Artificial Intelligence

Artificial Intelligence (AI) is the simulation of HUMAN-LIKE INTELLIGENCE by computer systems. The hallmarks:

• Pattern recognition — recognise faces, classify emails, identify cancer in scans.
• Decision-making — choose the best action given uncertain inputs.
• Language understanding — interpret natural-language input, generate human-like text.
• Learning — improve performance with more data or feedback.
• Adaptation — adjust behaviour as conditions change.

Components of an AI system.

• Data. The fuel — text, images, sensor readings, historical decisions. Quality and bias of data heavily affect outcomes.
• Learning algorithm or rule base. The 'how it knows' — either rules from human experts, or patterns learned from data.
• Reasoning / inference. Drawing conclusions from inputs given the rules or learned patterns.
• Adaptation / feedback. Improving over time as more data arrives.

AI ≠ robot. AI is the SOFTWARE / decision-making side. Robots are physical machines. AI runs in your spam filter, Netflix recommendations, Google Maps, voice assistants — none of those need a body.

Cambridge tip. Mark scheme expects 'simulates human-like intelligence' AND at least two named components. Pure 'AI is intelligent computers' answers score poorly.

An expert system is a kind of AI that captures the knowledge of HUMAN EXPERTS in a specific domain (medicine, law, engineering, geology) and uses it to give advice or diagnosis.

The four components.

1. Knowledge base. A large database of facts and rules elicited from human experts. For medical diagnosis: symptoms, conditions, test results, drug interactions, the relationships between them. Typically encoded as IF-THEN rules.

2. Inference engine. The reasoning component. Applies the rules in the knowledge base to the user's inputs to draw conclusions. Uses forward-chaining (start from facts, derive conclusions) or backward-chaining (start from goal, work back).

3. User interface. Lets the user enter information (symptoms, observations, history) and read the system's conclusions and recommendations. Usually question-and-answer style.

4. Explanation system. Tells the user WHY a particular conclusion was reached — which rules fired, which facts mattered. Crucial for medical, legal, financial uses where blind trust isn't acceptable.

Famous examples. MYCIN (medical diagnosis, 1970s). DENDRAL (chemistry). Modern expert systems power some clinical decision-support tools, tax-prep software and engineering diagnostics.

Cambridge tip. Mark schemes expect ALL FOUR components. The 'explanation system' is the one most commonly omitted — and the one Cambridge marks most heavily.

Machine learning (ML) is a different approach to AI. Instead of being PROGRAMMED with explicit rules, an ML system LEARNS patterns from large amounts of training data.

The basic recipe.

1. Collect training data. Pairs of inputs and known outputs (supervised learning), or just inputs (unsupervised), or rewards from interactions (reinforcement learning).
2. Choose a model. A function with adjustable parameters — a decision tree, a neural network, etc.
3. Train. An optimisation algorithm tunes the parameters so the model's predictions match the training data as closely as possible.
4. Evaluate. Check how well it performs on data it hasn't seen.
5. Deploy. Put it into production; collect more data; iterate.

Where ML wins. Tasks where the rules are too complex or vague for humans to write down — image recognition, language translation, voice transcription, recommending what to watch.

Where ML struggles.

• Bias from training data. If the training data over-represents some groups, the model produces biased decisions affecting people unfairly. Famous failures in hiring, lending, sentencing tools.
• Poor edge-case handling. ML models extrapolate badly outside their training distribution.
• Black-box decisions. Many ML models can't explain WHY they reached a conclusion.
• Hallucinations. Generative AI fluently produces plausible but wrong answers.
• Energy / data hunger. Training large models takes huge compute and electricity.

Cambridge tip. Don't conflate AI and ML. AI is the umbrella; ML is one approach to AI; expert systems are another (rule-based) approach.

Advantages.

• Processes vast amounts of data in seconds — finds patterns humans would miss.
• Operates 24/7 without fatigue, mood swings or distraction.
• Lower cost per decision for routine tasks (fraud detection, image triage, language translation).
• Improved diagnoses — AI screening tools have helped detect cancers earlier in some studies.
• Personalisation — tailored content, recommendations, education.
• Accessibility — speech-to-text, visual descriptions, real-time translation.

Disadvantages.

• Bias and discrimination. AI inherits the biases in its training data; biased decisions can affect hiring, lending, sentencing, healthcare unfairly.
• Job displacement. Cognitive jobs (translation, customer service, paralegal work, parts of medicine and finance) are increasingly automatable.
• Accountability gap. When an AI gets a decision wrong, who is responsible? The user, the deployer, the developer, the data provider?
• Privacy concerns. AI systems often collect more data than users realise.
• Environmental cost. Training large models takes huge amounts of electricity (and water for cooling).
• Hallucinations. Generative AI produces fluent text that can be confidently wrong.
• Misuse risks. Deepfakes, autonomous weapons, personalised disinformation.

Cambridge tip. Top-band evaluation answers acknowledge BOTH the productivity gains AND the bias/jobs/accountability costs. One-sided answers (either techno-utopian or AI-as-existential-threat) lose marks.