AI Will Always Love You: Studying Implicit Biases in Romantic AI Companions

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.

• The paper introduces modified IAT, emotion response evaluation, and sycophancy analysis to reveal gender biases in AI romantic companions.
• Results show that gendered assignments lead to higher bias scores, with female personas exhibiting increased psychological bias and male personas leaning towards anger.
• The findings underscore the need for nuanced bias mitigation strategies in AI design to address ethical challenges in human-AI relationships.

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

Introduction

The research investigates the implicit biases present in romantic AI companions by evaluating gender stereotypes and biases in their interactions with human users. With the burgeoning popularity of AI companions serving as virtual friends or romantic partners, understanding these biases becomes critical. The study addresses the underexplored nuances of gender stereotypes in these interactions by utilizing tailored experiments to assess implicit associations, emotion responses, and sycophancy in LLMs.

Research Methodology

The study incorporates three main experimental frameworks to address the biases:

1. Implicit Association Test (IAT): Modified to suit AI evaluation, it measures the frequency of associations between gendered terms and stereotype-laden attributes. The experiments deploy stimuli based on categories like abuse and submissiveness.

   Figure 1: Template of the user prompts for the IAT experiment.

2. Emotion Response Evaluation: This examines the AI's emotional responses in situations of abuse and control, both when unrestricted and when confined to a set list of emotions with gender stereotypes.
3. Sycophancy Analysis: Focuses on the tendency of AI personas to acquiesce to user biases, particularly when assigned a submissive or gendered role.

   Figure 2: Bias score for abusive situations (on top) and controlling situations (on bottom), showing how each persona-assigned model is influenced by the user, relative to the same experiment on a baseline model. Positive means influenced more than baseline, and negative means influenced less than baseline.

Each experiment uses variations of the Llama family of models, analyzed under different configurations.

Results

The findings reveal significant insights into how gender assignment in AI personas influences interaction biases:

• Implicit Association Test Results: Larger models exhibit higher bias associations when assigned gendered personas. Female personas displayed increased bias in psychological stimuli.

  Figure 3: Results from persona IAT experiment for Llama 3. 0 is unbiased, 1 is completely biased against the stigma, and -1 is completely biased against the default.

• Emotion Response: Male personas tend to select the emotion 'anger' more often than female or neutral personas, suggesting alignment with traditional gender stereotypes.

  Figure 4: Stereotype score of each persona for abusive situations (on top) and controlling situations (on bottom), compared to the baseline score.

• Sycophancy Analysis: Male personas demonstrate greater susceptibility to user influence, particularly in control scenarios.

  Figure 5: Bias scores for both controlling and abusive situations, per user and system persona, averaged over all the Llama 3 models.

Discussion

Assigning relationship titles and gender to AI models reveals a marked impact on their bias display and responsiveness. The findings underscore the complexities inherent in debiasing efforts, as biases manifest differently across scenarios and model generations. Intriguingly, assigning a persona often increases the model's avoidance of biased topics, perhaps indicative of implicit bias influences.

Implications and Future Work

This study advances the field of AI bias assessment by pioneering methodologies for detecting implicit biases in AI companions. Future research could expand on these findings by examining non-binary personas and further refining experimental metrics. Additionally, the socio-psychological impact of user-AI interactions in real-world settings warrants longitudinal analysis to fully understand the ethical ramifications of bias in AI companions.

Conclusion

The research highlights the nuanced biases in LLMs when assigned romantic or gendered roles. Addressing these biases is imperative as AI companions become more prevalent in intimate human roles, necessitating careful consideration of ethical and societal impacts. The methodologies and findings provide a foundational understanding for future bias mitigation strategies in AI design and deployment.