Principles of Human-Centered Design for Explainable AI
Human-centered design
The story: A car dashboard is designed for the driver, not the engineer. It shows the few things you need, clearly, at a glance, and warns you in a way you'll understand.
In AI/AWS terms: Human-centered design (HCD) makes AI explanations and interfaces clear, useful, accurate, and fair for the people who use them. It has three principles, below.
For the exam: HCD for explainable AI has three principles: amplified decision-making, unbiased decision-making, and human and AI learning.
1. Design for amplified decision-making
The story: A pilot landing in a storm has seconds to decide. The cockpit display shows only what matters, in plain words, with controls that are hard to press by mistake. A checklist makes them pause and confirm, and every action is logged.
In AI/AWS terms: Supports people making high-stakes decisions under stress or pressure, maximizing the benefit of the technology while minimizing errors. Key aspects: clarity, simplicity, usability, reflexivity (prompting users to reflect on their decisions, like the checklist), and accountability (the log).
For the exam: High-stakes decisions under pressure → design for amplified decision-making.
2. Design for unbiased decision-making
The story: A company wants fair hiring. It looks for where bias creeps in, like interviewers favoring people from their own school. It uses a written scoring sheet everyone follows. And it trains interviewers to notice their own blind spots.
In AI/AWS terms: Keeps decision processes and tools free of bias:
- Identify and assess potential biases.
- Design transparent, fair processes and tools.
- Train decision-makers to recognize and reduce bias.
Key aspects: transparency, fairness, training.
For the exam: Unbiased decision-making = find bias, design fair processes, train the decision-makers.
3. Design for human and AI learning
The story: An apprentice carpenter learns by watching a master. A good teacher adjusts to each student's pace. And the workshop has ramps and signs in several languages so everyone can take part.
In AI/AWS terms: Creates learning environments that work for both people and AI:
- Cognitive apprenticeship: AI learns from human experts, the way apprentices learn from mentors.
- Personalization: learning adapted to each learner.
- User-centered design: accessible to everyone, including people with disabilities or language barriers.
For the exam: AI learning from human experts like an apprentice = cognitive apprenticeship.
RLHF
The story: A new waiter tries two ways of greeting guests. After each shift, the manager says "the second one was better". Over time, the waiter learns what guests like, not from a rulebook but from those rankings.
In AI/AWS terms: Reinforcement learning from human feedback (RLHF) puts human feedback into the reward function so the model aligns with human goals. It improves performance and user satisfaction, and it's used in both traditional and generative AI.
Amazon SageMaker Ground Truth provides the managers: human-in-the-loop labeling, including RLHF. Annotators rank or classify model responses, and that ranking data becomes the reward model.
For the exam: RLHF = human rankings become the reward. On AWS, SageMaker Ground Truth provides the human feedback.