A Proposed Framework for Human-Centered AI Design Principles

Over the past year, Duke has made significant investments in AI infrastructure, launching new internal tools and adopting vendor platforms that could significantly change the nature of our work, education, research and service missions at Duke. As we now move into the next phase, we’re confronted with the task of actually putting that capacity to work across the countless ideas people have for integrating AI and automation into their daily lives. The shift is exciting and the ideas are endless, but it also creates a new challenge - how do we prioritize, design, develop, assess, and deploy all of these ideas responsibly and effectively? The principles below are a first attempt at naming a shared starting point for doing this work in a way that stays grounded, human-centered, useful, and realistic.

Human Centered and Participatory Design

Projects are created with the people who experience the problem. End users help define the problem, test prototypes, provide feedback, and validate accuracy. We cannot build AI tools for people without building them with people that understand the value and can evaluate their effectiveness.  Stakeholders must be committed to helping develop content, gathering data, and providing iterative feedback.

Human in the loop

AI supports human decision making rather than replacing it. AI can reduce cognitive load, organize information, reformat or re-frame materials, and accelerate work, but humans must remain responsible for final decisions.

Clear success criteria and shared definition of success

Every project begins with an agreement about what success looks like. AI solutions rarely need to be 100% perfect. A tool that performs well enough to deliver meaningful value can be considered complete once the team and stakeholders have aligned on an acceptable performance threshold.

Designing for the present

AI is evolving quickly. These solutions are intended to address current needs with current tools. We are intentionally NOT building with longevity in mind. It is acceptable (and likely, desirable) for solutions to be replaced as technology changes. However, baseline security, privacy, and operational guardrails remain non-negotiable.

Empowering stakeholder ownership

When possible, solutions should be designed so stakeholders can take ownership and make continual adjustments or improvements without requiring heavy engineering support. The people closest to the work should be able to refine the tool as their needs change.

Minimal intrusion and maximum value

AI should fit smoothly into existing workflows and reduce friction. A good solution saves time and effort without asking users to change established processes more than necessary.

Iterative development and velocity over perfection

Where possible, we build and release in small increments so we can learn early, adjust quickly, and prioritize steady forward progress over perfect solutions. This approach helps us provide value throughout the project rather than only at the end.  It also gives stakeholders room to refine their goals or success measures as we clarify the problem, understand technical constraints, and see what AI can or cannot accomplish.

Accessibility and inclusivity

AI tools should be usable by the broadest range of people, including those with disabilities. Inclusive design is required, not optional.

Transparent and predictable communication

We maintain open communication, provide shared visibility into progress, and establish reliable communication rhythms so stakeholders always know where the work stands. At the beginning of each project, we also agree on how communication will happen, including meeting cadences and participation expectations, to ensure clarity and reduce uncertainty.

Flexible adaptation to change

AI projects evolve as we learn more about the problem and the data. Changes in understanding are expected. We design our workflow to allow for constructive adjustments.

These principles represent a starting point, not a finished doctrine. We're figuring it out as we go along, too. I'd love to discuss examples or thoughts about how these concepts flourish or breakdown under real world pressure. What do you think?