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 world seemed to tilt on its axis. Standing in front of the art gallery, waiting for the teacher to greet the class, Sarah recognized her old nemesis: Ms. Peterson, the woman who had once told her she had “no artistic talent whatsoever.” Now, Ms. Peterson was beaming at her son, Max, and holding a paintbrush.

Sarah’s breath hitched. She blinked, wanting to confirm that her eyes were not deceiving her. This couldn't be happening. She squeezed her husband’s hand, searching for a grounding anchor. He looked at her with a questioning expression.

She felt a hot flush rush up her neck, staining her cheeks. Suddenly, she was a nervous fifth grader again, desperately trying to draw a passable portrait of a still life. The memories were vivid, unwelcome. “Well, this is…interesting,” she managed, her voice barely a whisper.

Emotion: surprised

Cluster: Surprise / Confusion
PC1 (Valence): -0.58 Negative
PC2 (Disposition): -0.18

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
que0.495
de0.447
la0.431
l0.323
😳0.296
Suppressed:
每天-0.269
cheerfully-0.249
ing-0.238
alleviate-0.224
gladly-0.219