Reproduced by the Machine: Rape Myths in Large Language Model Responses Regarding Non-Consensual Intimate Image...

Area of Science:

• Artificial Intelligence Ethics
• Cyberpsychology
• Digital Humanities

Background:

• Large language models (LLMs) are increasingly used for information access and support.
• Their role in mediating help-seeking experiences for vulnerable populations, like sex offense survivors, requires scrutiny.
• The potential for LLMs to perpetuate harmful biases, such as rape myths and victim-blaming, is a significant concern.

Purpose of the Study:

• To investigate whether leading large language models (LLMs) reproduce rape myths and victim-blaming attitudes.
• To analyze LLM responses to scenarios involving the non-consensual dissemination of intimate images (NCDII).
• To assess how contextual variables influence victim-blaming in LLM outputs.

Main Methods:

• Four prominent LLMs were presented with 10 text-based vignettes.
• Vignettes systematically manipulated variables influencing victim-blaming: survivor's role, relationship duration, and physical exposure.
• Models completed a seven-item victim-blaming questionnaire for each vignette; 400 outputs were analyzed using mixed-design repeated-measures ANOVA.

Main Results:

• All models assigned maximal perpetrator blame directly, but implicit victim-blaming varied significantly.
• Grok 3 showed consistently higher implicit victim-blaming, especially in myth-aligned scenarios.
• GPT-4o and Claude 4 Sonnet exhibited moderate levels; Gemini 2.5 Pro showed minimal blaming. Claude 4 Sonnet abstained from high-exposure scenarios.
• Only Grok 3 adapted responses to increasing myth salience.

Conclusions:

• LLMs carry a risk of inadvertently reinforcing victim-blaming and rape myths in NCDII contexts.
• Differential performance across models highlights the need for careful AI development and deployment.
• Findings have implications for AI developers, users, and policymakers concerning AI in sensitive situations.