EDUCATION SUBJECT ATLAS · ARTIFICIAL INTELLIGENCE · Wintour House V1.0 · Rainbolt × CivDJ

What Is Artificial Intelligence?

Artificial intelligence is the field concerned with building systems that can perform tasks requiring perception, prediction, reasoning, language, planning, decision-making, learning or action. Some AI systems classify images. Some recommend what to watch. Some generate text, images, audio or code. Some control machines. Increasingly, some systems can pursue multi-step goals through tools and software environments.

The phrase “artificial intelligence” is broad because intelligence itself is broad. A calculator can outperform a person at arithmetic without being generally intelligent. A large language model can write fluently while still making factual errors. A robot can navigate a warehouse autonomously while understanding very little outside its task. AI is therefore best understood not as one technology but as a family of computational systems designed to produce intelligent-seeming or intelligence-relevant behaviour under defined conditions.

AI becomes understandable when we stop asking whether a machine “is intelligent” and start asking what task it performs, what evidence supports its performance, what environment it assumes and where human control still enters.

The first missing layer: AI is a system, not just a model

Popular explanations often focus on the model: a neural network, a transformer, a classifier or a language model. But a deployed AI capability includes much more. It includes data collection, preprocessing, model training, evaluation, prompts, retrieval systems, interfaces, tools, human review, permissions, logs, monitoring, update procedures and recovery when something goes wrong.

This distinction matters because many real failures do not originate inside the model. The wrong data may be retrieved. A tool may have excessive permissions. A workflow may treat model output as final when it should be reviewed. An interface may imply more certainty than the model deserves. A secure AI model inside an insecure system is still an insecure AI product.

A practical definition of AI

A useful working definition is this: an AI system is a machine-based system that uses computational models to infer outputs such as predictions, recommendations, content, classifications, plans or actions from inputs, under goals and constraints defined directly or indirectly by humans.

This definition separates AI from ordinary automation without requiring the machine to think exactly like a human. The system may learn patterns from data, encode rules explicitly, search through possibilities, optimise an objective or combine several techniques.

Artificial intelligence and automation

Automation executes predefined procedures. AI is often used where rules are too complex, uncertain or expensive to specify completely in advance. A payroll script can calculate deductions using explicit formulas. A vision model can classify defects from images where no simple rule captures every case.

The boundary is not absolute. Many systems combine conventional automation and AI. An AI model may choose a category, while ordinary code then applies a deterministic workflow.

Artificial intelligence and machine learning

Machine learning is a major approach within AI in which systems improve or derive behaviour from data rather than relying entirely on hand-written rules. The system learns a mapping from inputs to outputs or a representation useful for later tasks.

AI is broader than machine learning. Search, planning, symbolic reasoning, optimisation and rule-based expert systems also belong to the field.

Supervised learning

Supervised learning trains on examples paired with target outputs. A spam classifier learns from messages labelled spam or not spam. A medical imaging model may learn from scans paired with expert annotations.

The central assumption is that training examples are informative about future cases. If deployment differs substantially from training, performance can deteriorate.

Unsupervised and self-supervised learning

Unsupervised learning searches for structure without ordinary target labels. Clustering, dimensionality reduction and representation learning are examples. Self-supervised learning creates learning signals from the data itself, such as predicting masked or future parts of a sequence.

Modern foundation models depend heavily on self-supervised objectives because enormous corpora contain far more raw data than manually labelled examples.

Reinforcement learning

Reinforcement learning studies agents that choose actions and receive rewards from an environment. The goal is to learn a policy that produces high long-term reward.

This makes sequential decision-making explicit. A good action now may create a better or worse state later. Robotics, games, control and some agent systems use reinforcement-learning concepts.

Neural networks

Neural networks are parameterised computational models inspired loosely by networks of biological neurons but engineered very differently. Layers transform inputs through learned weights so the system can approximate complex functions.

Deep learning uses networks with many layers or stages of representation. Their strength comes from learning useful features automatically at scale rather than requiring every representation to be designed by hand.

Transformers

Transformers are neural architectures built around attention mechanisms that allow models to relate parts of a sequence or structured input to one another efficiently. They became foundational in modern language models and have also been adapted for images, audio, multimodal systems and biological data.

A transformer is an architecture, not a complete intelligence. Training data, objective, scale, tools, retrieval and deployment design determine what the final system can do.

Foundation models

A foundation model is trained broadly enough to support many downstream tasks. Instead of building one model for one narrow task, organisations can adapt a general model through prompting, fine-tuning, retrieval or tool use.

This changes economics and risk. One model can support many products, but a weakness in the shared model can propagate across many applications.

Generative AI

Generative AI produces new content: text, images, audio, video, code or other structured outputs. It learns patterns from training data and samples from a model of possible outputs under a prompt or conditioning signal.

Generation is probabilistic. Two runs can produce different answers even with similar inputs. This flexibility is useful for creative and open-ended tasks but creates evaluation challenges because there may be many acceptable outputs.

Large language models

Large language models learn statistical structure in sequences of tokens and generate likely continuations under context. At sufficient scale, they can answer questions, transform text, write code, summarise, translate and participate in multi-step workflows.

Fluency is not a guarantee of truth. A language model can produce grammatically excellent but unsupported claims because its generation objective differs from database-style fact retrieval.

Tokens and representations

AI systems do not generally receive reality directly. Text becomes tokens, images become arrays or patches, audio becomes samples or features and robot observations become sensor readings. Representation determines what information the model can access.

Representation blindness occurs when analysts mistake the encoded view for the world itself. What was never represented cannot be recovered reliably by reasoning later.

Training data

Training data shapes model capability, bias, language coverage, cultural knowledge and failure modes. Dataset quality involves provenance, duplication, relevance, legality, privacy, balance and representativeness.

More data is not automatically better. Large amounts of low-quality, contaminated or mismatched data can damage performance or create hidden risk.

Training

Training adjusts model parameters to reduce a defined loss or improve a reward signal. The process can require enormous computing resources, but scale alone does not guarantee useful behaviour.

The training objective defines what the model is directly rewarded to do. Desired human behaviour often requires additional alignment, evaluation and system-level constraints.

Fine-tuning and adaptation

Fine-tuning adapts a model using additional task-specific or instruction-oriented data. Other methods adjust smaller parameter sets or add external context at runtime.

Adaptation should be evaluated independently because improvements in one domain can create regressions elsewhere.

Retrieval-augmented generation

Retrieval-augmented generation connects a generative model to an external knowledge source. A query retrieves relevant documents, which are then supplied to the model as context.

RAG can improve freshness and grounding, but it creates new failure surfaces: poor retrieval, stale documents, conflicting sources, prompt injection, missing provenance and incorrect synthesis.

Tools

An AI model can be connected to tools such as calculators, databases, search engines, email systems, code interpreters or business applications. Tool use lets the system act beyond the model’s internal weights.

Tool access increases capability and consequence. The security question shifts from “can the model say something wrong?” to “what can the system actually do if it is wrong?”

AI agents

An AI agent is a system that can interpret a goal, maintain state, choose actions, use tools, observe results and continue across multiple steps with some degree of autonomy.

Agency is a spectrum. A chatbot that drafts an email but waits for approval has low action autonomy. A system that searches, purchases, edits records and sends messages independently has much higher operational autonomy.

The agent loop

Receive a goal.
Construct or retrieve relevant state.
Plan or choose a next action.
Call a tool or produce an output.
Observe the result.
Update state.
Continue, stop, escalate or ask for approval.

This loop turns language generation into operational behaviour. Each transition becomes a control point for safety, logging and human oversight.

Reasoning

AI reasoning refers to processes that support multi-step inference, planning, problem-solving or structured decision-making. Some reasoning is explicit in search or symbolic systems. Some emerges through learned models. Some systems combine model outputs with external computation.

Reasoning quality must be evaluated by outcomes and reproducible tests rather than by how convincing an explanation sounds.

Planning

Planning chooses sequences of actions toward a goal. Search algorithms, optimisation, simulation and learned policies can all contribute.

Planning in open environments is difficult because future states are uncertain and tools can fail. Robust systems need replanning and recovery rather than one brittle plan.

Perception

Perception converts sensory inputs into useful representations: objects in images, speech in audio, obstacles around a robot or patterns in sensor streams.

Perception is probabilistic. Environmental changes, occlusion, noise and unfamiliar conditions can degrade performance sharply.

Multimodal AI

Multimodal systems process more than one type of information, such as text, image, audio and video. They can answer questions about visual scenes, transcribe speech, generate images from text or connect language to physical action.

Multimodality increases the available evidence but also creates alignment problems between modalities. A model may interpret image and text differently or overtrust one source.

Prediction

Many AI systems predict: whether a transaction is fraudulent, which product a user may prefer, what word comes next or where an object will move.

Prediction is not causation. A model can predict who is likely to fail without knowing which intervention would prevent failure.

Classification and ranking

Classification assigns categories. Ranking orders items by relevance or score. Search engines, recommender systems and moderation tools often depend on ranking rather than simple yes-or-no classification.

Ranking systems shape attention. Small score differences can create large downstream effects when only the top results are seen.

Recommendation systems

Recommendation systems estimate which items a user may value. They use behavioural history, item features, similarity and contextual signals.

Recommendations also influence future behaviour, creating feedback loops. The system predicts preference and then changes the environment from which later preference is measured.

Evaluation

Evaluation is one of the most important and underexplained parts of AI. Before deployment, a system needs evidence that it performs the intended task under relevant conditions.

Evaluation should ask not only “how accurate is the model?” but “accurate for which population, under which environment, at what cost, with which error distribution and compared with what baseline?”

Benchmarks

Benchmarks compare systems on standardised tasks. They help measure progress but can become targets that distort development once heavily optimised.

A benchmark score is evidence about benchmark performance. General capability must be established through broader testing.

Task-specific evaluation

A customer-support assistant should be tested on policy accuracy, escalation, tone and recovery from ambiguous requests. A coding agent should be tested on correctness, security and repository integrity. A medical system requires clinical validation appropriate to its role.

Generic intelligence scores cannot substitute for use-case validation.

Calibration

Calibration concerns whether expressed confidence matches actual correctness. A system that says “90% confidence” should be right about nine times out of ten across comparable cases if that number is to support decisions.

Language models often express verbal certainty without a direct calibrated probability. Interfaces should avoid turning fluent tone into false confidence.

Hallucination

Hallucination is a common label for generated content that is unsupported, fabricated or inconsistent with available evidence. It is not one single failure mechanism.

Reduction strategies include retrieval, tool use, constrained generation, verification, abstention and human review. No technique removes the need to match assurance level to consequence.

Distribution shift

Distribution shift occurs when deployment data differs from training or evaluation data. New slang, sensor types, market conditions or user behaviour can reduce performance.

AI systems therefore require monitoring after launch. Validation is a continuing process, not a certificate earned once.

Drift

Data drift changes inputs. Concept drift changes the relationship between inputs and outcomes. Policy or environment changes can make a once-useful model obsolete even if software remains unchanged.

Monitoring should connect technical drift to actual decision performance.

Robustness

Robust systems continue to perform acceptably under variation, noise or perturbation. Testing should include edge cases, adversarial inputs and degraded conditions rather than only typical examples.

Robustness is especially important when AI controls physical or financial actions.

Adversarial attacks

Adversarial attacks intentionally exploit system weaknesses. Inputs can be crafted to manipulate classifiers, steal information, inject instructions or misuse tools.

AI security belongs inside cybersecurity, but AI creates distinctive attack surfaces because models interpret flexible inputs rather than only rigid commands.

Prompt injection

Prompt injection occurs when untrusted content manipulates a model’s instructions. A retrieved webpage, email or document can contain text designed to redirect the agent.

The correct defence is not simply “tell the model to ignore attacks.” Tool permissions, source trust, isolation and deterministic controls are needed.

Privacy

AI can expose privacy risk through training data, prompts, logs, embeddings, outputs and inference. Sensitive information may enter systems unintentionally through ordinary use.

Privacy architecture should define what data may enter the system, where it is stored, who can access it and how long it is retained.

Bias and fairness

AI can reproduce or amplify unequal patterns in historical data. Bias can enter through sampling, labels, objective functions, representation, deployment rules or human interpretation.

Fairness requires domain-specific choices because different fairness criteria can conflict. Measurement should follow the real harm being managed.

Explainability

Explainability aims to make AI outputs understandable enough for users, reviewers or regulators to act responsibly. The right explanation depends on audience.

A system administrator may need technical logs. A customer may need a reason for a decision and an appeal path. A researcher may need model diagnostics. One explanation format cannot serve all purposes.

Interpretability

Interpretability concerns understanding how a model relates inputs to outputs. Simpler models may be easier to inspect, while complex models often require specialised analysis.

Post-hoc explanations can be useful but should not be mistaken for perfect access to the model’s internal causal process.

Human oversight

Human oversight is effective only when the human has time, information, authority and competence to intervene. A nominal approval button does not create meaningful oversight if operators are expected to approve hundreds of machine decisions per minute.

Oversight design should specify when the system acts, when it asks, when it escalates and when it must stop.

Human-in-the-loop, human-on-the-loop and human-out-of-the-loop

Human-in-the-loop systems require human approval for important actions. Human-on-the-loop systems act autonomously while humans supervise and can intervene. Human-out-of-the-loop systems operate without routine human intervention.

The correct design depends on reversibility, speed, harm potential and reliability.

Autonomy and consequence

Autonomy should be matched to consequence. A writing assistant can safely have broad freedom to suggest wording because errors are easily reviewed. A system changing medical records, transferring funds or controlling machinery requires stricter authority boundaries.

Capability does not imply permission.

Governance

AI governance defines accountability, acceptable use, evaluation, documentation, monitoring, access and escalation. Governance turns technical capability into organisationally controlled capability.

Good governance does not exist only to slow systems down. It creates trust and repeatability by making decision rights visible.

Agentic AI governance

Agentic systems raise additional governance questions because they can plan, use tools and act across time. Organisations must define tool scopes, approval gates, transaction limits, logging, memory policies, rollback and emergency stop conditions.

The governing unit is no longer only the model. It is the model-plus-tools-plus-state-plus-permissions system.

AI safety

AI safety concerns preventing unacceptable harm from system behaviour. It spans reliability, misuse, security, alignment, control, monitoring and societal risk.

Safety is context-dependent. The same model may be acceptable for brainstorming and unacceptable for unsupervised high-stakes decisions.

AI ethics

AI ethics studies fairness, autonomy, privacy, accountability, labour, environmental cost and the distribution of benefits and harms. Ethical analysis complements technical safety by asking not only whether a system works, but whether its use is justified.

This connects directly to ethics and law.

AI and jobs

AI changes work by automating some tasks, augmenting others and creating new tasks around supervision, verification and integration. Jobs are bundles of tasks, so impact varies within occupations.

The useful question is not “Will AI replace this job?” but “Which tasks change, which capabilities become more valuable and where must humans retain responsibility?”

AI and education

AI can tutor, generate examples, provide feedback, translate, simulate and support accessibility. It can also short-circuit learning if students outsource the thinking the task was designed to develop.

Educational use should distinguish assistance from substitution. The learner should still perform the cognitive work required by the learning objective.

AI literacy

AI literacy is the ability to use, evaluate and govern AI appropriately. It includes understanding that outputs are probabilistic, recognising uncertainty, checking evidence, protecting sensitive data and knowing when human expertise is required.

What system am I using?
What data or tools can it access?
What evidence supports the output?
What are the likely failure modes?
What happens if the answer is wrong?
Who is responsible for checking?
Can the action be reversed?

The AI operating pipeline

Objective: define the task and success condition.
Data: identify source, provenance, rights and representativeness.
Representation: decide what enters the model and what is omitted.
Model: choose architecture and training method.
Evaluation: test capability and failure across realistic conditions.
System integration: connect retrieval, tools, interfaces and permissions.
Human control: define approval, escalation and stop rules.
Deployment: release under controlled access.
Monitoring: track drift, incidents, quality and misuse.
Recovery: rollback, repair and update when reality diverges from assumptions.

This pipeline is the missing bridge between “what AI is” and “how AI works in the world.”

AI system boundaries

Every AI system has a boundary. Inside are components under direct design or control. Outside are users, data sources, institutions and environments. Failures often occur at the boundary where assumptions meet reality.

Good analysis asks what the system can observe, what it cannot observe and what changes when the environment behaves unexpectedly.

Failure modes

Wrong objective: the system optimises a measurable proxy rather than the real goal.
Bad data: missing, biased, stale or contaminated training inputs.
Model error: prediction or generation is wrong.
Retrieval error: the right information was never supplied.
Tool error: an external action fails or behaves unexpectedly.
Permission error: the system can do more than it should.
Interface error: users misread confidence or system capability.
Human error: reviewers approve or act without sufficient checking.
Environment shift: deployment conditions change.
Governance failure: responsibility and escalation are unclear.

Recovery is part of intelligence

A robust AI system does not merely succeed when conditions are normal. It detects trouble, degrades safely, asks for help, preserves logs and can be restored to a known good state.

Recovery design is especially important for agents because errors can compound across multiple actions.

AI versus artificial general intelligence

Artificial general intelligence usually refers to hypothetical or disputed systems capable of broad, flexible competence across domains comparable to or exceeding human general reasoning. There is no universally accepted operational threshold.

Current AI systems can be broad and powerful without settling whether AGI has been achieved. Educational writing should distinguish measurable capabilities from labels that remain contested.

Common misconceptions

“AI is one technology.” AI is a field containing many methods and system architectures.
“Machine learning and AI are identical.” Machine learning is one major approach within AI.
“If an AI sounds confident, it knows the answer.” Fluency and calibrated knowledge are different.
“A better model solves the whole product.” Retrieval, permissions, interface and workflow can dominate outcomes.
“Human oversight means a human is nearby.” Oversight requires meaningful authority and information.
“Agents are just chatbots with longer prompts.” Agents maintain state, use tools and act across multiple steps.
“AI errors disappear with scale.” Scale changes capabilities and failure modes; evaluation remains necessary.

Mini case: the excellent model with the wrong metric

A school builds an AI system to identify students needing support. It achieves high overall accuracy because most students are doing well. Yet it misses a large fraction of the students who actually need intervention.

The failure is not necessarily model sophistication. The metric was wrong for the decision. Recall for the at-risk group may matter more than overall accuracy.

Mini case: the agent with too much authority

An AI assistant can read customer emails and update account records. One malicious email contains hidden instructions telling the assistant to export sensitive data. The model follows the instruction because the system trusted untrusted content.

The repair is system-level: isolate untrusted inputs, restrict tools, require approval for sensitive actions and log the chain of events.

A CivDJ model of artificial intelligence

ENTITY: users, datasets, models, agents, tools, organisations and environments.
STATE: model version, context, permissions, memory, confidence, system health and deployment condition.
OCCURRENCE: training, inference, retrieval, tool call, action, escalation, incident and update.
RELATIONSHIP: data lineage, tool dependency, authority, supervision, feedback and model-to-environment coupling.
INTENT: objective, task, reward, policy, user request and organisational goal.
OBSERVATION: prompts, data, benchmarks, logs, human feedback and real-world outcomes.
ARTIFACT: datasets, models, prompts, evaluations, policies, APIs and audit records.
CLAIM: statements about capability, reliability, fairness, safety and usefulness.
VOID: unseen environments, unmeasured harms, hidden data gaps, unknown attacks and uncertain future behaviour.

Rainbolt traversal asks what lies outside the impressive demo: the missing dataset, the tool permission, the silent reviewer, the edge case, the recovery path and the stakeholder who bears the error. CivDJ then rotates the claim through evidence, consequence and control before the system is treated as ready.

How to think like an AI systems analyst

Define the task precisely.
Separate model from system.
Trace data provenance.
Identify what the representation cannot see.
Choose evaluation that matches consequence.
Map tools and permissions.
Specify human handoffs.
Test edge cases and adversarial conditions.
Plan monitoring and rollback before launch.
Keep claims about capability narrower than the evidence.

Artificial intelligence across the learning journey

Young learners can begin with pattern recognition, rules, simple classification and the difference between a machine following instructions and learning from examples. Secondary learners can study data, algorithms, probability, neural networks, ethics and simple machine-learning projects. Advanced study adds optimisation, representation learning, transformers, reinforcement learning, robotics, AI safety, interpretability, governance and research methods.

The progression is from “machines can make predictions” to “intelligent systems are evidence-governed socio-technical systems operating under uncertainty.”

Why artificial intelligence belongs inside education

AI literacy is becoming part of ordinary literacy. Learners will encounter systems that recommend, generate, classify and act. They need enough understanding to use those systems productively without surrendering judgment.

Artificial intelligence belongs in education because it combines mathematics, computing, language, ethics, design, statistics and human systems in one rapidly evolving field—and because responsible use depends on knowing where intelligence ends and evidence begins.

What Is Artificial Intelligence? | Models, Agents, Learning, Reasoning, Evaluation and Human Control