Artificial Intelligence

Artificial intelligence covers different ways to represent problems, search for solutions, and learn. Studying these approaches together helps show why AI is not one single technique. The AI foundations section introduces the field’s basic distinctions; history of AI provides a route through its changing methods and expectations.

Use this cluster to form a map before specializing. Ask what information a system receives, what task it is meant to perform, and whether its method depends on explicit representations, learned patterns, or both. Then pick a branch: the next step could be machine learning, AI agents, or AI and society.

• Elements of AI

  A broad introductory route for learners without previous AI knowledge or coding.

Machine learning concerns models built from data; AI training and adaptation focuses on how those models are prepared or adjusted. Google’s introductory course covers model building, data, neural networks, evaluation, and production systems. Follow the math for AI link when you want to understand the mathematical tools behind these ideas; this hub points to that connection rather than repeating a full mathematics course.

For the claim-assessment outcome, pay attention to how a model is tested. A training result is not the same as a result on held-out examples, and repeatedly using one test set to guide changes can make it less useful as an independent check. Practice asking what examples the system saw during training and what evidence supports its performance on new cases.

• Datasets: Dividing the original dataset

  Explains training, validation and test sets, including how repeated use can affect evaluation.

Foundation models are one path into broadly trained systems; prompting and context looks at the instructions and information supplied for a task. The Google course introduces large language models through tokens, Transformers, training, and text generation. Use that material to ask what input a system receives and what kind of output it is designed to produce.

Then follow multimodal AI to explore systems that work across more than one kind of media. As you assess a claim about a model, identify the input type, task, and test conditions it refers to. A result about one task or input does not automatically answer a different question. Keep the scope of the claim visible as you move from a demonstration to an evaluation.

• AI competency framework for students

  Its framework includes AI techniques and applications alongside human-centred and ethical dimensions.

AI systems are used in different settings, so it helps to ask who uses a system, who is affected, and what decision or activity it supports. Explore AI for coding as an application area, and embodied AI for systems that perceive and act in physical environments. The AI and society and AI safety and governance pages connect technical choices to impacts, responsibilities, and oversight.

NIST describes trustworthiness as involving several characteristics, including validity, safety, accountability, transparency, explainability, privacy, and fairness. It also notes that these characteristics must be considered in context and may involve trade-offs. When evaluating an AI claim, ask which people and settings were considered, what harms or benefits were examined, and who is responsible for oversight.

• AI competency framework for students

  A human-centred framework connects AI knowledge and system design with ethics and critical judgement.

An AI application includes choices about how a model is accessed, what information it can use, and how its outputs fit into a larger workflow. Explore AI deployment, AI integration standards, and AI agents for those system-level questions. Retrieval-augmented generation is one approach that retrieves external information for a generated answer.

RAG can fail at several distinct points: retrieval quality, answer faithfulness, factual correctness, citation support, and permissions. Checking one point does not settle the others. For local and on-device AI, ask where inference runs and inspect the data path: local apps can offer access to cloud models, so local use alone does not establish where a request is processed. Follow the local AI page to compare these boundaries.

• Offline Operation

  Describes offline use of downloaded models and the operations that still require connectivity.

• Cloud models

  Explains that cloud models can be used through familiar local tools.

Evaluation starts by defining the task and deciding what evidence would answer the question. AI evaluation covers test design, metrics, uncertainty, and whether an evaluation represents the intended use. NIST cautions that benchmark results can rely on assumptions, combine different notions of performance, or leave uncertainty unclear. A score therefore needs its task and test conditions to be interpretable.

When reading an AI claim, separate what was measured from what someone concludes. Ask whether the evaluation used relevant examples, whether its measure matches the claim, and whether the source reports limitations. Then describe what the evidence supports and what remains open. This habit applies to claims about model capability, deployment, or social effects without requiring a ranking of systems.

• Datasets: Dividing the original dataset

  Shows how training, validation and test sets play different roles in model development and evaluation.