Artificial Intelligence
This domain tests what AI is, how machine-learning models are built and judged, and where AI creates risk. CET questions emphasize supervised versus unsupervised learning, classification versus regression, evaluation metrics, data quality, and generative-AI limits.
Core rules and concepts
- Scope: today's systems are narrow AI; general AI remains hypothetical. Rule-based systems follow explicit logic; ML systems infer patterns from data.
- Learning paradigms: supervised learning uses labeled examples; unsupervised learning finds structure without labels (clustering); reinforcement learning learns from reward and penalty.
- Task types: classification predicts a category; regression predicts a number.
- Data splits: training data fits parameters; validation data tunes choices; test data estimates final performance and must not be used for tuning.
- Fit: overfitting means good training performance and poor generalization; underfitting means poor performance on both.
- Metrics: accuracy can mislead on imbalanced data. Precision = TP/(TP+FP); recall = TP/(TP+FN); F1 balances them. A confusion matrix shows all four outcomes.
- Generative AI: large language models predict likely text and can produce fluent but false statements (hallucinations); outputs need verification.
Worked example
A spam filter reports 95% accuracy, but 40% of actual spam reaches inboxes. The dataset is 95% legitimate email. Which metric should guide improvement?
Reasoning: with 95% legitimate mail, a model that labels everything not spam also scores 95% accuracy while catching zero spam. The business cares about missed spam, so recall for the spam class is the key metric; precision matters because false positives hide real mail. The team should inspect the confusion matrix, rebalance or reweight classes, and tune the decision threshold. Reporting only accuracy hides the failure.
Traps
Do not treat accuracy as sufficient on skewed data. Do not tune on the test set. Correlation in training data is not causation. More data does not fix systematically mislabeled data. A chatbot's confident tone is not evidence of truth.