ICDL AI Essentials cheat sheet

AI is the field of building systems that perform tasks associated with human perception, reasoning, learning, language, or decision-making.

Machine learning is a subset of AI; it learns patterns from data instead of relying only on explicit hand-coded rules.

Training adjusts model parameters using data; inference applies the trained model to new input.

Supervised learning uses labeled examples, unsupervised learning finds structure in unlabeled data, and reinforcement learning learns from reward signals.

Neural networks are layered sets of weighted connections; deep learning uses neural networks with many layers.

Generative AI produces new text, images, audio, or code by modeling statistical patterns in training data.

Deployed AI systems are usually narrow: they are scoped to specific tasks rather than general human-like intelligence.

AI outputs are probabilistic; a confident-sounding response or high confidence score is not proof of correctness.

Recommendation systems rank products, videos, or posts by predicting relevance from user behavior and item similarities.

Smartphones use AI for photo enhancement, keyboard prediction, voice transcription, and face or scene recognition.

Navigation tools combine map data with predicted traffic conditions to estimate travel time and suggest routes.

Spam and phishing filters classify messages using patterns in senders, wording, links, and user reports.

Fraud detection flags transactions that deviate from a customer's normal pattern or match known abuse signals.

Machine translation converts text or speech between languages by modeling patterns in multilingual training data.

Voice assistants use speech recognition and natural language processing to interpret requests and trigger actions.

Customer-service chatbots handle routine queries and route complex or sensitive cases to human agents.

Generative AI appears in writing assistants, search summaries, image editing, and code completion tools.

A complete prompt states the task, context, constraints, and desired output format.

Few-shot prompting supplies example input-output pairs so the model can infer the expected pattern.

Complex requests are easier to control when split into smaller steps with a check after each step.

Lower randomness settings favor consistent factual answers; higher randomness favors varied brainstorming.

Model outputs require verification against reliable sources because fluent text can contain fabricated facts.

Rule: enter sensitive personal, financial, or confidential data only when the tool's data handling is approved.

System or role instructions set an assistant's boundaries and response style before user requests are processed.

High-stakes outputs such as medical, legal, or financial decisions require qualified human review before use.

Algorithmic bias is systematic unfairness in AI outputs caused by biased data, flawed labels, or design choices.

Fairness evaluation compares outcomes across groups rather than assuming neutral intent is sufficient.

Transparency makes a system's purpose, data use, limitations, and decision process visible to affected people.

Accountability assigns responsibility for AI-assisted decisions to a person or organization that can explain and correct outcomes.

Data minimization limits personal data collection to what is necessary; informed consent requires voluntary agreement after understanding the processing.

Copyright risk arises when training data or generated outputs reproduce, imitate, or improperly incorporate protected creative work.

Hallucination is fluent but false or unsupported output, so factual claims require verification against reliable sources.

Data leakage and prompt injection are security risks: sensitive prompts can be exposed, and hidden instructions can override intended rules.

Automation bias, deepfakes, and model drift create safety risks through over-trust, deceptive synthetic media, and performance decline after data changes.