From Patterns to Causes

Dr Ahmed T. Hammad

Decision Making for AI — Week 5 of 13



“What can you do that AI cannot?”

Where We Are

Last week — More data ≠ more knowledge.

More data -> more noise

Noise -> less reliable conclusions

= worse decision making


Today — Give AI every advantage:

Perfect data. Clean data. No noise. No missing values.

Even then — what can AI still not do?

AI is extraordinarily good at one thing

Finding patterns in data.

Give it a million data points — it finds relationships no human could spot manually.

But let’s look at what that actually means.

Case 1 — Ice Cream and Drowning Deaths

Case 1 — What the Machine Missed

Should we ban ice cream to prevent drownings?

The pattern is real. The relationship is false.

A third variable — summer heat — drives both.

Term Meaning
Correlation Two things moving together
Causation One thing producing the other
Confounder A hidden variable driving both

Case 2 — Internet Access and Life Expectancy

Case 2 — The Hidden Driver

Countries with more internet access have higher life expectancy.

Wealth buys both — better healthcare and internet infrastructure.

Internet is not keeping you alive…

Case 3 — Wolf or Husky?

An image classifier trained to distinguish wolves from huskies — predictions shown below each image

Case 3 — What the Algorithm Actually Learned

The highlighted patches show what drove each prediction. The model learned: snow in background = wolf. Not the animal.

The confounder was the background, not the subject. The model found that pattern and ran with it.

Source: Ribeiro, Singh & Guestrin. “Why Should I Trust You?” KDD 2016. (LIME paper)

Be Careful

Statistical Learning ≠ Causal Learning

Real Stakes — A Healthcare Case

A hospital uses an AI system to decide which patients should receive an aggressive treatment.

The AI is trained on historical records:

  • whether the patient received the treatment
  • whether the patient survived
  • gender, age, bmi, BP

What is missing from the list?

Real Stakes — What the Data Looks Like

What the Algorithm Could Not See

The sickest patients were the ones who received the treatment.

Of course their outcomes were worse — they were already critical.

  • The AI saw the pattern ✓
  • It could not see the selection process that created the pattern ✗

The machine could not ask: why were these patients selected for treatment?

An intervention is justified when the potential benefits in severe outcomes outweigh the uncertain risks of side effects.

Are we really that good?

Causation is hard for humans too.

We confuse correlation and causation every day — in medicine, in policy, in personal life.

So what is actually different between us and the machine?

We know it is hard. That’s why we have:

  • Randomised controlled trial
  • Scientific consensus
  • Legal standard of proof

the key difference is that we design methods to go beyond our limits.

Why.


That single question is what changes everything.


Correlation tells you what the world looks like.

Causation tells you what you can change about it.

“What can you do that AI cannot?”

You can ask why.

You can suspect a pattern and say: something feels off — let me find the mechanism.

You can sit with uncertainty.

“The most dangerous thing about AI is not that it is wrong. It is that it is confidently wrong in ways it cannot detect. Your job, going forward, is to be the one who can.”

You can doubt — and build on top of your doubts.

For Next Week

If AI cannot reliably distinguish correlation from causation, what are the implications for its use in your respective fields of study?




Thank you

Dr Ahmed T. Hammad

ahmed.t.hammad@gmail.com

WA: +393349715555

Decision Making for AI — Acacia College, NUS