Human Factors in AI 

Designing a product

• Need a human centered approach
• Think user and problem
• 3 Es - Empathy, Expanse  Thinking, Experimentation
• Stanford' Design Thinking Process:
• Empathy->Define->Ideate->Prototype->Test->Repeat
• Empathy
• Gather insight into users' needs
• Observe (listen), Engage (talk)
• Define
• Define product statement - point-of-view
• 3 elements: user, need, insight
• Measurable, narrow enough to recommend a solution
• Ideate
• Generate ideas to solve the problem
• Prototype
• Build to answer questions and test hypothesis
• Quick and cheap: Application, Data , Model
• Get in front of users asap
• Test
• Get user feedback, adjust prototype
• Task Analysis
• Why ? Ensure underspending of the problem
• How? Observe
• Optimize - work backwards from the user task

 

User Experience Design

• Don Norman’s 6 Principles of Interaction Design:
• Visibility: more important->more visible
• Feedback: infirm on the action taken
• Constraints: simplify, limit options
• Mapping: clearly show controls->effects
• Consistency throughout user experience
• Affordance: clarity, user is clear on items purpose

 

User Input

• Data collection should be an integrated workflow
• Ideally - show user immediate benefit
• Cold Start Problem - new user with no use or preference history
• Communicate clearly to user what user is needed and what it will be used for

Transparency

• Explain to the user where and how they are interacting with AI
• How much transparency to provide depends on the use case: text autocorrect vs. loan approval

Communicate Uncertainty

• present the probabilities of each class
• explanation for the predicted class

Feedback Loop

• Feeding user interaction with the model back into the model
• Explicit
• Based on user feedback back on the quality of model output
• Used to retrain the model
• Implicit
• Based on user actions resulted from interaction with the model
• Common in recommendation systems - get items, keep what you like, return the rest
• Need to be careful not to inject user biases

 

DATA PRIVACY

• Right of users to have control over how their information is collected, used and shared
• PII - Personally Identifiable Information
• Nonpublic info, can be tied back to a person
• Subset of this is sensitive data (SSN, etc.)
• Directly Identifiable - name, address
• Identifiable - data points that can be collectively used to re-identify a person
• Must follow the country laws if:
• Offer online services in the country
• Analyze or process data on users from the country
• FIPS - Fair Information Practices
• Guidelines behind data privacy since 1970s
• Rights of Individuals
• Notice
• Choice and consent
• Data access - users nee dot see their info
• Controls on Information
• Info security
• Info quality
• Information lifecycle
• Collection
• Use & Retention
• Management of PII
• Accountability
• Enforcement
• US Approach to privacy
• Modern Hippocratic Laws
• HIPPA - health Insurance Portability and Accountability Act
• FERPA - Family Education Rights and Privacy Act, applicable to public schools
• FCRA - Fair Credit Reporting Act of 1970, see also FACTA
• GLBAU - Gramm-Leach-Bliley Act 1999
• European Union
• GDPR 2018 - applies to companies that have assets, sell to users or store data in EU; fines can be severe
• AI - data needed vs. data available due to privacy
• Need to be careful collecting large sets of data for a model
• Privacy protection - Tech Approaches
• Federated Learning - run the model in silo with sensitive data; shar output only
• Differential privacy - sterilize the input data to remove sensitive user data

 

Bias in AI

• Allocative harm - resources are available to specific group of people only
• Ex: scan of image of hospital staff identifies all women as nurses
• Fairness:
• Individual Fairness
• Group Fairness
• Accountability
• Transparency
• Interpretable models
• Feature importance
• Simplified approximations
• Counterfactual Explanations
• Algorithmic Bias
• Sources of Bias
• Data collection
• Historical bias
• Representational bias
• Defining Features / Labels
• Measurement bias (GPA as measure of learning success)
• Training and evaluating
• Leaning bias
• Deployment
• Deployment bias
• Feedback loop
• Feedback loop bias

 

Mitigating Potential Ethical Risks

Dataset Datasheets

Ethical Checklist

Ethical pre-mortem

 

 

ARTIFICIAL  INTELLIGENCE

• Artificial general intelligence - original concept originated in 1955
• Ability of an intelligent agent to learn a task a human can perform
• "Narrow AI" - ability to accomplish a pre-learned task
• Differentiating b/w Human and AI predictions:
• Human
• Limited memory
• Decisions are limited by factors - bias, mood
• Learn and make decisions based on "rule of thumb"
• Can apply commonsense to a new situation
• AI up until recent
• Capable of processing large amounts of data
• Decisions are consistent - not influenced by eternal factors
• Decisions are based on data only
• Unable to operate outside of trained data
• Automation - replacing humans in performing tasks
• Augmentation - supporting humans