Path7.Mod1.a - Responsible AI Dashboard - General Requirements and Goals Flashcards
Added due to updates to DP-100 on Oct 18th 2023 https://learn.microsoft.com/en-us/training/modules/manage-compare-models-azure-machine-learning/
Define Fairness
Treat everyone fairly, Avoid different treatment for similar groups
Define Reliability and Safety
Manipulation is different from hacking.
Reliable, safe, and consistent. Operate as intended, handle unexpected situations well, and resist harmful manipulation.
Define Privacy and Security
Transparency about data collection, use, and storage, empower users with control over their data. Treat data with care to ensure privacy.
Define Transparency
When influencing life affecting, important decisions, people need to understand how those decisions were made and how the model works.
Define Accountability
Be accountable for decisions that models may influence
MAINTAIN HUMAN CONTROL (Was part of AI-900)
Define Inclusiveness
Systems should empower everyone and engage people
ImpAss OverSigAdvImp FitforPur DatGovMgt HumOvandCon
The five Accountability Goals
- Impact Assessment: identify potential for significant adverse impact on people, organizations and society
- Oversight of Significant Adverse Impacts: Review and apply oversight on Impact Assessment
- Fit for Purpose: Systems provide valid solutions for problems they’re designed to solve
- Data Governance and Management
- Human Oversight and Control: Capabilities for supporting informed human interaction and control (trouble shooting, operating, alerting, etc)
SysIntelDecMak ComtoStak DisAIInt
The three Transparency Goals
- System Intelligibility for Decision Making: Ensure systems are designed to support stakeholder needs for intelligibility
- Communication to Stakeholders: Info on capabilties and limitations so support informed choices about those systems
- Disclosure of AI Interaction: Inform people they are interacting with an AI system
QoS AllReOpp MiniSterDemErasOut
The three Fairness Goals
- Quality of Service: Provide similar quality of service to all groups
- Allocation of Resources and Opportunities: Allocate to minimize disparities in outcomes for all groups
- Minimize Stereotyping, Demeaning and Erasing Output: Systems that describe, depict or represent people, cultures or societies should minimize these things.
RSGuid Fai&Rem OngMonFeedEval
The three Reliability and Safety Goals
- Reliability & Safety Guidance: Operational factors are expected to perform reliably and safely, remediate issues and relate info to users
- Failures and Remediation: Minimize time to remediation for predictable or known failures
- Ongoing Monitoring, Feedback and Evaluation: Systems that describe, depict or represent people, cultures or societies should minimize these things.
PSC SPC
The Two Privacy and Security Goals
- Privacy Standard Compliance: Protect privacy. Simple.
- Security Policy Compliance: Systems must be secure w.r.t. Cybersecurity
The One Inclusiveness Goal
Don’t exclude anyone! LOL