Human-Centered AI Stakeholder Design
Design AI with every affected constituency from the start.
- Difficulty
- Expert
- Time to result
- ~months to results
- Steps
- 6
- Confidence
- 98%
Human-Centered AI Stakeholder Design expands product discovery beyond the direct user. Begin by identifying every constituency affected by the system's decisions, including subjects of decisions, families, past or future victims, institutions, and society. Map the benefits, burdens, and definitions of fairness for each group. Bring multidisciplinary and multicultural participants into objective-setting before engineers lock in the architecture, because a homogeneous team can mistake internal agreement for correctness. Reconcile competing goals explicitly, design autonomy and control around affected people, and test outcomes by constituency rather than only measuring usability for the operator. The mechanism prevents fairness from becoming a cosmetic post-build patch and produces a system whose objective, interface, and operating policy reflect the full field of human consequences.
Origin
Norvig distinguishes software engineering, which fights complexity, from AI, which fights uncertainty, and human-centered AI, which must build systems that do the right thing fairly for everyone involved.
Core principles
- 01Human-centered AI should do the right thing for everyone affected.
- 02The direct user is only one constituency.
- 03Fairness cannot be bolted on after engineering is complete.
- 04Diverse disciplines and cultures reveal missing objectives.
- 05Affected groups should shape the system from the beginning.
How to run it
- 1
Identify all constituencies
List direct users, decision subjects, indirect stakeholders, and wider groups affected by the system.
Pro tip Follow each output outward to the people who experience its consequences.
Watch out Do not stop at the person who operates the interface.
- 2
Map effects and values
Document benefits, harms, control, and competing definitions of fairness for each constituency.
Watch out A benefit to one group may be a burden to another.
- 3
Broaden the design group
Include relevant disciplines, cultures, and affected perspectives before defining the objective.
Pro tip Invite people able to challenge what the engineering team treats as obvious.
- 4
Reconcile objectives
Surface disagreements about what to optimize and document how the chosen objective balances them.
Watch out Consensus inside a homogeneous team is weak evidence of fairness.
- 5
Co-design the system
Let the stakeholder analysis shape data, objectives, controls, interfaces, and operating policy from the start.
Pro tip Tie each major design choice to a mapped constituency effect.
Watch out Do not postpone fairness work until after the core system is built.
- 6
Test outcomes by group
Evaluate usability and real consequences across constituencies, then monitor them after launch.
Watch out A polished experience for the operator can coexist with severe harm elsewhere.
In the wild
A conventional user-centered approach might optimize charts and case information for the judge. Norvig's human-centered approach also examines effects on the defendant and family, past and potential victims and families, and society through incarceration and discrimination.
→ The system is designed around all affected constituencies rather than only its operator.
A hiring team includes recruiters, candidates, accessibility specialists, managers, and legal experts before defining an AI screening objective. They test not only recruiter speed but also candidate exclusions and appeal paths.
→ Efficiency is balanced against fairness, access, and accountability before launch.
Common mistakes
Designing only for the operator
A great interface for the direct user can hide serious consequences for people subjected to the decision.
Bolting fairness on later
Late-stage review cannot easily repair an objective, dataset, or architecture built around incomplete values.
Mistaking agreement for truth
A homogeneous team may agree because everyone shares the same blind spots.
Is it for you?
Best for
AI systems with consequential effects beyond the person operating the interface.
Not ideal for
Isolated, low-stakes personal tools with no meaningful impact on other people.
From the transcript
“And then in human-centered AI, the goal is to build systems that do the right thing for everyone. And do that fairly.”
“So, you're not just serving one user. You're You're serving all these different constituencies.”
“You've got to really bring in all these people right from the start.”
From the episode
Peter Norvig: Simple Ways to Grow Your Business with AI
Peter Norvig