Artificial intelligence did not begin with systems that learned from billions of examples. Many early AI systems depended on rules written explicitly by people. Modern AI takes a different approach: instead of describing every decision in advance, we train mathematical models to discover useful patterns from data, feedback and experience.
That change — from programming individual decisions to training systems from examples — is one of the most important shifts in modern computing.
It has enabled systems that can recognise images, transcribe speech, translate languages, recommend products, detect unusual activity and generate text. But it is also important to be precise about what “learning” means.
AI does not learn in exactly the same way a person does, and a model does not necessarily possess human-like understanding. Instead, training adjusts large collections of numerical parameters so that the system becomes better at producing useful outputs for particular tasks.
Older systems often tried to encode knowledge as rules. Modern machine learning usually learns statistical relationships from examples.
From hand-written rules to learned patterns
Rule-based systems were an important part of early artificial intelligence. Developers and subject-matter experts attempted to represent knowledge using explicit logical statements.
A simplified rule might look like:
The behaviour is written directly into the system. The machine applies the rule; it does not discover the rule from examples.
This approach can work well when a domain is stable, the decision logic is clear and the number of possible situations is manageable.
The difficulty appears when reality becomes uncertain or highly variable. If every exceptional case requires another rule, the system can become increasingly difficult to maintain.
Rules can be inspected, tested and traced directly to explicit decision logic.
Unexpected situations may fall outside the rules developers anticipated.
Teams can specify exactly which conditions should trigger a particular action.
Complex environments may require large numbers of overlapping rules and exceptions.
Machine learning changes the question
Machine learning asks a different question.
Instead of:
What instructions should we write so the computer knows what to do?
machine learning asks:
What examples, objective and learning process will allow the model to discover a useful pattern?
Machine learning has a much longer history than the recent AI boom, but improvements in data availability, algorithms and computing power have made increasingly capable models practical at much larger scales.
The trained model can then apply what it has learned to new inputs it did not see during training.
A simple example: detecting spam
Consider an email spam detector.
A rule-based version might contain instructions such as:
- flag messages containing certain phrases;
- flag messages from specific senders;
- increase the risk score when particular links appear;
- allow messages from known trusted domains.
A machine-learning system can instead be trained on many examples of messages labelled as spam or not spam. During training, the model learns statistical relationships between characteristics of those messages and their labels.
The important distinction is that the relevant relationships are learned from examples rather than all being specified manually.
A trained model does not usually keep rewriting itself every time it sees new data. Many production models remain fixed until a controlled retraining or updating process takes place.
Deep learning: learning representations through layers
Deep learning is a branch of machine learning based on neural networks with multiple processing layers.
Neural networks are sometimes described as “mimicking the brain”, but that comparison should not be taken too literally. Their historical inspiration includes biological neurons, yet modern neural networks are mathematical computational systems rather than digital replicas of a human brain.
Their power comes from learning useful internal representations across multiple layers.
In computer vision, earlier layers may become sensitive to relatively simple visual features, while deeper layers combine information into more useful representations for the final task.
In language models, words or tokens are represented numerically, transformed through many layers and used to predict likely continuations or other outputs.
How modern AI learns: the main approaches
There is no single way in which every AI system learns. Different tasks use different training approaches, and many real systems combine several techniques.
Models learn from examples where the desired output is already known, such as images labelled with their correct category.
Algorithms discover structure, clusters or relationships without being given a correct label for every example.
Training creates useful prediction tasks from the structure of the data, making it possible to learn from enormous unlabelled datasets.
An agent learns through interaction by optimising behaviour according to rewards, penalties or other feedback signals.
1. Supervised learning
Supervised learning uses examples that contain both an input and a known target.
For example:
The model produces a prediction, compares it with the known target and adjusts its parameters to reduce the difference.
Applications can include image classification, risk prediction, forecasting and many other tasks where labelled examples are available.
2. Unsupervised learning
Unsupervised learning works with data that does not contain an explicit correct label for each example.
The aim may be to discover structure in the data: groups of similar items, unusual observations or lower-dimensional representations.
Imagine a dataset containing customers with different behavioural patterns. An unsupervised algorithm may identify clusters of customers that appear similar without first being given predefined customer categories.
Group examples according to similarities discovered in the data.
Identify observations that differ substantially from expected patterns.
Represent complex high-dimensional information using a smaller number of useful dimensions.
Reveal structure that may not have been obvious before analysis.
3. Self-supervised learning
Self-supervised learning has become particularly important in modern AI because it makes it possible to learn from enormous amounts of data without requiring people to label every example manually.
The training signal is derived from the data itself.
For language models, a simplified task is predicting a missing or subsequent token from surrounding context.
Training teaches the model which continuations are more probable in a given context. Repeating this process at enormous scale produces rich internal representations of language patterns.
For a sentence such as “The cat sat on the ___”, the model may assign a higher probability to “mat” than to an unrelated continuation. The important point is not that the system stores one rule saying cats sit on mats. It learns distributions and relationships across huge numbers of examples.
Large language models learn statistical representations of language and relationships within their training data. That capability can look remarkably like understanding, but it should not automatically be equated with human comprehension or consciousness.
4. Reinforcement learning
Reinforcement learning involves an agent interacting with an environment and receiving feedback connected to its actions.
This approach has been used in areas such as game playing, robotics, optimisation and control.
Modern AI systems can also use reinforcement-learning techniques during post-training, although the precise methods vary between systems.
Why deep learning became so important
Deep learning became transformative because it can learn complex representations directly from large datasets and can scale effectively with additional data, computing resources and model capacity.
But “more data means better performance” is not an unconditional rule.
Larger models can represent more complex relationships, but only when training and data support that capacity.
More data is not automatically better if it is biased, irrelevant, duplicated, inaccurate or poorly governed.
Large-scale optimisation requires substantial processing power, memory, engineering and energy.
A capable training process still needs robust evaluation against the actual purpose and operating context.
The hardware behind modern AI learning
Modern deep learning depends heavily on hardware capable of performing huge numbers of numerical operations efficiently.
Graphics Processing Units, or GPUs, became particularly important because many neural-network calculations can be performed in parallel. Other specialised accelerators, including TPUs and dedicated AI hardware, are also used.
Training and using a model are different stages. Training may require large clusters of specialised hardware, while inference can sometimes be performed using much smaller infrastructure or even on-device hardware, depending on the model.
Does modern AI really understand?
This is where the title of this article needs an important qualification.
Modern AI systems can build extremely rich internal representations. They can respond to context, combine information, generalise across examples and produce outputs that appear highly knowledgeable.
But the word understanding carries different meanings.
A model may demonstrate functional competence — for example, answering a question correctly, translating a passage or reasoning through a structured problem — without possessing human experience, awareness or comprehension in the same sense that a person does.
The model learns patterns and representations from data and training objectives.
The model can generalise those learned representations to new prompts, examples or tasks.
Useful behaviour should not automatically be treated as proof of consciousness, intent or human-like understanding.
Modern AI has moved from explicit rule-following towards increasingly powerful learned representations. Whether those capabilities should be called “understanding” depends on what we mean by the word.
Learning does not remove the need for rules
The shift to machine learning does not mean modern AI services contain no rules.
In practice, AI-enabled services often combine learned models with conventional software, business logic, policies and operational controls.
That combination is often desirable. Deterministic rules can remain appropriate where behaviour must be explicit and predictable, while machine learning can handle problems where useful patterns are too complex to specify manually.
From model learning to service design
Understanding how a model learns is only part of the challenge.
When AI becomes part of a real service, teams must decide how the model interacts with users, operational staff, policies, systems and existing decision processes.
Training data, objective, model behaviour, limitations and evaluation.
User journeys, operational workflows, hand-offs, exceptions and human intervention.
Purpose, controls, monitoring, logs, risk, challenge and decision ownership.
An AI model can be technically impressive while still being the wrong solution for a service.
Teams therefore need to ask:
- What problem are we actually trying to solve?
- Why does the service need machine learning?
- What data was used to train or adapt the model?
- How do we know the output is good enough?
- What happens when the model is uncertain or wrong?
- When should a person override or review the output?
- How will users understand what the system is doing?
- What evidence should be logged?
- How will the model be monitored after deployment?
- Who owns the outcome when an AI-supported decision causes harm?
The ability to learn from data makes AI more flexible than a fixed rule system, but it also creates new questions about evidence, control, explainability and accountability.
From rules to learned systems: the bigger picture
The history of AI is not a simple story in which rules disappeared and learning replaced them overnight.
Rule-based reasoning, statistical machine learning, neural networks, probabilistic methods and optimisation have developed alongside one another for decades. Modern AI systems often combine several of these traditions.
What has changed dramatically is the scale at which learned models can now operate.
Final reflections
Modern artificial intelligence learns by adjusting mathematical models in response to data, objectives and feedback.
That is fundamentally different from writing every decision as an explicit rule.
Supervised learning uses labelled examples. Unsupervised learning looks for structure. Self-supervised learning creates training signals from the data itself. Reinforcement learning uses feedback from interaction. Deep learning provides architectures capable of learning complex internal representations at scale.
Together, these approaches have made AI systems dramatically more capable.
But capability should not be confused with certainty, autonomy should not be confused with accountability, and convincing behaviour should not automatically be interpreted as human-like understanding.
The important shift is not that machines have stopped following rules. It is that we can now build systems whose behaviour is shaped by learned representations as well as explicit instructions.
For organisations adopting AI, understanding that distinction is increasingly important. It helps us decide where learning systems add genuine value, where traditional rules remain safer, and where human judgement must remain part of the service.
Continue reading
To understand the mathematics underneath this learning process, continue with:
The Maths Behind Neural Networks: How Machines Learn Using Numbers →