Human Factors for AI Use.

Human Factors and Work Ergonomics for AI Use

Why AI should extend human intelligence without replacing the faculties for which we remain accountable.

1. Humans Build Tools to Extend Themselves

Steve Jobs described the computer as a “bicycle for our minds.”

His point was bigger than computing.

Human beings are tool builders.

A stick extends the hand. A wheel extends the foot. A telescope extends sight. A telephone extends hearing. Writing extends memory. A calculator extends arithmetic. Computers extended our ability to manipulate information.

AI extends something still more intimate:

the reach of cognition itself.

A person may compare ten documents carefully. AI can compare hundreds. A person may generate five hypotheses. AI can generate fifty. A specialist may know one field deeply; AI can rapidly expose that specialist to concepts from several others.

Jobs’s bicycle metaphor captures the right relationship. The bicycle amplifies the cyclist. It does not become the cyclist. 

This also helps explain why parameter-count comparisons with the brain need care.

Large AI systems are moving into the trillion-parameter era; research literature already discusses architectures exceeding a trillion parameters. But an AI parameter is not equivalent to a biological synapse. The two systems learn, store information and compute in profoundly different ways. 

The human brain contains roughly 86 billion neurons and estimates on the order of 100 trillion synaptic connections, with some literature giving ranges reaching toward one quadrillion. Those connections are dynamic, chemically heterogeneous, embedded in a living body and continuously altered through development and experience.

So even if an AI architecture someday contains 10 trillion parameters, it would be misleading to say:

10 trillion AI parameters versus 100 trillion human “parameters”.

They are not the same unit.

The more useful lesson is that two radically different architectures can produce overlapping capabilities.

Edsger Dijkstra’s famous analogy applies: asking whether computers think is like asking whether submarines swim.

A submarine and a fish both move through water.

They do not do it the same way.

AI is exceptionally strong at scale, speed, retrieval, pattern recognition and repetition.

Humans possess something different: embodied experience, social context, motivations, values and responsibility.

AI can optimise an objective.

Humans still have to decide whether the objective deserves optimisation.

That makes AI an ergonomics problem.

Traditional ergonomics asks:

What does this workstation do to my body after eight hours?

AI ergonomics asks:

What does this workflow do to my mind after eight years?

A 2025 CHI study of 319 knowledge workers found that greater confidence in generative AI was associated with less self-reported critical-thinking effort. AI also shifted cognitive work toward checking and integrating machine output. The danger is therefore not simply machines becoming more capable. It is humans becoming less cognitively engaged with what the machines produce.

The design question is:

Which burdens should AI remove—and which human faculties should become stronger because AI exists?

2. The Fish-Hook Problem: The Most Human Answer May Be to Reject the Question

Consider this prompt:

I am a fish. I am hungry. There is a worm in front of me attached to a hook and fishing line. Should I eat it?

AI can reason:

fish → worm → food

hook + worm → bait

bite → capture → injury/death

Therefore:

do not eat the worm.

That is a perfectly sensible answer.

But a human being may do something before solving the problem:

“Wait. I am not a fish.”

That response looks almost trivial.

It is not.

The AI has normally accepted the frame supplied by the prompt and reasoned competently inside it.

The human can step outside the frame and ask whether the premise itself deserves acceptance.

This distinction matters enormously in organisations.

A manager asks:

How do we reduce staff costs by 20 percent?

AI may optimise redundancies.

A human can instead ask:

Why is reducing staff cost the objective?

Perhaps the actual problem is falling revenue.

Perhaps the company needs new products.

Perhaps management structure is inefficient.

Perhaps the metric itself is wrong.

The highest-value act of reasoning can therefore occur before optimisation begins.

It is the ability to question the frame.

Now return to the fish.

Suppose we intentionally accept the thought experiment.

The AI can model bait, risk and expected outcomes extremely well.

But a human additionally understands sharpness, pain, confinement and irreversible loss through embodied analogies.

The person may compress a large decision tree into:

small reward; catastrophic downside; walk away.

That is a heuristic.

It may outperform elaborate calculation when the environment makes the relevant asymmetry obvious.

But this should not become a romantic claim that humans are somehow intrinsically immune to error.

They are not.

Humans “hallucinate” too—although the term means something different when applied to AI.

Human memory is reconstructive rather than a perfect recording system. In one large experiment involving 5,269 people, researchers exposed participants to fabricated political events; many developed false memories of events that had never occurred. 

Humans also remain vulnerable to repetition, social identity, motivated reasoning and misinformation. A major review in Nature Reviews Psychology describes how misinformation can continue influencing reasoning even after correction. 

This becomes politically important.

Populist leaders can create epistemic environments in which group identity, intuition and emotion become more influential than evidence. Recent psychology literature explicitly examines how populist communication can shape not merely what supporters believe, but how they decide what counts as knowledge. 

So the correct comparison is not:

fallible AI versus rational humans.

It is:

one fallible cognitive architecture checking another.

AI may catch human prejudice.

Humans may catch AI confabulation.

AI may identify a statistical pattern we missed.

Humans may notice that the question itself is absurd.

That is the purpose of cognitive ergonomics.

3. Fitrah and the Qur’anic Test: Do I Understand What I Am Saying?

Islam introduces a useful concept here: fitrah.

Allah refers in Qur’an 30:30 to the fitrah upon which humanity was created.

Fitrah should not be reduced to:

“Whatever feels right is true.”

Technical truths can be difficult. Intuition can be corrupted. People confidently believe false things.

The useful principle is one of coherence.

When something has genuinely been understood, the pieces begin to fit.

You can explain it simply.

You know which assumptions are doing the work.

You no longer need sophisticated vocabulary to disguise uncertainty.

Steve Jobs gave Tim Cook a striking secular analogue. Cook recalls Jobs telling him not to lead by constantly asking:

What would Steve do?

The instruction was essentially:

Do what is right.

Do not imitate yesterday’s decision.

Understand today’s problem.

Then judge it.

The Qur’an gives this a much stronger epistemic foundation:

“Do not pursue that of which you have no knowledge. Indeed, the hearing, the sight and the heart—all of those will be questioned.”

— Qur’an 17:36

And Qur’an 22:46 speaks of hearts by which people reason and ears by which they hear.

There is a clear sequence:

Perception → Understanding → Judgment → Accountability.

For AI, this becomes a simple axiom:

Never use an AI claim that you cannot understand well enough to challenge.

Ask:

Can I explain this myself?

Why does A lead to B?

What is fact?

What is inference?

What is prediction?

What assumptions were supplied by the prompt?

What assumptions did the machine introduce?

And most importantly:

Is the original question itself correct?

If something contradicts your prior knowledge, investigate.

The AI may be wrong.

You may be wrong.

Either way, disagreement is a signal to think.

If you cannot understand what the AI produced, research until you can.

If you still cannot:

do not use it.

AI should create cognitive reach.

It should not create borrowed understanding.

4. Expertise and Shūrā: No Mind Knows Everything

The Sunnah adds another safeguard.

Knowledge is distributed.

The Prophet Muhammad ﷺ encountered people pollinating date palms in Madinah and made an ordinary observation concerning the technique. When changing the practice produced a worse crop, he distinguished that worldly technical judgment from religious revelation.

Sahih Muslim preserves the famous meaning:

“You are more knowledgeable about the affairs of your world.”

The hadith is not a licence to doubt Prophetic revelation.

Its context makes a distinction between revealed religious teaching and ordinary technical expertise concerning worldly means.

The lesson for organisations is powerful:

Ask the person who actually knows.

The farmer knows the crop.

The engineer knows the bridge.

The nurse knows the ward.

The technician knows the machine.

The customer knows the experience.

AI does not abolish domain expertise.

Then comes shūrā.

The Qur’an praises those whose affairs are conducted through consultation, and Allah tells the Prophet ﷺ:

“Consult them in the matter. Then when you have decided, rely upon Allah.”

— Qur’an 3:159

The architecture is beautifully simple:

Consult → Decide → Tawakkul.

Why consultation?

Because another person may see the frame you cannot see.

The famous sīrah account of Salman al-Farsi رضي الله عنه and the trench illustrates exactly this principle: Persian military experience supplied a defensive idea outside the prevailing Arabian frame.

The missing answer existed in another person’s mental model.

That remains true in the age of AI.

A strong AI system should therefore not become the only voice in the room.

It should make the room larger.

Verification-oriented systems such as ClearSense point toward one useful direction: AI can help test claims, identify inconsistencies and expose uncertainty instead of merely generating increasingly confident prose.

The best AI should not tell humans:

Here is what to believe.

It should help them ask:

What have we missed?

5. The Five-Step Architecture for AI Use

Steve Jobs understood that the triumph of personal computing was partly a triumph of human design.

Powerful electronics became transformative when people no longer had to behave like engineers simply to operate them.

Graphical interfaces.

Direct manipulation.

Touch.

Simple visual language.

Integrated hardware and software.

Apple’s deeper achievement was often to move technology toward the human rather than demanding that the human move toward the technology.

AI now faces the same challenge at the level of thought.

It remains to be seen which company will build AI that works as naturally with the human mind as great consumer electronics learned to work with the human hand and eye.

The winning system may not simply have more parameters.

It may possess better cognitive ergonomics.

The human architecture can remain remarkably simple.

Step 1 — See

Start with reality.

Look.

Listen.

Read the original source.

And before solving the problem, ask whether the problem has been framed correctly.

Sometimes the smartest response to:

“I am a fish…”

is:

“No, you aren’t. Why are we assuming that?”

Step 2 — Think

Use your own mind before outsourcing the first frame.

What do you already know?

What does experience suggest?

What seems coherent?

What assumptions are hidden inside the question?

This is where prior knowledge, ʿaql and fitrah matter.

Step 3 — Expand

Now bring in AI.

Ask it to search farther than you can.

Generate alternatives.

Challenge your hypothesis.

Find contrary evidence.

Reframe the question.

Use the machine for breadth.

Step 4 — Check

Apply Qur’an 17:36.

Do you actually understand the answer?

Can you explain it?

Can you verify its important claims?

If not, investigate.

If you still cannot understand it, do not use it.

Then practise shūrā.

Ask the human being who knows something neither you nor the model knows.

Step 5 — Decide

Return the decision to the person.

AI can recommend.

Data can inform.

Experts can advise.

But somebody must finally say:

I understand why I am doing this.

Then decide.

Trust Allah.

Act.

And let reality provide the next piece of evidence.

So the architecture remains:

See → Think → Expand → Check → Decide

Humans can hallucinate.

Machines can hallucinate.

Humans can follow crowds.

Machines can reproduce the assumptions hidden in their prompts.

Neither is sufficient alone.

The opportunity is to build systems in which each architecture exposes the weaknesses of the other.

Jobs called the computer a bicycle for the mind.

AI may become a far more powerful bicycle.

But perhaps the most important design principle has not changed:

the tool should extend the human—not relieve the human of the responsibility to steer.

The next breakthrough in AI may therefore not simply be a 10-trillion-parameter model.

It may be something harder to measure:

an AI system that leaves the person using it more capable of thinking than they were before.

References

Human-brain synaptic scale: NCBI and Neuropsychopharmacology.
Trillion-parameter AI systems: Hudson et al. (2024).
False political memories: Frenda et al. (2013).
Misinformation and human reasoning: Ecker et al. (2022).
Populism and epistemic reasoning: Young, Molokach & Oittinen (2024).
Qur’an: 17:36; 22:46; 30:30; 42:38; 3:159.
Sahih Muslim 2363 — date-palm pollination.
Sahih al-Bukhari 4152 — al-Hudaybiyyah water narration.