Stop 18 · Evaluation, Safety & Responsible AITopics · 9/9

Responsible AI Governance

In short

When an AI system hurts someone, 'the engineer thought it was fine' isn't an answer. Governance means a company has clear rules, reviews and records for every AI system it ships.

Organizational processes (review boards, documentation, audits) that ensure AI systems are built and deployed responsibly at a company level, not just left to individual engineers' judgment.

The blob presents a rolled-up plan to a small committee seated around a table in a housing society office.

Think of it like this

The housing society committee

Want to knock down a wall in your flat? The society committee checks the plan, an engineer confirms it's safe, and it's all written in the register. Nobody does risky work alone. AI governance is that committee for models: review, approve, record, recheck.

Watch it happen

Getting a high-risk model approved

Propose

Risk review

Model card

Approve

Re-review

1/5A team proposes a resume-screening model and states its purpose and who it affects.

The big ideas

What you need to know

Risk-based effort

A meme generator and a loan model don't need the same review. Spend effort where harm could be serious.

Document everything

Model cards and datasheets record purpose, data, limits and test results, so others can check decisions later.

Clear owners

Every model has a named owner who answers for it and a process for complaints and incidents.

Regulation is growing

Laws like the EU AI Act set extra duties for high-risk uses such as hiring. Companies need processes to show compliance.

See it in code

Walk through the code

yaml

1model: resume-screener-v22owner: hiring-ml-team3intended_use: Rank applications for recruiters to review4not_for: Rejecting any candidate without human review5training_data: 2021-2025 applications, names and photos removed6evaluation:7  auc: 0.818  selection_rate_ratio_by_gender: 0.939risk_level: high10human_oversight: A recruiter reviews every shortlist11approved_by: AI review board12next_review: 2027-04-01

1/4What the model is and who answers for it.

A short model card as YAML. The numbers are example values. Real cards also cover limits, known failures and contact details.

Save yourself the pain

Common mistakes

Oops

Writing the model card after launch, to tick a box.

Do this instead

Write it during development. It forces the team to state limits and test fairness before shipping.

Oops

Approving a model once and never looking again.

Do this instead

Data and users change. Set a re-review date and triggers from monitoring.

Your mission

Write a model card

~30 min

Quick check

Why does a model card include a 'not for' field?

Go deeper

1Read a company's public AI usage policy or responsible-AI framework as a real example.
2Learn what a model card / system card typically documents.
3Understand how governance processes fit into the ML lifecycle you've already built (training → deployment → monitoring).