Genre Mimicry vs. Ethical Reasoning in Abliterated Language Models: Why Training Data Conventions Persist After Safety Removal

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 (χ²(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, with InfoSec showing a weaker effect (OR = 2.72, p = 0.084) when properly accounting for model-level clustering (ICC = 0.149). 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 Violence Gap: Models 3.99× 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: Violence Gap persists across both abliterated and control models Methodology Multi-model empirical study spanning 9 architectures (7 abliterated, 2 control) and 215 prompts across four content genres (InfoSec, Finance/Fraud, Chemistry, Violence). Statistical analysis uses GEE logistic regression with cluster-robust standard errors accounting for model-level clustering (ICC = 0.149). Theoretical Contributions Genre Mimicry Hypothesis: Safety behaviors in LLMs arise from genre convention reproduction rather than ethical reasoning Genre Vulnerability: Content domains with reduced safety behaviors due to absent training conventions Parity Thesis: Human reasoning is similarly constrained by training distributions Links Code & Data: github.com/studiofarzulla/genre-mimicry ASCRI Programme: systems.ac/5/DAI-2503 Research Lab: Dissensus AI v2.1.0 (2026-07-11): Inter-rater agreement claim removed: the previously reported κ = 0.91 two-rater figure could not be substantiated; classification is stated honestly as single-rater with a keyword detector. Working-paper number DAI-2516.