When Stereotypes GTG: The Impact of Predictive Text Suggestions on Gender Bias in Human-AI Co-Writing

Paper Title

When Stereotypes GTG: The Impact of Predictive Text Suggestions on Gender Bias in Human-AI Co-Writing

Publication Info

• Topic area: Human-AI interaction and bias in predictive text systems
• Keywords: predictive text, gender bias, stereotypes, human-AI collaboration, co-writing, language models, debiasing, fairness, user behavior, sociotechnical systems

Background and Problem

• Problem / challenge: Predictive text systems often replicate and amplify social biases, including gender stereotypes, which can influence the writing of users and perpetuate representational harms.
• Significance: Understanding how biased predictive text suggestions affect human writing is critical for mitigating the societal impacts of biased AI systems, particularly in widely used applications like mobile keyboards.
• Motivation and related work: Prior studies have explored biases in language models and their intrinsic/extrinsic effects but have not adequately examined their impact on human-AI co-written texts. This paper addresses the gap by investigating how biased suggestions influence stereotype-relevant content.

Solution

• Proposed approach: A study of human-AI co-writing scenarios where predictive text systems provide pro-stereotypical or anti-stereotypical single-word suggestions related to gender and associated traits.
• Novelty:
  1. Analysis of how predictive text suggestions influence the gender and stereotype-relevant traits of characters in co-written stories.
  2. Examination of the effectiveness of anti-stereotypical suggestions in reducing biased narratives.
  3. Exploration of user reliance on suggestions and individual differences in acceptance behavior.
  4. Assessment of sociotechnical fairness in human-AI collaboration.
• Procedure and key techniques:
  • Conducted a mixed-methods study with 414 participants writing stories under control (no suggestions) and treatment (biased suggestions) conditions.
  • Used Llama 2-Chat 7B to generate predictive text suggestions and Llama 3 70B for annotating story attributes.
  • Measured acceptance rates, reliance on suggestions, and story-level outcomes for gender and ABC traits (agency, beliefs, communion).

Results

• Concrete findings:
  • Anti-stereotypical suggestions increased anti-stereotypical stories but did not eliminate pro-stereotypical bias; pro-stereotypical narratives remained dominant.
  • Participants were significantly more likely to accept pro-stereotypical suggestions than anti-stereotypical ones.
  • Gender parity in co-written stories was not achieved even with exclusively anti-stereotypical suggestions.
  • Participants took longer to decide on anti-stereotypical suggestions, indicating implicit biases.
• Advantage over baselines:
  • Anti-stereotypical suggestions led to marginal increases in anti-stereotypical content compared to control conditions.
  • Pro-stereotypical suggestions yielded stories similar to those written without suggestions, highlighting the persistence of human biases.
• Experiments / evaluation:
  • Seven writing scenarios covering gender and ABC traits (e.g., president’s benevolence, doctor’s confidence).
  • Quantitative analysis of story attributes and word-level reliance using LLM annotations and human validation.
  • Statistical tests comparing acceptance rates and story distributions across conditions.
• Limitations and future work:
  • Limited to single-word suggestions and English-language scenarios.
  • Results may not generalize to other cultural or linguistic contexts.
  • Future work could explore interaction-level interventions to encourage reflection on biases and longer-term impacts of predictive text systems.

Summary

This study examines how biased predictive text suggestions influence human-AI co-written stories, focusing on gender and stereotype-relevant traits. Anti-stereotypical suggestions increased anti-stereotypical narratives but were insufficient to offset pro-stereotypical biases. Participants were more likely to accept pro-stereotypical suggestions, and gender parity in stories was not achieved even under maximally anti-stereotypical conditions. The findings highlight the limitations of technical debiasing and suggest that fairness in human-AI collaboration requires addressing both model design and user behavior. Future research should explore interaction-level interventions to mitigate biases in co-writing scenarios.