Enleashed

We use the word data constantly.

Companies collect data.

Scientists analyse data.

Governments publish data.

Artificial intelligence is trained on data.

Our phones generate data.

Sensors stream data.

But what actually is data?

At its simplest:

Data is a representation of something we have observed, measured, recorded or constructed about the world.

That definition is deliberately broad.

Data might be:

a number,
a word,
a photograph,
a sound recording,
a location,
a timestamp,
a transaction,
a sequence of measurements,
an answer to a survey,
a click on a website.

Modern artificial intelligence can learn from all of these.

But before anything becomes data, something has to turn part of the world into information that can be recorded.

That process matters enormously.

From the world to data

Imagine you are standing outside.

The world around you contains an extraordinary amount of information.

There is:

temperature,
wind,
sunlight,
sound,
movement,
humidity,
air pressure,
people,
vehicles,
buildings,
smells,
conversations.

Now place a thermometer beside you.

The thermometer might produce:

18.4°C

An enormously complicated physical environment has been reduced to one number.

That number is data.

But notice what happened.

The world did not naturally contain a little label saying:

temperature = 18.4°C

We used:

an instrument,
a measurement process,
a unit,
a location,
a moment in time

to create that observation.

A useful way to represent this is:

WORLD → OBSERVATION → MEASUREMENT → DATA

Data is therefore connected to reality.

But it is not reality itself.

Data represents something

Consider:

Is that data?

Potentially.

But without context, it tells us very little.

It could mean:

42°C,
42 kilometres,
42 people,
€42,
42 MW,
42 years,
42 website visits.

The number only becomes useful when we know what it represents.

For example:

temperature = 42°C

is more informative.

Add time:

temperature = 42°C at 14:00

Add location:

temperature = 42°C at Location A at 14:00

Now we have a much richer observation.

This gives us an important principle:

Data requires meaning.

A value without information about what it represents can be difficult or impossible to interpret.

Data usually has context

A useful observation often contains several pieces of information.

For example:

Time	Location	Temperature
09:00	A	12°C
10:00	A	14°C
11:00	A	16°C

The temperature values are data.

But so are:

the times,
the locations,
the relationships between the rows.

Sometimes information describing other data is called metadata.

A photograph might have metadata describing:

when it was taken,
where it was taken,
which camera produced it,
its resolution,
its file format.

Context can radically change what data means.

Data can represent physical quantities

Some data comes from measuring physical properties.

For example:

temperature,
voltage,
speed,
mass,
pressure,
distance,
sound intensity,
electricity consumption.

Sensors allow machines to convert aspects of the physical world into numerical information.

A temperature sensor might produce:

21.2°C

A smart electricity meter might record:

0.43 kWh

A speed sensor might record:

62 km/h

A GPS system might estimate:

latitude = 53.3498

longitude = -6.2603

These observations create numerical representations of physical conditions.

Data can represent human activity

Much of the world's data does not come directly from physical sensors.

It comes from people and the systems they interact with.

For example:

purchases,
searches,
messages,
journeys,
votes,
medical records,
examination results,
employment histories,
social-media interactions.

Suppose someone watches a video online.

A platform might record:

user → video → time → duration → interaction

That event becomes data.

Millions or billions of such events can reveal patterns of behaviour.

Recommendation systems can then learn from those patterns.

Data can be deliberately provided

Sometimes people intentionally provide information.

A survey might ask:

How satisfied are you with this service?

with answers:

1, 2, 3, 4 or 5

A person selects:

Their subjective experience has now been converted into a number.

But something important has happened.

The survey designer decided:

what question to ask,
how to phrase it,
which answers were allowed,
what numerical scale to use.

The data therefore reflects not only the person answering the question.

It also reflects the design of the measurement system.

Data can be created by machines

Machines increasingly generate data about other machines.

A wind turbine might continuously record:

wind speed,
rotational speed,
electrical output,
vibration,
temperature,
component status.

A data centre might record:

processor utilisation,
memory usage,
electricity consumption,
network traffic,
temperature,
cooling demand.

A robot might record:

camera images,
distance measurements,
motor positions,
battery state,
acceleration.

These streams allow machines to observe their own condition and the environment around them.

They also provide the information from which predictive and autonomous systems can learn.

Data can be derived from other data

Not all data is directly measured.

Sometimes we calculate new information from existing observations.

Suppose we record:

distance travelled = 100 km

and:

time taken = 2 hours

We can derive:

average speed = 50 km/h

The speed is still data.

But it was calculated rather than directly observed.

Likewise, a company might calculate:

average customer spending,
monthly revenue,
growth rates,
risk scores.

An AI system might create:

embeddings,
classifications,
predicted probabilities,
estimated states.

So datasets often contain a mixture of:

observed data

and:

derived data

Understanding which is which can be important.

Data can describe things that are difficult to observe directly

Not everything we care about can be measured directly.

Consider:

intelligence

How do we measure it?

Or:

creditworthiness

Or:

health

Or:

social deprivation

Or:

customer satisfaction

These are complex concepts.

We often create measurable variables that attempt to represent them.

For example, health might be represented using:

blood pressure,
heart rate,
test results,
diagnoses,
reported symptoms.

Creditworthiness might be represented using:

income,
repayment history,
outstanding debt,
missed payments.

The measurable variables act as representations of a more complicated underlying concept.

This introduces an important problem:

What we can measure is not always identical to what we actually care about.

A dataset is a collection of observations

When many observations are organised together, we have a dataset.

Suppose we measure electricity demand every hour:

Time	Demand
08:00	30 GW
09:00	32 GW
10:00	34 GW
11:00	35 GW

Each row is an observation.

Together they form a dataset.

We can then use the dataset to ask questions.

For example:

When is demand highest?
How quickly does demand change?
Does demand follow daily patterns?
Can tomorrow's demand be predicted?

Datasets allow observations from the past to become evidence for reasoning about the future.

But datasets do not simply appear

This is one of the most important ideas in this module.

A dataset is the result of decisions.

Someone or something decided:

what to measure

whom or what to measure

when to measure

where to measure

how to measure

how often to measure

how to store the result

what not to record

Consider a dataset of household electricity consumption.

Someone had to decide:

which households were included,
what type of meter was installed,
how frequently measurements were taken,
whether appliance-level information was recorded,
how missing readings were handled,
whether location was stored,
whether household characteristics were collected.

The resulting dataset reflects those decisions.

Data is not a naturally occurring neutral substance

We sometimes talk about data as though it were a raw material waiting to be discovered.

You may hear phrases such as:

Data is the new oil.

But data is different from a naturally occurring resource.

Oil exists underground whether or not anyone measures it.

A dataset exists because some system produced a representation of something.

That system might involve:

sensors,
databases,
software,
surveys,
administrative processes,
people,
categories,
definitions.

This means data has a history.

Before asking what a dataset tells us, we should ask how the dataset came to exist.

What we choose to measure matters

Imagine trying to understand a school.

We could collect:

examination results,
attendance,
class size,
teacher experience,
student age.

A model trained on this information might predict examination performance.

But perhaps we do not record:

whether students have somewhere quiet to study,
whether they have caring responsibilities,
whether they are hungry,
whether they enjoy the subject,
whether they feel safe,
whether they have access to tutoring.

The model cannot learn directly from information that was never recorded.

This gives us a principle that will appear repeatedly throughout the course:

What is missing from a dataset can matter as much as what is present.

Who gets measured matters

Suppose we want to understand how people travel around a city.

We collect location data from a smartphone application.

We might obtain millions of observations.

That sounds impressive.

But who uses the application?

Perhaps:

younger people use it more,
some people do not own smartphones,
some users disable location tracking,
tourists use it differently from residents.

The dataset may be enormous and still fail to represent the whole population.

More data does not automatically mean more representative data.

We will explore this later when we examine sampling and selection.

When we measure matters

Suppose we measure traffic only between:

10:00 and 15:00

We might conclude that the road network rarely experiences serious congestion.

But we missed:

the morning commute,
school drop-off,
the evening commute.

Our conclusion may be wrong because the measurement window excluded the periods that mattered.

Time is therefore part of the data-generating process.

As we saw in Module 1:

Every observation belongs to a particular moment.

Where we measure matters

Suppose a country has 100 weather stations.

If 90 are located in cities and only 10 are located in rural areas, the resulting dataset represents some places much better than others.

The same issue can occur with:

pollution sensors,
traffic cameras,
healthcare records,
broadband measurements,
electricity meters.

Spatial coverage shapes what the dataset can tell us.

Again:

The data we have depends partly on where we decided to look.

How we measure matters

Suppose two thermometers measure the same room.

One reports:

20.0°C

The other reports:

21.3°C

Which is correct?

Perhaps one sensor is poorly calibrated.

Perhaps one is beside a radiator.

Perhaps one measures every second while another measures every ten minutes.

Perhaps they have different levels of precision.

Measurement is not automatically perfect.

Data can contain:

noise,
bias,
errors,
rounding,
delays,
calibration problems.

We will examine measurement error later in this module.

Categories shape data

Suppose a hospital database records patients as:

low risk

medium risk

high risk

Where do those categories come from?

Someone had to define the boundaries.

Perhaps:

0–20 → low

21–50 → medium

51–100 → high

Another organisation might choose different boundaries.

The underlying patients have not changed.

The representation has.

Categories are useful because they simplify information.

But they also impose structure.

And that structure can influence what a model learns.

Labels can contain judgement

Machine-learning datasets often contain labels.

For example:

Email	Label
Message A	Spam
Message B	Not spam
Message C	Spam

The model learns to predict the label.

But where did the label come from?

Perhaps:

a person assigned it,
a rule-based system assigned it,
another model assigned it,
it was inferred from later behaviour.

Labels can therefore contain:

human judgement,
organisational policy,
measurement error,
historical decisions.

A model trained on labels learns the patterns associated with those labels.

It does not automatically know whether the labels themselves were correct or fair.

Historical data contains history

This is particularly important when data describes people and institutions.

Suppose we train a hiring model using historical employment records.

The dataset might contain:

who applied,
who was interviewed,
who was hired,
who was promoted,
who was paid more,
who became a manager.

These are facts about what happened.

But what happened historically may have been influenced by:

discrimination,
unequal opportunities,
social norms,
economic inequality,
institutional practices.

The dataset records those outcomes.

An AI system can learn them.

This gives us one of the central themes of the course:

Historical data does not merely contain information about the past. It can contain the structures, choices and inequalities of the past.

That will become extremely important when we later examine algorithmic fairness.

Data can be correct and still mislead us

Imagine a company reports:

Average salary = €60,000.

That number may be perfectly correct.

But perhaps:

most employees earn €35,000,
a small number of executives earn millions.

The average is accurate.

Yet it may create a misleading impression of what a typical employee earns.

Data does not need to be false to mislead us.

The way it is:

selected,
aggregated,
summarised,
presented

can change the story we infer from it.

Later in this module we will examine how data can lie without lying.

Data becomes the world the model sees

For an AI system, the training dataset is extraordinarily important.

The model does not experience the world directly.

It experiences a representation of the world through data.

Conceptually:

WORLD

↓

MEASUREMENT

↓

DATA

↓

MODEL

The model learns patterns in the data.

So if the data:

excludes important information,
overrepresents some groups,
underrepresents others,
contains measurement errors,
contains historical biases,

the model encounters the world through those distortions.

This is why many apparent "AI problems" actually begin much earlier.

They begin with the data-generating system.

More data is not automatically better

Modern AI often benefits enormously from scale.

Large datasets can expose models to more:

examples,
situations,
languages,
images,
behaviours,
patterns.

But quantity is not the same as quality.

A billion poor measurements do not necessarily become good measurements because there are a billion of them.

A huge biased dataset can still be biased.

A huge historical dataset can still describe a world that has changed.

A huge dataset can still omit the variable that actually matters.

So we should distinguish between:

How much data do we have?

and:

How informative is the data we have?

Data is evidence, not truth

A useful way to think about data is as evidence.

Evidence can be:

strong,
weak,
incomplete,
noisy,
biased,
contradictory.

We use evidence to reason about the world.

But evidence is not identical to the world.

This distinction will become especially important when we move into probability and inference.

We will ask:

Given the observations available to us, what should we believe?

And later:

Given what we believe, what should we predict?

Data sits at the beginning of the AI pipeline

We can now return to the framework introduced in Module 1:

PAST → DATA → MODEL → PREDICTION → DECISION → ACTION → FUTURE

Data sits near the beginning of this chain.

That means decisions about data can propagate through everything that follows.

If measurement is poor:

poor measurement → poor data

If the data is unrepresentative:

unrepresentative data → distorted model

If the model is distorted:

distorted model → distorted predictions

Those predictions may influence decisions.

Those decisions may affect people.

And those actions create the future data from which the next model learns.

So the data-generating process is not merely a technical detail.

It can shape the entire system.

Ask where the data came from

Whenever you encounter a dataset, develop the habit of asking:

What does each observation represent?
What was measured?
How was it measured?
Who or what was measured?
Who or what was not measured?
When were the measurements taken?
Where were they taken?
How accurate are they?
Who defined the categories?
Who created the labels?
What information is missing?
What system produced this dataset?

These questions often reveal more than simply looking at the numbers.

The central idea

Data is how we create usable representations of observations about the world.

But the process is not neutral.

Before a model sees a dataset, decisions have already been made about:

what to observe

what to measure

who to include

when to measure

where to measure

how to represent the result

what to leave out

Those decisions create the informational world from which the model will learn.

Data is not reality. It is a representation of selected aspects of reality, produced through a system of observation, measurement and recording.

This is why understanding artificial intelligence must begin before the model.

It begins with understanding how the world becomes data.

In the next lesson, we will look more closely at that transformation:

measurement and observation — how physical events, human behaviour and abstract concepts become information that machines can process.