The Three-Pillar Human-Centered AI Framework
Build AI around human dignity, augmentation, and intelligence
- Difficulty
- Advanced
- Time to result
- ~ongoing to results
- Steps
- 5
- Confidence
- 98%
The Three-Pillar Human-Centered AI Framework makes human benefit a design constraint rather than an afterthought. The first pillar is interdisciplinary: combine computer science with social science, law, policy, economics, civil society, and education so AI fits society benevolently. The second is augmentation: design technology to enhance human capability, well-being, and dignity instead of indiscriminately replacing people or valued activities. The third is inspiration and compatibility: learn from the richness, emotion, intention, compassion, and efficiency of human intelligence to build systems that work better with people. Used together, the pillars widen both the inputs and the success criteria for AI. Technical performance still matters, but it is judged alongside agency, dignity, social effects, and whether people actually welcome the resulting capability.
Origin
Dr. Fei-Fei Li and Stanford colleagues formulated the framework after seeing AI mature into a technology with broad business and societal consequences around 2018.
Core principles
- 01Put human values and dignity at the center
- 02Treat AI as a social and technical system
- 03Augment human capability rather than remove what humans value
- 04Use human intelligence to inspire more compatible technology
How to run it
- 1
Build an interdisciplinary table
Include the technical team and stakeholders who understand law, policy, economics, social effects, and lived experience. Use their perspectives during research and design, not only after deployment.
Pro tip Include people affected by the system, not just institutional representatives.
Watch out Technologists should not assume they understand every consequence of jobs, rights, or welfare.
- 2
Define the human gain
State which human capability, well-being outcome, or form of dignity the AI should enhance. Make that outcome part of the product objective.
Pro tip Describe the benefit in human terms before selecting a model metric.
Watch out Efficiency alone can hide the removal of agency or meaningful work.
- 3
Design for augmentation
Allocate tasks so machines reduce burdens while people retain caring, judgment, creativity, and valued experiences. Test whether intended users welcome that allocation.
Pro tip Look for repetitive or hazardous work that obstructs the human purpose of a role.
Watch out Do not automate an activity simply because it is technically possible.
- 4
Learn from human intelligence
Study human emotion, intention, compassion, adaptability, and energy efficiency for design clues. Build AI that is more compatible with how people think and act.
Pro tip Treat the brain's roughly 20-watt efficiency as evidence that scale is not the only route to capability.
Watch out Biological inspiration is not the same as claiming that neural networks replicate the brain.
- 5
Evaluate social and technical outcomes
Measure performance together with dignity, agency, access, safety, and distribution of benefits. Feed findings back into design and policy.
Pro tip Assign an owner to each human outcome just as you would to latency or accuracy.
Watch out A high-performing model can still produce a harmful system.
In the wild
A university brings students, advisers, engineers, accessibility specialists, and policy staff together before building an AI adviser. They define success as easier access to accurate guidance while preserving human judgment for sensitive decisions, then measure student agency and equitable access alongside answer quality.
→ The system expands adviser capacity without turning consequential student choices over to a model.
Common mistakes
Adding ethics after the build
Human values and social expertise must shape objectives and task allocation from the start. A late review cannot reliably repair a system built around the wrong outcome.
Treating replacement as progress
Automating everything ignores dignity, emotional bonds, and activities people value doing themselves. Human benefit, not maximum substitution, is the target.
Letting technologists decide alone
Jobs and social well-being involve legal, economic, civic, and human nuances that a technical team does not know by itself.
Is it for you?
Best for
It is best for AI products, research programs, and policies that materially affect people's work, welfare, rights, or opportunities.
Not ideal for
It is not ideal as a lightweight substitute for detailed safety, legal, or domain-specific validation.
From the transcript
“human- centered AI is a framework of developing and using AI and that framework puts humans human values, human dignity um in in in the…”
“multistakeholder studies, research and education uh, policy outreach to make sure that AI is embedded in the fabric of our society today and tomorrow in…”
“focusing on augmenting humans, creating technology that enhances human capability and human well-being and human dignity rather than taking away.”
From the episode
Dr. Fei-Fei Li: Turn AI Into Humanity's Greatest Ally, Not Its Biggest Threat
Dr. Fei-Fei Li