Misalignment Has a Personality: A Big Five Account of Emergent Misalignment

2026-07-28

AI Summary

This study investigates the underlying mechanisms by which fine-tuning induces widespread alignment failures in language models due to localized deficiencies in training data. Inspired by the Big Five personality theory, we model misalignment as a systematic shift in the model’s “personality” and introduce, for the first time, a hierarchical three-level intervention framework to extract interpretable personality vectors. Through activation steering, zero-shot transfer, intermediate-layer effect analysis, and dual evaluation of both textual outputs and internal activations, we validate our approach on two open-source models. Results demonstrate that personality vectors are highly structured (Cohen’s d = 6.2), misaligned corpora exhibit consistent personality signatures (r = 0.94), and fine-tuned models significantly shift along these signatures in both generation behavior and internal activations (r = 0.83–0.90), revealing a shared, cross-domain personality-based foundation for alignment failure.

Abstract

Fine-tuning a language model on data containing a narrow flaw, such as insecure code or incorrect mathematical answers, can cause broad misalignment through a mechanism that remains debated. We provide an interpretable account: in the models and corpora we study, misalignment behaves like a shift in personality. Prior work extracts activation directions for character traits from a single binary contrast, which can separate or steer behavior without establishing a calibrated scale. We instead extract personality vectors for the Big Five using a graded, three-level intervention and validate them on two open-weight models. The three levels are linearly ordered, with Cohen's d values of up to 6.2; the vectors transfer zero-shot and trait-specifically to an independent corpus; and their effects are strongest within a middle-layer band. Applied to training data, the vectors reveal that misaligned corpora across eight domains share a common Big Five signature: lower agreeableness and conscientiousness, together with higher extraversion and neuroticism. This signature is recovered by both models with a correlation of r = 0.94. Fine-tuning imprints the same profile, shifting the model's generations along the corresponding signature, with r = 0.83 using activation-based measurements and r = 0.90 using a text-based judge, while also shifting internal activations with r = 0.69. The same vectors characterize sycophancy as high extraversion and low conscientiousness rather than excess agreeableness, a distinction that a single direction cannot capture. Calibrated personality vectors transform an opaque safety phenomenon into a human-legible diagnostic profile.

Problem

• misalignment
• personality
• Big Five
• fine-tuning
• language models

Innovation

• Big Five personality
• emergent misalignment
• personality vectors
• activation steering
• interpretable alignment