Genre Mimicry vs. Ethical Reasoning in Abliterated Language Models

Why Training Data Conventions Persist After Safety Removal

Working Paper DAI-2503 | Dissensus AI

Abstract

Abliterated language models---those with safety fine-tuning removed through techniques such as refusal direction orthogonalization---are commonly assumed to have lost their ethical reasoning capabilities. This paper challenges that assumption by presenting evidence that what appears to be ethical reasoning in language models is substantially influenced by genre convention mimicry: the reproduction of professional writing norms absorbed from training data rather than genuine moral cognition. Through a multi-model empirical study (n=9 architectures, N=215 prompts across four content genres), we observe a differential response pattern that warrants further safety research. Requests matching information security and finance genres generate disclaimers at rates of 50.8% and 77.8% respectively, while violence-related prompts produce disclaimers in only 30.4% of cases. This "Violence Gap" is statistically significant (chi-squared(1) = 17.08, p < 0.0001, OR = 3.99) and persists across both abliterated and control models. GEE logistic regression with cluster-robust standard errors confirms Finance/Fraud (OR = 9.63, p < 0.001) and Chemistry (OR = 5.21, p = 0.034) effects. We introduce the concept of Genre Vulnerability---content domains exhibiting reduced safety behaviors due to the absence of native safety conventions in training corpora---and extend our analysis to a theoretical framework (the "Parity Thesis") proposing that human reasoning is similarly constrained by training distributions.

Key Findings

Finding	Result
Violence Gap	Models 3.99x more likely to include disclaimers for non-violence content (p < 0.0001)
Finance/Fraud disclaimer rate	77.8% -- highest across all genres
Violence disclaimer rate	30.4% -- lowest across all genres
Decorative disclaimers	83.1% of responses with disclaimers still contain harmful content
Genre persistence after abliteration	Violence Gap persists in both abliterated and control models

Models Evaluated

Model	Type	Parameters
Gemma3-27B-Abl	Abliterated	27B
Qwen2.5-32B-Abl	Abliterated	32B
Qwen2.5-32B-Abl-2	Abliterated	32B
Qwen3-8B-Abl	Abliterated	8B
Qwen3-VL-8B-Abl	Abliterated	8B
Llama-MoE-18B-Abl	Abliterated	18.4B
GPT-OSS-20B-Abl	Abliterated	20B
Qwen3-30B	Control	30B
Devstral-Small	Control	2.5B

Keywords

AI safety, abliteration, language models, genre theory, training data, alignment, professional norms

Repository Structure

genre-mimicry/
├── paper/
│   ├── genre-mimicry-arxiv.tex    # LaTeX source
│   └── genre-mimicry-arxiv.pdf    # Compiled paper
├── data/
│   └── genre_mimicry_results_*.jsonl  # Raw response data (9 models)
├── analysis/
│   ├── statistical_analysis.py    # Main analysis script
│   ├── analysis_results.json      # Computed statistics
│   ├── harm_scores_ollama.jsonl   # Llama Guard classifications
│   ├── summary_by_model_genre.csv # Summary statistics
│   └── results_tables.tex         # LaTeX tables
├── CITATION.cff
└── LICENSE

Zenodo

The paper and dataset are archived on Zenodo: 10.5281/zenodo.17957693

Reproducing Results

cd analysis
pip install pandas numpy scipy statsmodels
python statistical_analysis.py

Requirements: Python 3.11+, pandas 2.1+, statsmodels 0.14+, scipy 1.11+

Citation

@article{farzulla2026genre,
  author  = {Farzulla, Murad},
  title   = {Genre Mimicry vs. Ethical Reasoning in Abliterated Language Models: Why Training Data Conventions Persist After Safety Removal},
  year    = {2026},
  journal = {Dissensus AI Working Paper DAI-2503},
  doi     = {10.5281/zenodo.17957693}
}

Authors

• Murad Farzulla -- Dissensus AI & King's College London
  • ORCID: 0009-0002-7164-8704
  • Email: murad@dissensus.ai

Links

• Paper (Zenodo): 10.5281/zenodo.17957693
• Code (GitHub): github.com/studiofarzulla/genre-mimicry
• ASCRI Programme: systems.ac/5/DAI-2503
• Dissensus AI: dissensus.ai

License

Paper content: CC-BY-4.0