Back to library ↗SUBJECT COLLECTION · 01
Understand how artificial intelligence and machine learning systems work, how they are evaluated and where their limits matter.
Begin with the wider idea of artificial intelligence, then look closely at machine learning: how data trains a model, what evaluation can reveal and why a strong benchmark is not a guarantee of real-world performance.
01 / COLLECTION NOTES
Artificial intelligence is an umbrella term for computer systems built to perform tasks such as recognizing patterns, making predictions, processing language or selecting actions. Machine learning is one approach within that field: instead of specifying every rule by hand, developers fit a model to examples. Deep learning uses neural networks with multiple layers; generative AI refers to systems that produce text, images, audio or other outputs. These terms overlap, but they are not interchangeable. Naming the specific method and task gives a clearer picture than a product label.
02 / COLLECTION NOTES
A useful way to evaluate a technology claim is to trace the lifecycle. What problem is being solved? What data represents that problem? What output does the system return? How was performance measured before launch, and who monitors it later? A polished demonstration may show that a system can succeed on one prompt or one image. It does not show how often it fails, whether the test matches everyday use or what happens when a person relies on a wrong answer.
03 / COLLECTION NOTES
Machine-learning models reflect their examples and objectives. If important situations are missing from the data, the model may perform unevenly. If a team measures a convenient proxy rather than the outcome people care about, a system can optimize the wrong thing. A single accuracy score can also hide different kinds of mistakes: for a rare but serious event, false negatives and false positives may have very different costs. Evaluation needs to fit the task and the people affected.
04 / COLLECTION NOTES
Start with the artificial-intelligence guide for terminology and a broad mental model. Then read the machine-learning guide for data, training, inference, generalization and common evaluation failures. Together, they answer both “what does AI mean?” and “how do we judge a particular machine-learning claim?” As the collection grows, new guides should fill a specific gap, such as neural networks or model evaluation rather than repeat the same introduction under a new search phrase.
05 / COLLECTION NOTES
Before trusting an automated output, ask what input the system received, what information may be stored, whether the result can be checked against evidence, and who is responsible for fixing an error. For consequential uses, ask how people can appeal a decision and whether the model has been evaluated on the population and conditions where it will operate. These questions apply to generative assistants, recommendation systems and older rule-based automation alike.
Choose an AI product you have used. Write down its input, output, likely failure mode and what evidence would convince you it works for your task. If you cannot identify how the output is checked, that uncertainty is part of your assessment.
These institutional references support the background in this collection. Follow a guide’s individual source links for evidence related to a specific claim.
Learn what artificial intelligence means, how machine-learning systems are built, where generative AI is useful and how to judge its limitations.
Read the guide ↗Understand how machine-learning models learn patterns from examples, how training differs from inference, and how to spot overfitting and weak evaluation.
Read the guide ↗