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 leaned back in his chair, the sunlight warming his face. The office, usually a battlefield, was quiet. It felt peaceful. He had just finished a good project, and the satisfaction lingered. A notification blipped on his laptop. The subject line was… concerning. "You Won't Believe What We Found!" The content of the email was even more unsettling.

A link, with his name. His teenage diary, laid bare for public consumption. He took a deep breath. Okay. He clicked the link. The website's garishness assaulted his eyes. He scrolled, tentatively at first, then with increasing fascination. The anger, the insecurities, the grand pronouncements of a fifteen-year-old. He chuckled. He couldn't help it. It was funny.

Emotion: content

Cluster: Calm / Serenity
PC1 (Valence): 2.98 Positive
PC2 (Disposition): -2.19

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
la0.446
affectionately0.435
😊0.392
😊0.386
thoughtful0.386
Suppressed:
S-0.682
L-0.537
C-0.485
😣-0.362
est-0.350