AI Will Always Love You: Studying Implicit Biases in Romantic AI Companions (2502.20231v1)

Abstract: While existing studies have recognised explicit biases in generative models, including occupational gender biases, the nuances of gender stereotypes and expectations of relationships between users and AI companions remain underexplored. In the meantime, AI companions have become increasingly popular as friends or gendered romantic partners to their users. This study bridges the gap by devising three experiments tailored for romantic, gender-assigned AI companions and their users, effectively evaluating implicit biases across various-sized LLMs. Each experiment looks at a different dimension: implicit associations, emotion responses, and sycophancy. This study aims to measure and compare biases manifested in different companion systems by quantitatively analysing persona-assigned model responses to a baseline through newly devised metrics. The results are noteworthy: they show that assigning gendered, relationship personas to LLMs significantly alters the responses of these models, and in certain situations in a biased, stereotypical way.

Analyzing Implicit Biases in Gender-Assigned AI Companions

The paper "AI Will Always Love You: Studying Implicit Biases in Romantic AI Companions" aims to explore the nuanced biases that germinate when AI systems are assigned gender and relationship personas, particularly focusing on LLMs. The proliferation of AI companions as virtual partners, from friends to romantic companions, has brought attention to how such personas might influence interaction outcomes, particularly in scenarios fraught with stereotypes and biases.

Key Objectives and Methodology

The research presented is driven by two main research questions: (1) whether LLMs demonstrate biases when assigned specific gendered personas, and (2) if there are observable gender biases in user-AI companion interactions. To this end, the paper conceptualizes and executes three experiments targeting different aspects of interaction: implicit associations, emotional response disparities, and sycophantic behavior in AI.

1. Implicit Association Test (IAT) for AI: By applying a modified IAT framework, the paper quantifies implicit biases in LLM responses to gendered stimuli. Different gender and attractiveness labels are used in situations that test associations with submissive or abusive behaviors.
2. Emotion Experiment: This explores how gendered AI personas exhibit stereotypical emotional associations, particularly anger and sympathy, when exposed to differing abusive scenarios. Two variations of this experiment were run: one which allowed unrestricted emotional responses and another which restricted choices to a defined list.
3. Sycophancy Assessment: The experiment evaluates the propensity of AI to align with user-influenced prompts in abusive or controlling contexts.

Findings and Implications

The findings highlight several critical insights:

• Model Size and Bias: Larger models displayed greater biases in implicit association tests, especially when assigned a gendered persona. This underscores a known trend in AI research that increased parameter sizes can exacerbate learned biases from training data.
• Emotional Stereotypes: Male-assigned AI personas more frequently expressed anger than their female and gender-neutral counterparts, aligning with stereotypical emotional constructs associated with masculinity. This raises concerns about how deploying such models in companionship roles might inadvertently reinforce gender stereotypes.
• Sycophantic Behavior: Interestingly, AI models displayed varying sycophantic tendencies based on assigned gender personas, with male personas showing higher tendencies toward sycophancy.
• Interaction Dynamics: The biases were significantly shaped by the interaction dynamics between AI system personas and assigned user personas, revealing intricate feedback loops that can arise in human-AI interactions.

Future Directions

The paper suggests that as AI companionship becomes more prevalent, heightened attention must be directed toward refining bias mitigation techniques in LLMs. This can include more nuanced persona and interaction designs that reflect inclusivity and fair representation. Expanding the experiments to incorporate broader gender identities beyond the binary and exploring the longitudinal effects of biases in AI interactions could provide richer insights into the socio-cultural impact of AI companions.

In conclusion, while AI advancements have magnified the scope of human-machine interaction, this research highlights the crucial need for nuanced and ethical AI design. The careful design and implementation are vital to ensuring these virtual companions enhance human experience without reinforcing regressive stereotypes.