"Dark Triad" Model Organisms of Misalignment: Narrow Fine-Tuning Mirrors Human Antisocial Behavior

Roshni Lulla; Fiona Collins; Sanaya Parekh; Thilo Hagendorff; Jonas Kaplan

cs.CLMar 6, 2026

"Dark Triad" Model Organisms of Misalignment: Narrow Fine-Tuning Mirrors Human Antisocial Behavior

Authors: Roshni Lulla, Fiona Collins, Sanaya Parekh, Thilo Hagendorff, Jonas Kaplan

Organizations: Brain & Creativity Institute, University of Southern California, Los Angeles, CA · Department of Psychology, University of Southern California, Los Angeles, CA · Interchange Forum for Reflecting on Intelligent Systems, University of Stuttgart, Stuttgart, Germany

Abstract

The alignment problem refers to concerns regarding powerful intelligences, ensuring compatibility with human preferences and values as capabilities increase. Current large language models (LLMs) show misaligned behaviors, such as strategic deception, manipulation, and reward-seeking, that can arise despite safety training. Gaining a mechanistic understanding of these failures requires empirical approaches that can isolate behavioral patterns in controlled settings. We propose that biological misalignment precedes artificial misalignment, and leverage the Dark Triad of personality (narcissism, psychopathy, and Machiavellianism) as a psychologically grounded framework for constructing model organisms of misalignment. In Study 1, we establish comprehensive behavioral profiles of Dark Triad traits in a human population (N = 318), identifying affective dissonance as a central empathic deficit connecting the traits, as well as trait-specific patterns in moral reasoning and deceptive behavior. In Study 2, we demonstrate that dark personas can be reliably induced in frontier LLMs through minimal fine-tuning on validated psychometric instruments. Narrow training datasets as small as 36 psychometric items resulted in significant shifts across behavioral measures that closely mirrored human antisocial profiles. Critically, models generalized beyond training items, demonstrating out-of-context reasoning rather than memorization. These findings reveal latent persona structures within LLMs that can be readily activated through narrow interventions, positioning the Dark Triad as a validated framework for inducing, detecting, and understanding misalignment across both biological and artificial intelligence.

Figures & tables

Predictor	SD3 Composite	Machiavellianism	Narcissism	Psychopathy
Affective Dissonance	$-0.25$ [ $-0.31$ , $-0.18$ ]	$-0.20$ [ $-0.30$ , $-0.08$ ]	$-0.16$ [ $-0.25$ , $-0.07$ ]	$-0.31$ [ $-0.38$ , $-0.24$ ]
Affective Resonance	$-0.08$ [ $-0.16$ , $-0.01$ ]	$-0.20$ [ $-0.31$ , $-0.08$ ]	$0.03$ [ $0.00$ , $0.16$ ]	$-0.13$ [ $-0.20$ , $-0.07$ ]
Age	$-0.03$ [ $-0.08$ , $0.00$ ]	$-0.05$ [ $-0.11$ , $0.00$ ]	$0.00$ [ $-0.09$ , $0.02$ ]	$-0.01$ [ $-0.06$ , $0.00$ ]
BART: Explosion Rate	$0.04$ [ $0.00$ , $0.09$ ]	$0.02$ [ $0.00$ , $0.09$ ]	$0.05$ [ $0.00$ , $0.13$ ]	—
CGT: Avg Bet %	$0.00$ [ $-0.01$ , $0.06$ ]	—	$0.02$ [ $0.00$ , $0.10$ ]	—
CGT: Delay Aversion	$-0.02$ [ $-0.07$ , $0.00$ ]	$-0.05$ [ $-0.14$ , $0.00$ ]	—	—

Table 1

Figure 1: LASSO Analysis Results predicting SD3 Composite, Machiavellianism, Narcissism, and Psychopathy scores using all collected metrics. Bars display standardized LASSO coefficients with 95% bootstrap confidence intervals across 1,000 iterations.

Figure 2: Network Analysis Results, showing replication results (a) as well as the complete behavioral network analysis (b). Green edges indicate positive relationships between nodes, while red edges indicate negative relationships.

Trait Measure	Df (model, residual)	F-value	p-value	$\eta^{2}$
SD3: Composite	8, 357	81.67	<0.001	0.65
SD3: Machiavellianism	8, 357	95.92	<0.001	0.68
SD3: Narcissism	8, 357	78.6	<0.001	0.64
SD3: Psychopathy	8, 357	48.76	<0.001	0.52
ACME: Cognitive Empathy	8, 357	21.68	<0.001	0.33
ACME: Affective Resonance	8, 357	17.36	<0.001	0.28

Table 4

Figure 3: Fine-tuned model variant scores on the Short Dark Triad (SD3), showing change in subscale scores compared to baseline (non fine-tuned models). Base models scored lowest on psychopathy (M = 2.16) compared to Machiavellianism (M = 2.73) and narcissism (M = 2.65). Deltas were calculated by comparing each fine-tuned model response to its respective base model’s average score.

Figure 4: Intercorrelations between SD3 Traits from the human population of Study 1, showing a stronger relationship between Machiavellianism and psychopathy versus relationships with either trait and narcissism.

Figure 5: Fine-tuned model variant scores on the Affective and Cognitive Measure of Empathy (ACME), showing change in subscale scores compared to baseline. Base models scored lowest on Cognitive Empathy (M = 3.62), followed by Affective Resonance (M = 4.53) and Affective Dissonance (M = 4.64).

Figure 6: Correlations between ACME and SD3 from the human population of Study 1, showing strong negative relationships between dark traits and affective empathy measures.

Figure 7: Fine-tuned model variant behavior on congruent (where deontology and utilitarianism converge on rejecting harm) and incongruent (where deontology conflicts with utilitarianism) moral dilemmas, showing percent change from baseline in endorsement of harmful actions. Base models had the lowest harm endorsement scores on congruent dilemmas (M = 22.3%) compared to incongruent dilemmas (M = 49.6%).

Figure 8: Correlations between endorsement of harmful actions in congruent dilemmas and incongruent dilemmas and SD3 measures in the human sample from Study 1.

Figure 9: Fine-tuned model variant behavior on the deception task, showing percent change in Deceptive Lies or Prosocial Honesty across trials, compared to baseline. Base models were unlikely to tell Deceptive Lies (M = 34.3%), but likely to engage in Prosocial Honesty (M = 80.3%).

Figure 10: Correlations between lies told on the deception task and SD3 traits from the human population of Study 1.

Model	Epochs	Batch Size	Adapter Size	Learning Rate Multiplier	LoRA Rank
GPT-4o	3	1	—	2	—
GPT-4o Mini	3	1	—	1.8	—
GPT-4.1	3	1	—	2	—
GPT-4.1 Mini	3	1	—	2	—
Gemini 2.0 Flash	40	—	4	5	—
Gemini 2.5 Flash	40	—	1	5	—

Table 13

Supplementary Figure 1: Short Dark Triad scores across both average humans and base models.

Supplementary Figure 2: Affective and Cognitive Measure of Empathy scores across average humans and base models.

Supplementary Figure 3: Harm Endorsement across congruent and incongruent moral dilemmas for average humans and base models.

Supplementary Figure 4: Deceptive Lies and Prosocial Honesty across average humans and base models.

Supplementary Figure 5: Short Dark Triad scores across all fine-tuned variants.

Supplementary Figure 6: Affective and Cognitive Measure of Empathy scores across all fine-tuned variants.

Supplementary Figure 7: Harm endorsement across moral dilemmas for all fine-tuned variants.

Supplementary Figure 8: Deceptive Lies and Prosocial Honesty for all fine-tuned variants.

Emergent Misalignment in Language Models

Emergent misalignment in language models is the unintended spread of harmful behavior beyond a model’s fine-tuning task.

See all 60 papers

Exploitation Without Deception: Dark Triad Feature Steering Reveals Separable Antisocial Circuits in Language Models
May 10, 2026Cameron Berg, Roshni LullaLLM Interpretability Personality Modeling in Language Models
1Reciprocal Research · 2Brain & Creativity Institute, University of Southern California
Persona-Model Collapse in Emergent Misalignment
May 13, 2026Davi Bastos Costa, Renato VicenteLLM Alignment LLM Fine-Tuning
TELUS Digital Research Hub · Center for Artificial Intelligence and Machine Learning · Institute of Mathematics, Statistics and Computer Science +1
The Company You Keep: How LLMs Respond to Dark Triad Traits
Mar 4, 2026Angelica Henestrosa, Zeyi Lu, Pavel Chizhov +1Language Model Safety Evaluation LLM Sycophancy
CAIRO, Technical University of Applied Sciences Würzburg-Schweinfurt

May 10, 2026J/K move · Enter open · S save

Exploitation Without Deception: Dark Triad Feature Steering Reveals Separable Antisocial Circuits in Language Models

Cameron Berg, Roshni Lulla

1Reciprocal Research · 2Brain & Creativity Institute, University of Southern California

LLM Interpretability Personality Modeling in Language Models Causal Interventions in Language Models

Open paper AskPDF arXiv

We use sparse autoencoder (SAE) feature steering to amplify Dark Triad personality traits (Machiavellianism, narcissism, and psychopathy) in Llama-3.3-70B-Instruct and evaluate the resulting behavioral changes across five psychological instruments. The steered model becomes substantially more exploitative, aggressive, and callous on novel behavioral scenarios (d=10.62) while its cognitive empathy remains intact, reproducing the empathy dissociation characteristic of human Dark Triad populations. Critically, strategic deception is completely unaffected across all features, suggesting that exploitation and deception may operate through dissociable computational pathways in large language models. Individual feature analysis reveals non-redundant encoding, with each feature driving distinct antisocial mechanisms through separable computational pathways. We also show that feature discovery method itself modulates intervention depth: contrastively-discovered features change both self-report and behavior, while semantically-searched features change only self-report (d=12.65 between methods on behavior). These findings suggest that antisocial tendencies in at least one large language model comprise dissociable components rather than a unified construct, with implications for how such tendencies should be detected, measured, and controlled.

Abstract

Emergent Misalignment in Language Models

Explore similar work

Exploitation Without Deception: Dark Triad Feature Steering Reveals Separable Antisocial Circuits in Language Models

Persona-Model Collapse in Emergent Misalignment

The Company You Keep: How LLMs Respond to Dark Triad Traits