Use this glossary to clean up high-confusion AI foundations language fast. This exam gets easier when you keep data quality, metric choice, model behavior, grounding, and governance in separate buckets.
High-value terms
| Term |
What to remember |
| bias |
systematic skew that leads to distorted or unfair outcomes |
| data leakage |
information in training or evaluation that would not exist at real prediction time |
| evaluation metric |
the number used to judge whether model behavior matches the task |
| F1 score |
precision-recall balance metric that becomes useful when both false positives and false negatives matter |
| GenAI |
systems that generate new content such as text or images |
| grounding |
supplying trusted external context so output is constrained by useful source material |
| imbalance |
class-distribution problem where one outcome appears much less often than another |
| overfitting |
learning training-specific noise too closely and generalizing poorly |
| precision |
share of predicted positives that are actually positive |
| recall |
share of actual positives that the model successfully finds |
| responsible AI |
the discipline of making AI systems safer, fairer, more private, and more accountable |
Common confusion pairs
| Pair |
Keep this distinction clear |
| precision vs recall |
precision asks how many predicted positives were right; recall asks how many real positives were found |
| leakage vs overfitting |
broken evaluation setup versus weak generalization |
| grounding vs fine-tuning |
context supplied at inference time versus behavior changed through additional training |
| bias vs variance |
systematic oversimplification versus instability across data samples |
| model quality vs retrieval quality |
learned behavior versus the quality of external context fed to the model |
| AI service vs AI concept |
a named OCI service is a product boundary; the AI concept is the underlying idea |
Fast recall anchors
| If you see… |
Think… |
| “great score but suspiciously easy win” |
leakage or bad evaluation setup |
| “class imbalance” |
accuracy may be misleading |
| “hallucinated but fluent answer” |
grounding, retrieval, or context-quality issue |
| “fairness or privacy risk” |
responsible AI and governance controls |
| “great train score, weak test score” |
overfitting or poor split logic |
If three terms blur together
| Blurry group |
Reset with this rule |
| accuracy, precision, recall |
overall correctness, false-positive control, and false-negative control are different decision lenses |
| leakage, overfitting, drift |
broken evaluation, weak generalization, and changing production reality are different failures |
| prompting, grounding, fine-tuning |
instruction quality, context quality, and model adaptation are separate controls |
| bias, privacy, safety |
fairness, confidentiality, and harmful behavior are related but not identical risks |
Route next
| Need |
Go here next |
| weekly pacing and weak-lane repair |
Study Plan |
| metrics, data, and GenAI tie-breaks |
Cheat Sheet |
| last-week decision cleanup |
FAQ |
| official Oracle and OCI sources |
Resources |