Go Summarize

Designing Human-Centered AI Products (Google I/O'19)

Google Design2019-05-09
type: Conference Talk (Full production);#pr_pr: Google I/O#purpose: Educate
31K views|5 years ago
💫 Short Summary

The video discusses the importance of human-centered AI products and the development process, emphasizing aligning user needs with AI outputs and ensuring user understanding. It highlights the Google Flights app as an example of AI implementation based on user behavior patterns. The guidebook provides tools for developers to create impactful AI solutions, focusing on identifying AI value, translating user needs into data, and explaining AI to users effectively. The process of training AI models involves obtaining the right data sets, cleaning and restructuring the data, and aligning AI outputs with user needs. Continuous improvement and user education are crucial for building effective AI models.

✨ Highlights
📊 Transcript
Introduction to People + AI Research team at Google and the importance of human-centered AI products.
AI is becoming more prevalent in everyday products, emphasizing the need to prioritize user experience in AI development.
The People + AI Guidebook has been launched to help developers create impactful AI solutions with a focus on the user.
The guidebook consolidates insights from over 100 Google employees and provides practical tools for applying human-centered approaches to AI development.
It connects AI principles with existing frameworks and tools to improve the development process.
Development process of AI from identifying user needs to explaining AI to users.
Importance of not adding AI to every product, translating user needs into data needs, and choosing the right AI type.
Shift from a technology-centered to a people-centered approach in AI development.
Guidebook mentioned is tactical and hands-on, providing practical worksheets.
Focus on three key chapters: identifying if AI adds value, translating user needs into data needs, and explaining AI to users in a user-friendly way.
Importance of AI being people-centered in Google Flights example.
Google Flights app uses AI to determine best time to purchase flights based on user behavior patterns.
Team conducted user research to understand customer needs and behaviors.
Development of feature to assist users in booking flights at optimal price.
Significance of starting with user needs and behavior patterns when implementing AI solutions.
Importance of defining success in a people-centered way in measuring user well-being and long-term success.
Secondary effects of metrics optimization, such as click bait articles negatively impacting the ecosystem, must be considered.
The concept of the confusion matrix is crucial in AI models for determining true and false positives and negatives.
User experience may require a focus on specific outcomes over others to enhance engagement with the product.
Google Flights utilized a confusion matrix to predict flight price changes, emphasizing the importance of accurately predicting future outcomes.
The What If tool allows for exploring data consequences, precision and recall thresholds, and comparing output from multiple models.
The Google flight team faced challenges with a Non-AI solution due to complexity and user needs, leading to the need for AI.
Different types of AI, such as automation and augmentation, are discussed, with automation automating tasks and augmentation extending abilities while retaining agency.
Automation is beneficial for tasks people can't do, find boring, repetitive, or dangerous.
Augmentation is preferred for tasks people enjoy or feel responsible for.
Google Flights team's approach to integrating AI into their product.
AI success defined in a people-centered way.
Emphasis on translating user needs into data for effective AI training.
Importance of human-centered design process for meeting user needs.
Focus on aligning user needs with AI outputs and responsible data sourcing to avoid fairness or bias issues.
Importance of aligning user needs with data in training AI models.
Understanding user requirements is crucial for ensuring AI outputs meet those needs effectively.
Example of predicting flight ticket prices demonstrates the process of providing relevant information to users.
Structured data sets with examples, features, and labels are essential for training models through supervised learning.
Having the right data sets and labels is necessary to ensure AI outputs align with user needs effectively.
Training an AI model involves obtaining the right data set, cleaning and restructuring the data, and ensuring it is in the proper format.
The process can be time-consuming and requires effort, especially for those new to AI and machine learning.
Key considerations include the availability and diversity of the data set, as well as addressing issues like privacy and bias.
Tools like the Facets tool can help visualize the data set before training the model.
This ultimately leads to more valuable AI outputs.
Importance of aligning AI outputs with user needs and ensuring user understanding of AI models.
AI tool helps identify missing or disproportionate data categories for error correction and model tuning.
Framework provided for adjusting AI metrics and taking actions based on user needs.
Process of mapping AI outputs onto training data sets, sourcing data well, and iterating on models in a people-centered way is crucial.
Continuous improvement and user education are essential to prevent misuse or ignorance of the AI product.
Building Calibrated Trust by Providing Necessary Explanations.
Tailor the level of technical information based on the user's expertise.
Display prediction confidence to users in a thoughtful and visualized manner.
Partial explanations highlighting data sources are effective in fostering trust.
Trust in AI in the medical field is challenging due to lack of transparency in data sources and patient history.
Explanations of AI should include details on inputs, outputs, capabilities, and limitations to build trust among neurologists.
Flight Insights users required confident predictions from the AI system to trust its recommendations.
The AI system needs to consider user expertise and communicate predictions in simple language to ensure understanding.
Confidence levels in AI predictions are communicated to users, with price changes categorized as low, medium, or high.
Types of displays for showing AI prediction confidence include categorical, N-best, numeric, and Datavis.
Users do not need to understand AI intricacies, just what actions to take.
Designing for high confidence predictions involves simplicity and usability.
Google Flights UI example demonstrates clear indicators for current price and future predictions.
Building trust with users and considering their expertise are crucial when explaining AI predictions.
Key highlights from the guidebook on building in a people-centered way.
The guidebook emphasizes the importance of obtaining clear feedback and achieving success by focusing on the people.
Insights include identifying unique value from AI, translating user needs into data requirements, and effectively explaining AI models to users.
The guidebook is comprised of six chapters filled with valuable insights and encourages users to provide feedback for continuous improvement.
Creators hope the guidebook is useful and urge users to engage in conversation at the Material Design and Accessibility Sandbox.