Pass 901

Mira Cole

Episode two. Responsible A.I. Microsoft lists six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.

Owen Park

If you only memorize the names, you will still miss questions. The exam gives a story. A hospital model. A hiring tool. A chatbot that will not work with a screen reader. You have to pick which principle is in play.

Mira Cole

Start with fairness. Fairness means the system treats people comparably. Similar candidates get similar outcomes. If a loan model approves one zip code and quietly rejects another with the same credit profile, that is a fairness failure, not a 'the math said so' shrug.

Owen Park

A picture that helps: two people standing on the same starting line, same race. Fairness is not giving one of them a secret head start because the training data was skewed. On the exam, look for words like bias, disparate impact, groups treated differently.

Mira Cole

Inclusiveness is the cousin people confuse with fairness. Inclusiveness is about whether people of different abilities and circumstances can actually use the system. Captions. Screen readers. Multiple languages. A kiosk that only works if you can stand and tap a small screen is an inclusiveness problem even if it scores every standing adult equally.

Owen Park

Write this on a sticky note. Fairness is comparable outcomes across groups. Inclusiveness is can everybody participate. Different questions, different answers.

Mira Cole

Reliability and safety. The system should work as intended, fail predictably, and not cause harm when it is wrong. A medical triage bot that hallucinates a diagnosis is a reliability and safety issue. So is a vision model that misses a stop sign in rain.

Owen Park

Exam language to watch: testing, monitoring after deployment, human review of high-stakes decisions, fallbacks when confidence is low. Safety is not a vibe. It is controls.

Mira Cole

Privacy and security. Data minimization, consent, encryption, access control, not stuffing customer medical notes into a public model. If the story is about leaking data, unauthorized access, or training on data you should not have, it is this principle.

Owen Park

A useful split: privacy is about people and their information. Security is about protecting the system from abuse. Microsoft bundles them because an unlocked database is both.

Mira Cole

Transparency. People should understand what the system is, what it does, and why it produced an output, at a level that matches the risk. Document the model. Disclose that the customer is talking to A.I. Explain a declined application in human language.

Owen Park

Accountability is the other mixed-up pair. Transparency is 'can we see and explain it.' Accountability is 'who owns the outcome.' A named human, a review process, an audit trail. The company does not get to say the model did it.

Mira Cole

Imagine a self-driving shuttle. Transparency is publishing how it decides to brake. Accountability is the operator who is legally on the hook when it does not.

Owen Park

On Foundry, these principles show up as product features, not posters. Content filters. Groundedness evaluation. Identity and role-based access. Tracing. You will meet those later. Today, match the story to the principle.

Mira Cole

Question one. A résumé screener ranks applicants. After launch, women with identical experience are scored lower than men. Which principle is primarily at issue?

Owen Park

Fairness. Comparable people, different outcomes. Not inclusiveness, unless the story was about a screen reader. Not transparency, unless nobody can explain the rank. The harm described is biased outcomes.

Mira Cole

Question two. A city deploys a benefits chatbot. It works for sighted mouse users, but the buttons have no names for a screen reader. Which principle?

Owen Park

Inclusiveness. The bot may be perfectly fair among people who can see the buttons. The problem is who can use it at all.

Mira Cole

Question three. A clinician-facing model sometimes invents lab values that were never in the record. The team adds a confidence threshold and a required human sign-off. Which principle are they strengthening?

Owen Park

Reliability and safety. Hallucinated labs can harm patients. Human review and thresholds are classic safety controls.

Mira Cole

Question four. A product manager wants to paste a spreadsheet of customer emails into a public playground to 'see what the model says.' Legal stops them. Which principle?

Owen Park

Privacy and security. That spreadsheet is other people's data, leaving a controlled environment. Easy point if you do not overthink it.

Mira Cole

Question five. After a credit model declines a customer, the bank cannot explain why, and no team owns the override process. Two principles are weak. Which pair?

Owen Park

Transparency and accountability. Cannot explain the output: transparency. No owner for the override: accountability. If you picked fairness, you invented a bias the question never stated.

Mira Cole

Carry those six into every later episode. Foundry content filters, evaluation, and identity are how the principles become buttons. Next time: how generative models actually work, and how you pick one from the catalog.

Owen Park

Tokens, temperature, context windows, serverless versus managed compute. Bring coffee.

The six responsible AI principles

On the mics