AI
New tool helps cross-functional teams spot AI harms early in design
Researchers developed AI LEGO, a visual collaboration platform that makes it easier for technical and non-technical team members to jointly identify potential risks in AI systems before they're built.
A persistent bottleneck in responsible AI development is getting the entire team on the same page about potential harms before code is written. A new research study reveals why this matters—and introduces a practical tool to address it.
The coordination gap in AI teams#
When large organizations build AI systems, they bring together people with vastly different expertise: engineers who make technical choices, designers, ethicists, policy specialists, and business leaders. The friction emerges in handoff: technical teams document their design decisions in ways optimized for other engineers, leaving non-technical colleagues struggling to understand the implications. Existing tools like JIRA or spreadsheets can track tasks, but they don't provide the shared visual language needed for collaborative harm-spotting.
Researchers conducted interviews with eight industry practitioners to understand how this plays out in practice. What they found was consistent: technical design choices rarely get communicated in a way that invites meaningful scrutiny from non-technical roles; teams lack structured workflows for collective deliberation; and non-technical participants often lack scaffolding to conduct systematic risk evaluation.
Introducing AI LEGO#
To bridge this gap, a team developed AI LEGO, a web-based platform grounded in boundary object theory—the idea that shared artifacts can help people with different expertise find common ground. The tool uses a modular, building-block metaphor that lets technical roles draft development plans using interactive components, while non-technical participants engage with the same blocks through role-specific checklists and AI-powered persona simulations.
The persona feature is particularly novel: the system prompts non-technical reviewers to imagine how different stakeholder groups might be affected by design choices and surfaces potential harms that might otherwise go unnoticed.
Testing with real teams#
The researchers tested AI LEGO with 18 cross-functional practitioners—a mix of engineers, product managers, designers, and ethics specialists. Compared to traditional worksheet-based harm identification, the platform increased both the volume of harms identified and the likelihood that teams caught design issues. Participants reported that the modular structure and AI-guided personas made harm evaluation feel more accessible and systematic, fostering more collaborative and transparent responsible AI practices during the crucial early design window.
