B) Supervised learning — the model learns from labeled examples (photo + correct answer). Unsupervised = no labels; reinforcement = learning by reward from actions; rule-based = humans write the rules by hand.

It will fail more often for children — the training data didn't represent them. This is dataset bias: models are only as fair as their data.

A false positive — the filter said "spam" (positive) when the true answer was "not spam." The opposite (letting real spam into your inbox) is a false negative. Knowing these two words is essential for Round 1.

C) The car's future decisions haven't happened yet — you can't train on data that doesn't exist. Training data is always collected beforehand. A, B, and D are all valid training data.

A hallucination — the model generates fluent, confident text that isn't grounded in truth. This is why you always verify AI output against reliable sources. Recognizing this is a core Responsible-AI idea Round 1 rewards.

Overfitting — the model memorized the training set instead of learning general patterns. Fixes: get more/varied data, simplify the model, use regularization, or hold out a validation set to stop training earlier.

df.isnull().sum() — one line, shows the count of missing values per column.

To measure how well the model generalizes to unseen data. If you test on the same data you trained on, a model that simply memorized would look perfect — the test set is the honest check.

It randomly splits your features X and labels y, holding out 20% for testing and keeping 80% for training. It returns four pieces: X_train, X_test, y_train, y_test.

Distribution shift (a.k.a. domain shift) — the real-world phone photos differ from the clean training images in lighting, angle, and quality. Fix: train on data that looks like the real deployment conditions.

Split train/test (stratified), handle missing values & encode categoricals, baseline (logistic regression), stronger model (random forest / gradient boosting), evaluate with precision, recall, and F1, not accuracy: predicting "no" for everyone already scores 92% accuracy while catching zero positives. A confusion matrix + ROC-AUC completes the honest picture.

Data augmentation (flip, rotate, crop, adjust brightness to multiply your effective dataset) and transfer learning (start from a model pretrained on millions of images, then fine-tune on your 500). Both are standard Round-2 moves.

It's cautious: when it does flag something as positive, it's usually right (high precision) — but it misses many real positive cases (low recall). For fraud or disease detection, low recall is dangerous because you let true cases slip through.

Root cause: biased training data — the model learned the historical human bias baked into past hiring. Fixes: remove gender and its proxies (e.g. certain clubs/keywords), rebalance the training data, and audit the model's decisions across groups before deploying. Naming the data (not "the algorithm") as the source is the mature answer.

Training teaches the model. Validation is used repeatedly to tune settings (which model, how long to train). Test is touched only once at the very end for an honest final number — if you tune against the test set, you secretly overfit to it and your reported score is a lie. Explaining this cleanly is a fast way to earn judge trust.

Under our Responsible AI Policy, AI tools are allowed in projects with disclosure — but you must be able to explain every line you submit. "I don't remember" signals you didn't understand your own solution, which is exactly what judges screen out. Know your whole codebase, and disclose your tools up front.

(1) The problem — what real need it addresses; (2) a live or recorded result — the model actually working on real input; (3) one honest limitation — where it fails and what you'd fix next. Judges reward honesty about weaknesses far more than a polished claim of perfection.