1. AI Can Make Almost Anything. Can We Still Tell What Is Good?

Part 3 of After Intelligence: From Capability to Agency in the Age of AI**

There is an experiment anyone can conduct with generative AI.

Ask it for ten ideas.

Then ask for ten more.

Then another ten.

It will not become tired.

It will not complain that the brainstorming session has gone on too long.

It will simply continue.

Names.

Strategies.

Research questions.

Images.

Business concepts.

Lesson plans.

Product features.

Possible solutions.

Alternative explanations.

More.

And more.

And more.

This abundance can feel exhilarating.

Until you have to choose.

That is the strange problem sitting underneath generative AI.

Generation is becoming extraordinarily cheap.

Selection is not.

In fact, selection may become harder precisely because generation has become easier.

If AI can make almost anything, one of the defining human questions becomes:

Can we still tell what is good?

Fluency is not quality

Generative AI has an extraordinary capacity to create outputs that look finished.

This matters psychologically.

A rough human idea announces its incompleteness.

The sketch is obviously a sketch.

The early draft looks unfinished.

The inexperienced speaker sounds inexperienced.

AI often obscures those signals.

It produces polished prose.

Beautifully formatted analysis.

Confident recommendations.

Convincing images.

Professional-looking slides.

Working code.

The result possesses what we might call surface authority.

It looks like something that deserves to be taken seriously.

But polish and quality are not the same thing.

A beautifully written argument may contain a poor assumption.

An impressive strategy may solve the wrong problem.

A striking image may communicate the wrong idea.

Elegant code may introduce security risks.

A confident research summary may omit the most important evidence.

This is why my earlier After Scarcity series argued that AI creates an educational paradox:

The less expertise required to produce an expert-looking answer, the more expertise may be required to evaluate one.

That paradox becomes even more important as AI creation expands.

We are moving from production to evaluation

Consider how many educational and professional activities have historically relied upon production as evidence of competence.

A good writer produced good writing.

A good designer produced good design.

A competent analyst produced a strong analysis.

A programmer produced working software.

A researcher produced a literature review.

The relationship between output and underlying capability was never perfect.

But there was a useful connection.

Generative AI weakens it.

Someone with limited knowledge can now produce a surprisingly sophisticated first draft.

This does not make expertise obsolete.

It makes the relationship between production and expertise less visible.

Which means we increasingly need to examine another capability:

Evaluation.

Can the person tell whether the output is good?

Can they improve it?

Can they identify what is missing?

Can they reject it?

Can they explain why?

Can they choose between alternatives?

Can they take responsibility for the final decision?

That is discernment.

What is discernment?

Discernment is more demanding than spotting errors.

It involves judgement about quality, relevance, consequence and meaning.

Discernment therefore sits at the intersection of several capacities.

Knowledge.

Experience.

Taste.

Values.

Critical thinking.

Context.

Ethics.

Curiosity.

Even humility.

It requires knowing enough to make a judgement while remaining willing to discover that your judgement is wrong.

Taste is knowledge with a history

We often talk about taste as though it were mysterious.

Someone “has a good eye”.

They “know good writing”.

They can “tell when something works”.

But taste rarely appears from nowhere.

A good editor has read enormous amounts of writing.

A good architect has encountered many buildings.

A good researcher has read strong and weak studies.

A good musician has listened carefully.

A good teacher has watched students misunderstand the same idea in many different ways.

Eventually, this accumulated experience becomes compressed.

The expert may recognise that something is wrong before they can fully articulate why.

We call that intuition.

But intuition is often experience travelling faster than explanation.

This creates an interesting educational problem.

If AI increasingly allows people to skip the difficult work of producing early, mediocre outputs, do they still develop the experience required to recognise excellent ones?

If a student never struggles to construct an argument, will they learn what makes an argument strong?

If someone rarely writes from scratch, will they develop editorial taste?

If a novice programmer delegates most code generation to AI, what experiences build their ability to recognise dangerous code?

This is not an argument for artificially withholding AI.

It is an argument for thinking much more carefully about how expertise develops in an AI-mediated environment.

Productive struggle still matters

Education has long contained an uncomfortable truth.

Some learning is inefficient.

Students attempt things badly.

They get confused.

They make mistakes.

They revise.

They discover why the obvious answer does not work.

From a productivity perspective, this can look wasteful.

AI offers to remove much of that waste.

Why spend three hours struggling with something an AI can explain in thirty seconds?

Sometimes that is exactly right.

There is no intrinsic virtue in unnecessary frustration.

But sometimes the struggle is where the mental model develops.

The challenge for educators is therefore not to preserve difficulty for its own sake.

It is to distinguish between:

friction that obstructs learning

and

friction that creates learning.

That distinction will become increasingly important.

The goal should not be:

“Can the student do this without AI?”

The more useful question may be:

“What does the student need to experience personally in order to develop the judgement required to use AI well?”

That is a much richer educational problem.

Abundance can flatten our standards

There is another danger.

When the volume of generated material becomes overwhelming, our criteria can gradually deteriorate.

We choose what is good enough.

We skim.

We accept the first plausible answer.

We allow fluency to substitute for thought.

We confuse convenience with quality.

AI systems encourage this because they are extremely accommodating.

Ask for an alternative and one appears.

Ask for more detail and it expands.

Ask it to sound authoritative and it will.

There is rarely a moment when the machine says:

“The problem here is not that you need another answer. You need to think harder about the question.”

Yet that may often be precisely what we need.

Discernment therefore includes knowing when not to generate.

When to stop.

When to sit with ambiguity.

When to seek another human perspective.

When to go to the original source.

When to conduct the experiment.

When to speak to the customer.

When to ask the student.

When reality deserves the final word.

Universities may become places where standards are formed

This suggests a different way of thinking about higher education.

Universities have traditionally been places where knowledge is transmitted and created.

Perhaps they are also, importantly, places where standards of judgement are formed.

Students encounter disciplinary traditions that say:

This is what counts as evidence.

This is how we distinguish a strong argument from a weak one.

This is what ethical practice requires.

This is what good design looks like.

This is what a credible experiment must demonstrate.

This is why this source deserves greater weight than that one.

These standards are not arbitrary barriers imposed by disciplines.

At their best, they are accumulated intellectual memory.

AI gives students extraordinary access to outputs.

Universities can help them develop the standards required to interrogate those outputs.

That could become one of higher education's strongest propositions in the AI age.

Assessment should reveal judgement

If this is true, assessment needs to evolve.

Imagine two students submit similarly polished reports.

One accepted almost everything AI suggested.

The other:

• challenged an assumption;
• verified key evidence;
• discarded three generated approaches;
• consulted an original source;
• changed their position;
• identified a stakeholder the model overlooked; and
• deliberately chose a less efficient solution because it produced a better human outcome.

The submitted documents might look similar.

The learning is not.

Perhaps assessment needs to make that difference visible.

Not by policing every interaction with AI.

But by asking students to demonstrate judgement.

For example:

What did you reject, and why?

Where was AI wrong?

What evidence changed your position?

Which choice was hardest?

What criteria did you use?

What would you defend if challenged?

What are you still uncertain about?

These questions move assessment away from merely producing an answer and towards taking intellectual responsibility for one.

The human advantage may be choosing carefully

There is something appealing about the idea that AI will make us all more creative.

I hope it does.

But creativity without discernment can simply increase noise.

The deeper human advantage may lie in the combination:

Generate widely.

Judge carefully.

Choose deliberately.

Take responsibility.

That responsibility matters.

Because the final decision cannot disappear into the model.

“The AI recommended it” is not a satisfactory ethical position.

Nor is:

“The algorithm selected it.”

Somebody still chose to use the system.

Somebody selected the criteria.

Somebody accepted the recommendation.

Somebody deployed the result.

Discernment ends not with judgement but with ownership.

But who gets to develop discernment?

This leads to a final complication.

It is tempting to imagine that abundant AI capability democratises opportunity.

Give everyone the same powerful tools and perhaps everyone becomes more capable.

There is truth in that.

But identical tools do not produce identical outcomes.

One person approaches AI with strong disciplinary knowledge, confidence, networks, economic security and years of experience.

Another approaches the same AI without those advantages.

The technology may be equally available.

Their capacity to judge, initiate and act may not be.

And that means the next important AI divide may not simply be access to technology.

It may be access to agency itself.

Next in the series

Part 4 — Agency Inequality: AI Gives Everyone More Capability. It May Not Give Everyone More Agency.

References

Farmer, I. (2026). After Scarcity: Rethinking the University in the Age of AI. Frontiering.