ML7: Responsible AI: fairness, explanations, privacy and security
Oversight, documentation and accountability
🎯You'll be able to
Write a model card, decide where people must stay in the loop, and set up a way for affected people to question a decision.
🏠The everyday picture
A new medicine comes with a leaflet: what it's for, who shouldn't take it, side effects, and what to do if something goes wrong. A model needs the same leaflet, and someone responsible for it.
📘What it really is
-
A model card records: intended use and uses it's not for; the data and dates; the measures, with baseline and ranges; results by group; limitations; ethical considerations; the owner; and the version. The kit writes
models/model_card.mdfrommetrics.json, so the figures can't drift apart from the real ones. -
Datasheets do the same for datasets: where the data came from, consent, what's missing.
-
Human oversight: decide which decisions a person must make or confirm. DakDost's flag is only a prompt for a call; nothing is refused or charged because of it. Higher-stakes uses (credit, jobs, health) need a person to review, and a way to override.
-
Contestability: a person affected should be able to ask why and get a human review.
-
Rules are growing. The EU AI Act sets duties by risk level for systems used in the EU; India's DPDP Act covers personal data; sector regulators add their own rules. Check the current requirements for your use; they change.
-
Accountability: name an owner for each model, a review date, and what triggers a review (drift, a complaint, a new use).
📝A worked example
Input: the reference solution's models/model_card.md, written by python -m dakdost.train (2 November 2026).
Expected output: sections for intended use ("a prompt for a person, not an automatic decision: no parcel is refused, charged more or
delayed because of this score"), uses it's not for (pricing, refusing service, judging couriers), the data and split dates, a table of
June measures (baseline 0.0968 PR-AUC; logistic regression 0.5818; this model 0.6235, precision 0.6647, recall 0.5349 at threshold 0.3),
results by zone, limitations (synthetic data; recall differs by zone; weather not in the data) and ethical considerations. Project 1's test
10 checks that the card quotes the same PR-AUC and threshold as metrics.json.
🌍Where it's used
- Model hubs: model cards come with published models (ML5 Lesson 5).
- Banks and insurers: model documentation and validation by a separate team.
- Public services: registers of algorithms used in decisions about people, in some countries.
📝Case scenario
A city used an algorithm to help decide which families got extra visits from social workers. Families couldn't find out why they had been chosen, and errors went unnoticed. After complaints, the city published what the tool used, added a review by a senior worker for each case, and gave families a way to ask for an explanation and a correction.
🖱️Try it
Open models/model_card.md after Project 1 works, and add a "Who to contact" line and a "Review date" line. Success: someone outside the
team could tell what the model is for, who owns it and when it will next be checked.
⚠️Common mistakes
-
Cards written once and forgotten. Sign: figures differ from the current model. Fix: generate the card from the metrics on every training run.
-
No owner. Sign: nobody answers when a problem is found. Fix: name an owner and a review date.
- Automating a decision that needs a person. Sign: no override. Fix: human review for decisions with serious effects.
🛠️For your project
Project 1's test 10 checks the model card's sections and figures; the capstone self-check asks who owns the model.
✅Quick check
What does a model card's "Not intended for" section do?
Show the answer
It warns against uses the model wasn't built or tested for.
Why generate the model card from the metrics file?
Show the answer
So its figures always match the real model's.