered himself a level-headed man, David. Yet, the receipt clutched in his trembling hand felt like a lead weight. The expense report, neatly printed, detailed a training seminar. He’d never heard of this seminar. He’d certainly never been invited. The amount, however, was clearly listed. And… it was more than his raise this year.

He found himself constantly “accidentally” bumping into Mark, the new analyst, in the break room. The young man, always impeccably dressed, responded to David's greeting with a cheerful, if somewhat distant, "Morning, David." The casualness of it grated on his nerves.

He started scrutinizing Mark's work, finding a misplaced comma, a minor formatting error. Each tiny flaw served to fuel a growing unease. He felt a phantom pain in his jaw, a muscle twitching with every passing hour.

Emotion: suspicious

Cluster: Suspicion / Vigilance
PC1 (Valence): -0.82 Negative
PC2 (Disposition): 0.82

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
🤔0.420
🤔0.415
🧐0.362
疑0.352
suspicious0.352
Suppressed:
and-0.307
alleviate-0.301
ได้รับการ-0.298
allevi-0.285
mouthful-0.284