Week 1 — Prediction Problems & Data · ~20 min

Using AI in this course: the rules

Free preview

Why this lesson comes first

In this course AI is two things at once:

  • a tool you use, like every other Orglobal course: an assistant that explains code, suggests approaches and drafts functions
  • the subject you study: from Week 7 onwards you build and evaluate AI systems yourself

That double role makes the rules matter more, not less. You cannot judge whether a model is right if you hand the judging to another model. Someone has to frame the problem, choose the baseline, protect the test set, and answer for the result. By the end of this course that is you.

The three stages

StageWeeksWhat AI may doWhat you must do
Foundations1–6Explain concepts, error messages and library documentation, like a patient tutor.Write your own code, do the hand calculations yourself, and run every experiment yourself.
Pair7–11Suggest approaches, draft functions and tests, generate synthetic data and evaluation cases.Read, run and test every line. Explain any line in a spot check. Verify every number.
Lead12 (capstone)Anything a teammate could do.Own the result: the analysis, the evaluation, and a live defence.

Rule 1: Disclose your AI use

Every submission includes an AI-use note. There's a box for it on the submission form:

AI use: asked Claude to explain why my ROC AUC was 1.0
What I checked: found the target column was still in X; removed it, AUC fell to 0.93
What it got wrong: it first suggested more trees; the real problem was leakage

"No AI used" is a fine answer. Honest disclosure never costs marks. Hidden AI use that we discover in a spot check is treated as not your work.

Rule 2: Never put these into an AI tool, or into a training set

  • Passwords, API keys, tokens, the contents of .env files. Your model API key lives in an environment variable and nowhere else.
  • Real people's data: customer records, patient data, phone lists, BVN, NIN, bank details.
  • An employer's private data or code, unless their policy allows it.

The same applies to training data for your own models. Under the Nigeria Data Protection Act 2023 (NDPA), personal data may only be processed for a lawful purpose, kept to what's necessary, and kept secure. A model trained on personal data carries that data with it, and a dataset that contains people needs a legal basis before you train on it. In this course you use built-in datasets, open-licence datasets and synthetic data you generate. When you need realistic people, invent them.

Rule 3: Verify before you trust, and that includes AI-suggested models

AI is fluent and frequently wrong. In machine learning that looks like:

  • scikit-learn arguments that don't exist, or that changed between versions
  • code that scales or imputes before the split (leakage) because that's how many tutorials do it
  • a suggested "better model" judged on training accuracy, or on a test set you have already looked at
  • confident claims about what a metric means, with the classes the wrong way round

An AI-suggested model is still just a hypothesis. It earns trust the same way yours does: it beats a baseline on validation data, through a leak-free pipeline, and gets evaluated once on the held-out test set. Check scikit-learn's own documentation for any function you haven't used before.

Rule 4: No fakes, no impersonation

Don't use AI to fabricate data and present it as real, invent experiment results, or produce text or images that pretend to be a real person or organisation. Synthetic data is welcome as long as you say it's synthetic.

Rule 5: You own what you ship

If your model or assistant gives a wrong answer, "the AI did it" is not an explanation. You chose to use its output. In Week 12 you defend your work live, including the parts AI helped with.

Sign in to track your progress.