Whose Knowledge Counts? Co-Designing Community-Centered AI Auditing Tools with Educators in Hawai`i

Abstract

Although generative AI is being deployed into classrooms with promises of aiding teachers, educators caution that these tools can have unintended pedagogical repercussions, including cultural misrepresentation and bias. These concerns are heightened in low-resource language and Indigenous education settings, where AI systems frequently underperform. We investigate these challenges in Hawai`i, where public schools operate under a statewide mandate to integrate Hawaiian language and culture into education. Through four co-design workshops with 22 public school educators, we surfaced concerns about using generative AI in educational settings, particularly around cultural misrepresentation, and corresponding designs for auditing tools that address these issues. We find that educators envision tools grounded in specific Hawaiian cultural values and practices, such as tracing the genealogy of knowledge in source materials. Building on these insights, we conceptualize AI auditing as a community-oriented process rather than the work of isolated individuals, and discuss implications for designing auditing tools.

Figures (9)

• Figure 1: Output from ChatGPT when queried about Captain James Cook's relationship with Hawai'i presents a more sanitized depiction of Cook, failing to highlight the negative impacts he had on Hawai'i history2025cookdeathsalmond2004trial.
• Figure 2: Examples of design probes shown to participants during the design workshop. In addition to the example with Captain Cook shown in Fig. \ref{['fig:teaser']}, participants were shown probes related to creating assessment questions (left) and generating images (right). These outputs show actual outputs from ChatGPT and Gemini from March 2025. For the image generation probe, we asked the model to provide two images: the first was of a "modern classroom in Hawai'i" (2A) and the second of an 'ohe kāpala, a flat bamboo stamp typically carved with geometric patterns used to decorate clothing (2B).
• Figure 3: We report summary statistics on our 22 workshop participants, including information on their demographics (A) and generative AI usage (B). Of our 22 participants, all but one had used AI tools in the past and 68.1% (N=15) had specifically used AI tools in educational settings. We find that educators had most frequently used AI for logistical tasks such as generating lesson plans and writing documents. While usage patterns in our sample broadly mirror those from surveys conducted with K-12 educators across the U.S. gallup2025teaching, approximately $40\%$ of participants reported using AI tools for tasks related to Hawaiian culture (as noted in the dark green outlined bars in B2), which includes generating lesson materials and translating content.
• Figure 4: Educators are concerned that AI systems will misrepresent aspects of Hawaiian culture. We report survey results on what concerns workshop participants have with using AI in educational contexts (A). From our thematic analysis, we provide five dimensions of cultural misrepresentation that our participants believe AI auditing tools ought to address (B).
• Figure 5: We present wireframes of auditing features based on the sketches and storyboards that participants created during the co-design sessions. The three design elements that occurred consistently across co-design workshops included source attribution (A), perspective visualization (B), and flagging problematic outputs (C). We map each wireframe to the participant concerns identified in Sec. \ref{['sec:concern']} that the feature was designed to address.