Can AI Develop Common Sense?

There are many definitions of intelligence. Common sense is harder to pin down. We recognise it when we see it — and perhaps even more quickly when we don’t.

Imagine you are about to leave the house. The sky has turned almost black. The wind is picking up. People walking toward you carry wet umbrellas.

It isn’t raining where you are standing. Nobody has told you it is going to rain. You haven’t checked the weather forecast. You take an umbrella.

That seems like common sense. But think about what just happened.

You observed several unrelated things. You connected them with previous experience. You formed an idea of what was probably happening nearby. You predicted what was likely to happen next. And you changed your behaviour accordingly.

Observe. Connect. Infer. Predict. Act.

Suddenly common sense doesn’t seem quite so simple.

More Than Intelligence

I’ve often been struck by situations where something seems perfectly obvious to me but not to someone else.

Everyone can have the same information. Everyone can understand the problem. Yet sometimes people keep analysing and complicating something when the practical answer seems remarkably simple. We’ve probably all known highly intelligent people who appeared to have surprisingly little common sense. And we’ve known people without great academic qualifications who possessed enormous practical judgement.

Perhaps intelligence and common sense aren’t the same thing. Intelligence can help us solve a problem.

Common sense may help us recognise what the problem really is.

The Invisible Invoice

I encountered an interesting example many years ago, when business computer systems were still relatively immature. Intercompany transactions traditionally generated invoices. The paper mattered. They provided a record—part of the audit trail accountants depended on. But computers were changing that.

If one company within a group created a transaction with another company in the same system, the computer could record both sides automatically. The transaction itself created the trail. So why produce the paper invoice? My suggestion was effectively an invisible invoice.

The accountants were horrified. Their reaction was understandable. Paper trails were fundamental to accounting controls, and computer systems were still new enough that abandoning them felt risky. Eventually, however, they tried the idea. The important point wasn’t really the invoice. It was recognising that technology had removed much of the reason for a familiar process. That suggests another way of thinking about common sense.

Common sense isn’t simply knowing the rules. Sometimes it is understanding why the rules exist — and recognising when the reason for them has disappeared.

Now Consider AI

In the previous article, How Does AI Learn?, we looked at a surprisingly simple process.

• AI predicts.
• It compares.
• It adjusts.
• And it repeats the process on an extraordinary scale.

Along the way, neural networks develop internal representations — relationships between words, objects, ideas and increasingly images, sounds, video and aspects of the physical world. That’s where our umbrella example becomes interesting. An AI capable of seeing the dark clouds, recognising wet umbrellas, understanding the significance of the wind and predicting rain could arrive at exactly the same conclusion:

Take an umbrella.

Would we call that common sense?

Perhaps.

But the invisible invoice is harder.

Knowing the Pattern — Or Seeing Beyond It?

Artificial intelligence is extraordinarily good at learning patterns.

Give it enough examples and it can discover relationships that humans may never notice.

But common sense sometimes requires something almost opposite.

Instead of learning the existing pattern, we have to ask:

Does this pattern still make sense?

The accountants knew the established process.

• Invoice created.
  Paper generated.
  Transaction recorded.
  Audit trail preserved.

The invisible invoice required questioning one of those steps. If the computer could reliably create the record, perhaps the piece of paper was no longer essential.

Could an AI reach the same conclusion?

Quite possibly.

Modern AI is already moving beyond text. Systems can combine language with images, video, and other forms of information, while researchers are developing what are sometimes called world models—systems intended to represent how environments behave and predict what may happen next.

That sounds surprisingly close to part of what we do when exercising common sense.

But there is still a difference worth considering.

We Live in the World

Human common sense comes from more than information.

• We fall over.
• We get wet.
• We burn our fingers.
• We misjudge people.
• We make decisions that seemed sensible and discover that they weren’t.

Over a lifetime, thousands of experiences accumulate into something we rarely consciously analyse.

We simply say:

That doesn’t make sense.

AI does not experience the world as we do.

But does it need to?

If an intelligent system can perceive enough of the world, build increasingly sophisticated representations of it, predict consequences, learn from errors and adjust its behaviour, perhaps something resembling common sense could emerge without human experience at all.

That leads to an uncomfortable possibility.
We tend to think common sense is uniquely human because it comes so naturally to us.

Perhaps it isn’t.

Perhaps common sense is partly what happens when enough knowledge about how the world works becomes available quickly enough to guide action.

And There Is Another Possibility

AI may eventually have an advantage humans don’t. We become attached to familiar ways of doing things.

“This is how we’ve always done it” is one of the most persistent forces in organisations. Experience gives us common sense. But experience also gives us habits.

An AI may learn our rules without becoming emotionally attached to them. Could that sometimes make it better at noticing when a rule has outlived its purpose?

The invisible invoice seemed wrong because the old process was familiar and trusted.

The technology had changed. The thinking hadn’t caught up.
As AI becomes increasingly capable of reasoning across different kinds of information, the question may not be whether machines can acquire our common sense.

It may be whether they occasionally reveal the limits of ours.

So Can AI Develop Common Sense?

We don’t yet have a simple answer.

Common sense includes prediction, perception, experience, context, judgement and an understanding of how things normally work. AI is becoming better at several of those things. But common sense also involves recognising exceptions, questioning assumptions and knowing when the normal rule no longer applies.

That is a much higher bar.

The umbrella is relatively easy. The invisible invoice is harder.

And somewhere between the two lies one of the most interesting questions in artificial intelligence:

If a machine can learn how the world works, can it also learn when the way we think it works no longer makes sense?

Perhaps that is where common sense really begins.

---

---