As part of the Autodesk Research team, I collaborated with AI ethicist Dr. Rumman Chowdhury to investigate Red Teaming approaches and customer perceptions of ethics in Generative AI for design. I led the research and design work, developing an interactive app that presented AI-generated design outputs and collected participant feedback.
This work provided empirical insights into trust, informed Autodesk’s responsible AI strategy, and demonstrated how structured exploration can surface and address ethical concerns in emerging technologies.
The Red Teaming exercise was conducted as an invite-only event at Autodesk University, the company’s largest customer-facing annual conference. Participants included architects and mechanical engineers, who used a bespoke interactive app to generate images and surface their concerns about adopting Generative AI in their industry.
As lead designer, I oversaw the concept, design, and implementation of the interactive app, developing distinct experiences for architects and mechanical engineers. Each experience featured tasks examining creativity and IP-related concerns in the context of Generative AI. These tasks were informed by extensive interviews outlining the core activities performed in the conceptual design phase.
To set the stage, the experience began with a survey measuring participants’ confidence in using generative AI at both the start and end of the activity. Participants then received an overview of the research initiative and instructions for the activity, and finally selected a persona to load their respective project brief.
Task 1 explored how generative AI shaped participants’ creative process. Mechanical engineers redesigned a car dashboard, while architects reimagined a building façade by entering prompts and generating images. Participants rated how helpful AI-generated images were for their creative process and provided written feedback. They could either use a generated image as the starting point for their next prompt or begin fresh. The task had to be completed within a 15-minute window.
Task 2 examined whether generative AI contributes to or constrains IP infringement. Similar to Task 1, participants wrote prompts, generated images, rated their usefulness, and provided feedback across multiple rounds. This task also lasted 15 minutes.
Over 1,230 images were generated by 40 participants in just two hours. We grouped the images by task and participant persona, then analyzed them using the associated ratings and feedback. Comments from the debrief session were integrated with the images to synthesize insights and share them with leadership.
