The Six Responsible AI Principles
Level Up Team
This post summarizes The Six Responsible AI Principles, module 4.1 of Level Up’s Azure AI Fundamentals series. The lesson is on YouTube as Azure AI Fundamentals AI-901 Exam: 4.1: The Six Responsible AI Principles. Module 3.4 is Configuration Parameters.
A capable model can still produce an irresponsible system
Responsible AI guides how organizations build and operate AI, not only how they choose a model. It covers how teams design solutions, select and prepare data, evaluate behavior, deploy and monitor applications, and respond when harm occurs.
Six principles, one primary risk
Several principles can apply to one failure. On the exam, select the principle that matches the question’s primary risk.
| Principle | Central question |
|---|---|
| Fairness | Does the system treat comparable people consistently? |
| Reliability and safety | Does it operate safely under expected and unexpected conditions? |
| Privacy and security | Does it protect people, data, models, and credentials? |
| Inclusiveness | Can people with different needs and abilities use it? |
| Transparency | Do people understand the system’s purpose and limitations? |
| Accountability | Who remains responsible for its decisions and effects? |
Fairness is comparable treatment
Ask whether differences have a valid, appropriate basis. Identical outcomes are not the test. A résumé-screening system that advances equally qualified applicants at different rates based on gender is a fairness failure. Measure performance across relevant populations, use sufficiently representative data, and investigate unequal error rates. Disparate treatment, demographic bias, or unequal outcomes usually signal fairness.
Reliability is performance. Safety is the harm when it fails.
A medical assistant that gives confident treatment advice when the evidence is incomplete is a reliability and safety problem. Systems should perform consistently in intended conditions, handle unusual inputs safely, fail without unacceptable harm, resist misuse, and support testing, monitoring, and recovery. Unsafe failures, inconsistent performance, adversarial inputs, or inadequate testing point here.
Privacy and security protect data and the system
Privacy governs appropriate data use. Security protects data and systems from unauthorized access. A support assistant that includes another customer’s account details in a response has failed both. Protect personal and confidential information, stored and transmitted data, endpoints, credentials, connected tools, and consent. Personal data, permissions, leakage, credentials, or unauthorized access usually indicate privacy and security.
Inclusiveness asks who the system leaves out
A customer assistant that supports typed input and offers no practical option for people who cannot use a keyboard excludes users. Inclusive AI considers different abilities and backgrounds, supports alternative interaction methods, avoids unnecessary barriers, and includes affected communities in design and testing. Accessibility, accommodation, or excluded users point to inclusiveness.
Transparency helps people understand the system
An assistant that presents an AI-generated eligibility decision as a final human judgment hides what the system is. Users should know they are interacting with AI, what it can and cannot do, what information influenced an output, when uncertainty matters, and how to question or appeal. Explanations, disclosure, traceability, or communicated limitations usually indicate transparency.
Accountability stays with people and organizations
An autonomous agent that issues refunds, with no team that owns its policies or reviews questionable actions, has no accountable owner. Assign ownership, define approval and escalation paths, review high-impact decisions, audit behavior, and correct failures. Ownership, oversight, governance, auditing, or responsibility point to accountability. Responsibility cannot be delegated to the model.
One failure can involve several principles
A voice assistant that performs poorly for some accents and gives no warning about its limitations can raise more than one principle. Identify the specific defect in the question.
| Problem | Principle |
|---|---|
| Unequal recognition performance | Fairness |
| Some users cannot complete the task | Inclusiveness |
| Limitations remain undisclosed | Transparency |
| No process exists for reviewing failures | Accountability |
| Scenario | Principle |
|---|---|
| A model reveals confidential records | Privacy and security |
| A hiring model disadvantages one demographic group | Fairness |
| A chatbot conceals that its answers come from AI | Transparency |
| A vision application cannot be used by people with impaired vision | Inclusiveness |
| An automated decision has no assigned human owner | Accountability |
| A system behaves dangerously when given an unexpected input | Reliability and safety |
Bring this to your team
Level Up can run a personalized Azure AI Fundamentals workshop for your team. Tell us a little about your group and we’ll reply with pricing.