Prediction versus decision

One of the most important distinctions in artificial intelligence is also one of the most frequently overlooked:

Prediction is not decision-making.

A machine-learning model can estimate what is likely to happen.

That does not tell us what we should do.

Consider a hospital.

A model analyses a patient's medical data and produces the following prediction:

Patient A has an 82% probability of benefiting from treatment.

That might be an excellent prediction.

But suppose the hospital has:

50 patients who could benefit from treatment
10 available beds

The prediction has not solved the hospital's problem.

It has merely provided information that might help solve it.

The hospital now faces a completely different question:

Who should receive the 10 available beds?

That is not primarily a prediction problem.

It is a decision and allocation problem.

What prediction does

Prediction attempts to estimate something we do not yet know.

Given some information (x), a model might estimate an outcome (y):

[ \hat{y} = f(x) ]

For example:

[ \text{patient data} \rightarrow \text{model} \rightarrow P(\text{benefit from treatment}) = 0.82 ]

The model is answering a question like:

Given what we know about this patient, how likely is a particular outcome?

Machine learning is extremely useful for these kinds of problems.

A model might predict:

whether a customer will repay a loan
how much electricity a household will consume tomorrow
whether a machine is likely to fail
how long a journey will take
whether a tumour is malignant
how much solar generation will be available
whether someone will click on an advertisement
how likely a student is to complete a course

These are all predictions.

They describe what the model believes is likely to happen.

They do not, by themselves, determine what should happen next.

Decisions are different

A decision asks:

Given what we know, what should we do?

That introduces things that may not appear anywhere in the prediction model.

For example:

limited resources
competing objectives
costs
benefits
risks
physical constraints
laws
fairness
priorities
uncertainty
the rights and preferences of different people

Return to our hospital.

Suppose the model produces predictions for five patients:

Patient	Predicted probability of benefiting
A	82%
B	79%
C	76%
D	65%
E	61%

If there is one available bed, should Patient A automatically receive it?

Perhaps.

But not necessarily.

What if Patient A requires the bed for three weeks while Patient B requires it for one day?

What if Patient B will die without immediate treatment while Patient A can safely wait?

What if the model is less accurate for Patient C's demographic group?

What if Patient D has a lower probability of benefit but a dramatically larger benefit if treatment succeeds?

What if some patients have already been waiting for months?

What if clinical guidelines impose additional priorities?

Suddenly:

[ \text{highest predicted probability} ]

is not obviously equivalent to:

[ \text{best decision} ]

That gap is fundamental.

Prediction provides information

A useful way to think about this is:

[ \boxed{\text{Prediction informs decisions.}} ]

It does not automatically make them.

The prediction might become one input into a larger decision process:

[ \text{Data} \rightarrow \text{Prediction} \rightarrow \text{Decision} \rightarrow \text{Action} \rightarrow \text{Outcome} ]

Machine learning often concentrates heavily on the first two stages.

Real systems have to deal with all five.

And the difficult social questions frequently appear after the prediction has been made.

A model can be accurate and the system can still be bad

This leads to an important consequence.

Imagine we build an extraordinarily accurate hospital model.

Its predictions are almost perfect.

We deploy it.

But the hospital then uses those predictions through a terrible allocation rule.

Perhaps it simply gives beds to whoever can pay the most.

Or whoever arrived first.

Or whoever has the highest predicted probability of recovery.

Or whichever patient happens to be closest to the hospital.

The prediction model could still be technically excellent.

The overall system could still produce outcomes we regard as:

inefficient
unfair
unsafe
discriminatory
wasteful

This is why evaluating an AI model is not the same thing as evaluating the system in which the model operates.

We have to ask two separate questions:

Is the prediction good?

and:

What are we doing with the prediction?

Those are different questions.

The electricity example

The distinction becomes particularly obvious in electricity systems.

Suppose an AI model predicts that electricity demand between 18:00 and 18:30 tomorrow will be:

[ 42.3 \text{ GW} ]

That may be a very useful forecast.

But the electricity system still has to decide:

which generators should produce
how much each should produce
which batteries should charge or discharge
whether flexible demand should move
how electricity should flow through the network
how congestion should be managed
what happens if there is insufficient supply
how scarce capacity should be allocated
who should pay
who should be rewarded

The prediction:

[ \hat{D}_{18:00} = 42.3 \text{ GW} ]

does not answer any of those questions.

It provides information to a system that must answer them.

The same distinction appears almost everywhere AI is deployed.

Predicting demand is not allocating resources

Consider a city trying to predict where ambulances will be needed.

Machine learning might estimate:

[ P(\text{emergency in location } i) ]

for every part of the city.

Excellent.

But there are only twenty ambulances.

Where should they wait?

That requires considering:

predicted demand
travel times
hospital capacity
geographic coverage
uncertainty
response-time targets
fairness between neighbourhoods
resilience if several emergencies occur simultaneously

The prediction problem is:

Where are emergencies likely to occur?

The decision problem is:

How should scarce ambulances be positioned?

These are related.

They are not the same.

Predicting risk is not deciding who gets a loan

A bank might estimate:

[ P(\text{default} \mid x) ]

for every applicant.

Suppose two people receive:

[ P_A(\text{default}) = 4% ]

and

[ P_B(\text{default}) = 7% ]

The model has estimated risk.

It has not determined:

whether either person should receive a loan
how much they should be allowed to borrow
what interest rate should apply
what level of risk the bank should tolerate
whether lending rules are fair
how regulations should constrain the decision

Those require a decision rule.

For example, someone might create:

[ \text{Approve loan if } P(\text{default}) < 5% ]

Now we have moved from prediction to decision-making.

The 5% threshold did not come from the prediction.

Someone chose it.

That choice embeds assumptions about:

acceptable risk
profitability
regulation
opportunity
fairness

The model estimates.

The surrounding system decides.

Classification thresholds are decisions too

We encountered this idea earlier when discussing classification.

A model might output:

[ P(y=1 \mid x)=0.63 ]

To turn that probability into a binary classification, we might decide:

[ \hat{y} = \begin{cases} 1 & \text{if } P(y=1 \mid x) > 0.5 \ 0 & \text{otherwise} \end{cases} ]

But why (0.5)?

There is nothing magical about it.

If false negatives are extremely dangerous, perhaps we use:

[ 0.2 ]

If false positives are extremely costly, perhaps we use:

[ 0.9 ]

The model gives us a probability.

Humans or another decision system determine what action follows from it.

That is one reason precision, recall, false positives and false negatives mattered so much in the previous module.

We were already beginning to move from:

How accurate is the prediction?

towards:

What happens when we act on it?

Decisions introduce objectives

Prediction usually asks us to estimate some unknown quantity.

Decision-making requires us to decide what outcome we are trying to achieve.

Suppose an electricity network has ten batteries.

What should they do?

We could try to:

minimise electricity cost
minimise carbon emissions
minimise network congestion
maximise battery lifetime
maximise owner revenue
improve reliability
protect vulnerable consumers

These objectives may conflict.

A battery might maximise its owner's revenue by discharging when prices are highest.

But the network operator might want it to discharge somewhere else to relieve congestion.

A household might want to keep the battery full in case of an outage.

There is no prediction model that magically resolves these competing objectives.

Someone has to decide:

What is the system trying to achieve?

This is where AI begins to collide with economics, engineering, ethics, law and politics.

Decisions introduce constraints

Real decisions also happen inside constraints.

A decision-maker cannot simply choose anything.

There may be:

Physical constraints

A power line cannot carry infinite electricity.

A hospital cannot put 50 patients into 10 beds.

An ambulance cannot be in Dublin and Cork simultaneously.

Financial constraints

A government has a budget.

A company has limited capital.

A household has limited income.

Legal constraints

Some actions are prohibited regardless of whether an algorithm thinks they are efficient.

Ethical constraints

Some outcomes may be technically possible but socially unacceptable.

Time constraints

A theoretically perfect decision tomorrow may be useless if action is required in ten seconds.

This means real decision problems often look more like:

[ \max_a U(a) ]

subject to:

[ g_1(a) \leq 0 ]

[ g_2(a) \leq 0 ]

[ \vdots ]

where:

(a) is an action
(U(a)) represents what we are trying to achieve
(g_i(a)) represents constraints

Prediction may help us estimate the quantities inside this problem.

But prediction is not the optimisation problem itself.

Prediction asks "what?"

Decision-making asks "what should we do?"

A useful distinction is:

Prediction	Decision
What is likely to happen?	What should we do?
What category does this belong to?	What action should follow?
How much demand will there be?	How should supply be allocated?
How risky is this patient?	Who receives treatment?
How likely is default?	Who receives credit?
Where will demand occur?	Where should resources be positioned?
What will the price be?	Should we buy, sell or wait?

The prediction can be extremely sophisticated.

The decision rule can still be extremely poor.

Why this distinction matters for AI

A great deal of discussion about artificial intelligence implicitly assumes:

[ \text{better prediction} \Rightarrow \text{better decisions} ]

But that implication does not automatically hold.

A more accurate prediction can improve the information available to a decision-maker.

Whether it improves the eventual outcome depends on:

the objective
the decision rule
the constraints
the incentives
the allocation mechanism
the institutional structure
the people affected

This becomes increasingly important as AI moves out of recommendation systems and into systems governing:

healthcare
energy
finance
transport
education
employment
public services
infrastructure

The difficult question is no longer merely:

Can AI predict this?

It becomes:

How should those predictions interact with the rules governing the system?

AI does not remove politics or values

There is another important consequence.

Suppose a government uses AI to allocate public resources.

It may be tempting to describe the result as:

data-driven

or:

algorithmic

or:

objective

But the system still contains human choices.

Someone chose:

what outcome to predict
what data to use
what errors matter
what objective to optimise
what constraints apply
how scarce resources are allocated
what counts as fair

AI can help us make those choices more intelligently.

It cannot make the underlying value choices disappear.

Sometimes an algorithm merely makes them harder to see.

From machine learning to system design

This is the transition this module will make.

So far, much of this course has concentrated on questions such as:

Can we learn useful patterns from data?

Can we make good predictions?

How do we know whether those predictions are good?

Those questions remain important.

But now we move one level higher.

We begin asking:

How are predictions used inside real systems?

That means thinking about:

decisions
objectives
constraints
allocation
incentives
feedback
fairness
human behaviour

This is where artificial intelligence becomes part of system design.

And it is where many of the most consequential questions about AI actually begin.

The key distinction

Remember this:

[ \boxed{\text{Prediction tells us what might happen.}} ]

[ \boxed{\text{Decision-making determines what we do about it.}} ]

And when resources are scarce:

[ \boxed{\text{Decision-making often becomes allocation.}} ]

A model might tell us that fifty patients could benefit from treatment.

It cannot make ten hospital beds become fifty.

Once scarcity appears, we have to decide how the available resources should be used.

That requires more than prediction.

It requires a system.

Key takeaway

AI models can provide extraordinarily useful information without telling us what the correct action is.

Prediction estimates outcomes.

Decision-making chooses actions.

Allocation determines who receives scarce resources.

Confusing these three things can lead us to believe that an AI system has solved a problem when it has actually solved only one component of it.

In the next lessons, we will start looking at what happens after prediction: objectives, optimisation, allocation, constraints, incentives and the design of the systems in which AI actually operates.