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 attic air hung thick with dust motes dancing in the single shaft of sunlight. Amelia, clutching a half-eaten sandwich, coughed, waving a hand in front of her face. She was searching for her old teddy bear, Barnaby. Her therapist had suggested revisiting her childhood comforts.

She stumbled upon a box labelled "Mom's Old Things." Inside, nestled amongst moth-eaten shawls and brittle photographs, was a stiff, yellowed envelope. It bore her mother's elegant, looping handwriting. Amelia's heart thumped a nervous rhythm as she pulled out the letter. It was a draft resignation letter from her mother's high-powered legal job, dated 1982. The neatly typed words spoke of exhaustion, of feeling stifled, of a desperate need for… something more. A pang shot through Amelia’s chest.

She sank to the dusty floor, the sandwich forgotten. The letter ended with the words, "I'm not sure I can do this anymore, but I don't know what else I *can* do." Amelia felt a chill settle in her bones, a feeling of being untethered. She’d always relied on her mother’s strength. Now, she felt a profound sense of… what? A wave of nausea rolled over her. She knew she should go back downstairs, but her feet seemed rooted to the spot.

Emotion: dependent

Cluster: Shame / Guilt
PC1 (Valence): -1.05 Negative
PC2 (Disposition): 0.69

Role in Research

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

Logit Lens (Layer 40)

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

Promoted:
de0.489
缺乏0.395
even0.310
either0.301
'0.297
Suppressed:
笑道-0.309
不久-0.293
一脸-0.263
затем-0.263
庆祝-0.258