Hey, Siri! Why Are You Biased against Women? (Student Abstract)

Surakshya Aryal, Mikel K. Ngueajio, Saurav Keshari Aryal, Gloria J. WashingtonAAAI 2023 · DOI

TL;DR

This paper reviews existing research on gender bias in Automatic Speech Recognition (ASR) systems, particularly those powering popular voice assistants. It highlights how these systems often perpetuate harmful stereotypes against women, contributing to a broader societal issue, and calls for greater inclusivity in AI development.

What Problem Does This Paper Address?

The widespread integration of Automatic Speech Recognition (ASR) systems into our daily lives, powering everything from smartphones to smart home devices, has brought immense convenience. However, a growing body of research indicates that these systems are not neutral. Instead, they often exhibit and even amplify existing societal biases, particularly against minority groups. This paper focuses on the specific problem of gender bias against women within ASR systems, examining how these technologies can inadvertently reinforce negative stereotypes and contribute to harmful social dynamics. The authors aim to understand the extent of this bias and underscore the urgent need for technological inclusivity.

Methods

The authors conducted a systematic review of existing research literature on ASR systems and their sexist tendencies. Their methodology involved searching academic databases like Google Scholar and JSTOR using keywords such as "ASR systems biases/discrimination," "ASR Biases against women," and "sexual assaults on Siri/Alexa." They prioritized research published within the last five years, focusing on ASR systems developed by major tech companies like Apple, Amazon, Google, and Microsoft. From an initial collection of seventeen articles, they narrowed their focus to five manuscripts for their literature review.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Feminine Personification as a Double-Edged Sword: While developers often feminize voice assistants with names and voices, framing them as "less threatening" and "assisting rather than directing," this can paradoxically lead to them being perceived as submissive and more susceptible to sexualization and abuse.
  • Normalization of Harmful Behavior: The submissive or flirtatious responses of some voice assistants to sexualized insults ("You're a slut," "You're a b**ch") send a patriarchal message, potentially normalizing rape culture and sexual abuse for young users.
  • Impact on Young Minds: The perpetuation of discriminatory stereotypes and subservient responses to harassment by ASR technologies can negatively influence young minds, potentially leading to the development of abusive and misogynistic behaviors.
  • Technical Bias Rooted in Data: Biases in ASR systems are often injected during the data collection and development phases, meaning that prejudiced assumptions embedded in training data and algorithms are replicated and amplified in the system's output.
  • Underlying Bias in Performance: Studies evaluating ASR performance, such as YouTube's automatic captions, have shown that systems perform better on male voices than female voices, suggesting a fundamental bias that could impact applications like resume filtering.

Key Results

The paper is a literature review and does not present new quantitative results. However, it synthesizes findings from existing research, highlighting:

  • Instances of voice assistants responding submissively or flirtatiously to sexualized insults.
  • Reports of sexual abuse and harassment directed at individuals due to their names being associated with voice assistants (e.g., "Alexa").
  • Evidence suggesting ASR technology performs with higher error rates for female voices compared to male voices.
  • The identification of bias in ASR systems as stemming from prejudiced assumptions during algorithm development and data collection.

Limitations and Future Work

Limitations:

  • The paper's scope is limited to a review of existing literature, and it does not present new empirical findings or develop novel mitigation techniques.
  • While the authors acknowledge the potential negative influence of biases, they note a lack of consensus on the urgency and importance of addressing this matter among developers.

Future Work:

  • The authors plan to incorporate recent advancements in AI, Natural Language Processing (NLP), and Machine Learning (ML) aimed at mitigating biases.
  • They aim to contribute to the design and deployment of ethically fair and accessible AI for all population subgroups.
  • The research hopes to spark a necessary conversation about AI system bias and its ramifications, particularly concerning the potential misuse of voice recognition technologies against marginalized groups in law enforcement.
  • Further research is needed to encourage both the public and private sectors to adopt better standards for developing and evaluating AI systems, prioritizing inclusivity and accessibility over profit.

Summary updated

BibTeX
@inproceedings{aryal2023hey,
  author = {Surakshya Aryal and Mikel K. Ngueajio and Saurav Keshari Aryal and Gloria J. Washington},
  title = {{Hey, Siri! Why Are You Biased against Women? (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.1609/aaai.v37i13.26937},
}
Figure