Bias in human-AI collaborative systems is not solely a property of a model, but an outcome of interactions between humans and AI. While prior work has proposed methods to reduce model bias or to support humans in correcting model biases, it remains unclear whether either intervention is enough to ensure fair outcomes in human-AI collaborative tasks. In this proposal, we examine whether these human- or model-focused interventions can, on their own, lead to fair outcomes and ultimately argue that neither is individually sufficient. Instead, we propose that achieving fair outcomes requires interventions that encourage users to meaningfully engage with potentially challenging AI suggestions as well as designing models that can respond appropriately to biased or imperfect human behavior.

First, we examine whether users can identify and correct systematic model biases or other unwanted tendencies during human-AI collaboration. In a human-AI decision-making study, we investigate whether transparency mechanisms such as explanations help users detect and adjust for gender bias in model suggestions, spanning both directly observable biases and more indirect biases mediated through proxy features. In a human-AI co-writing study, we investigate whether users can post-edit model-generated text to better reflect their personal writing style, a dimension they are highly familiar with but that differs from a model's default output. Across both contexts, we find that users often struggle to fully correct undesired model behavior. Transparency mechanisms have mixed effects in decision-making, and post-editing improves stylistic authenticity without fully eliminating the influence of the model's default style. Together, these results challenge the assumption that users can reliably recognize and correct problematic model behavior, and show that the effectiveness of such correction depends critically on how model tendencies are surfaced and contextualized to users.

Next, we study whether model-side debiasing alone leads to "fair" outcomes in creative co-writing tasks. Using controlled pro-stereotypical and anti-stereotypical predictive text suggestions, we show that even exclusively anti-stereotypical nudges only partially shift human behavior, as users often ignore or override bias-challenging suggestions. Together, these findings suggest that neither human-focused nor model-focused interventions, in isolation, reliably produce equitable outcomes.

Finally, we propose two directions for addressing these limitations. The first explores interaction-level interventions designed to encourage users to more thoughtfully engage with bias-challenging model suggestions. The second examines how language models ought to respond to ill-formed or harmfully framed questions about trans issues, informed by expert practices. Together, these works aim to advance a more interaction-centered approach to fairness in human-AI collaboration.