Why AI Feels Brilliant Until It Misses the Obvious

AI can solve advanced problems and still miss what a child would notice. That gap may explain many of today’s strangest AI failures.

There is a strange moment that happens after you use AI enough.

At first, it feels magical.

It can explain difficult concepts, write code, summarize documents, generate business ideas, draft strategy, debug problems, and reason through topics that would take a human hours to unpack.

Then suddenly, it misses something obvious.

Not complicated-obvious. Not expert-obvious. Human-obvious.

It may fail to count the number of letters in a word. It may overthink a question like, “Should I go to the car wash by car or by driving?” It may answer with impressive structure while missing the basic situation. It may give you a technically sophisticated response that still feels detached from what is actually happening.

That is the part that feels important.

The issue is not simply that AI makes mistakes. Humans make mistakes too.

The issue is the pattern of the mistakes.

Many AI failures do not feel like failures of intelligence. They feel like failures of sense.

AI is becoming extremely intelligent before it is becoming properly sensible.

And that distinction matters.

Intelligence and common sense are not the same thing

We often use “intelligence” as if it means one thing.

But the more I use AI, the more I think intelligence has layers.

One layer is pattern intelligence: the ability to recognize structure, predict language, connect ideas, and produce a coherent answer.

Another layer is analytical intelligence: the ability to reason through a problem, compare options, write code, solve math, or optimize for a goal.

But there is another layer that humans use constantly and barely notice:

Common sense.

Common sense is not just “knowing facts.” It is the ability to understand the situation you are in.

It is knowing what is plausible. It is knowing what is physically possible. It is knowing what someone probably means, even if they phrase it imperfectly. It is knowing when a question is malformed. It is knowing when the literal answer is less useful than the obvious one. It is knowing that context matters before computation begins.

A human hears, “Should I go to the car wash by car or by driving?” and instantly understands that the question is awkward. The sensible answer is: you go by car; driving is how you use the car.

A machine may process the sentence as a linguistic puzzle.

That difference is the gap.

Common sense is compressed contact with reality.

The silly examples are not actually silly

The “strawberry” example became popular because it is funny. Ask some models how many “r” letters are in “strawberry,” and they may get it wrong.

At first, this looks ridiculous. How can a system explain physics, law, medicine, or software architecture and then fail at counting letters?

But that example is more interesting than it looks.

Research on the “strawberry problem” connects this failure to tokenization: many language models process text in chunks or subword units, not always as individual visible characters. This can create a structural blind spot for character-level tasks that humans find trivial. [1]

That is not exactly the same as common sense, but it belongs to the same family of failure.

The human assumes the model sees the thing the way we see it.

It often does not.

The model may be fluent about the word without having reliable access to the word’s character-level structure. It can operate at the meaning layer while being weak at the substrate layer.

That is a powerful metaphor for AI more broadly.

AI can be fluent about a situation without being grounded in the situation.

The car wash example reveals a different version of the same pattern. The problem is not letters. It is pragmatic context. A human does not treat the phrase “by car or by driving” as two meaningful transportation options. We repair the question automatically. We infer intent. We map language to the world.

AI can do this sometimes. Increasingly often, in fact.

But it does not do it reliably.

That unreliability is what makes the issue important.

The research has been pointing at this for years

This is not a brand-new concern.

Commonsense reasoning has been a central challenge in AI for decades. Yejin Choi describes commonsense intelligence as a long-standing puzzle in AI and argues that, despite deep learning progress, AI systems remain narrow and brittle without the ability to reason intuitively about everyday situations and events. [2]

Benchmarks such as CommonsenseQA and HellaSwag were built specifically to test whether systems can use background knowledge and commonsense inference rather than merely pattern-match surface text. CommonsenseQA’s authors designed questions that require prior commonsense knowledge, and early strong baselines remained far below human performance. [3] HellaSwag showed that examples trivial for humans could still be difficult for state-of-the-art models, especially when wrong answers were adversarially constructed to be tempting to machines. [4]

The AI2 Reasoning Challenge took another angle: grade-school science questions that require knowledge and reasoning beyond shallow retrieval or word co-occurrence. [5]

More recent AI progress has not made the issue disappear. It has made it more interesting.

Stanford’s AI Index describes a jagged frontier: AI systems can perform extremely well on some difficult benchmarks while still struggling with tasks that feel ordinary or intuitive to humans. Its 2026 report gives the striking example that models can achieve impressive performance on advanced math while still failing at more everyday perception and reasoning tasks, such as reliably reading analog clocks. [6]

This is the shape of the problem:

AI is not uniformly smart or uniformly dumb.

It is uneven.

It can be superhuman in one direction and strangely fragile in another.

Why common sense is hard to build

Common sense feels easy because humans do it unconsciously.

That is exactly why it is hard.

We do not usually write down the obvious. We do not constantly state that cars are driven on roads, that water makes things wet, that people usually go to a car wash with a vehicle, that a glass can break if dropped, or that someone asking a poorly worded question may need intent-repair rather than a literal answer.

Human knowledge contains a huge amount of unstated background structure.

Much of it comes from living in the world.

We learn through bodies, objects, other people, constraints, accidents, habits, and repeated interaction with reality. We learn what usually happens next. We learn what cannot happen. We learn what would be strange. We learn when to pause.

Yann LeCun has argued that common sense can be viewed as a collection of world models that help an agent know what is likely, plausible, or impossible, and that much human common-sense knowledge comes from interaction with the physical world rather than from text alone. [7]

That point is crucial.

Modern language models are trained heavily on text. Text contains a lot of knowledge, but text is not reality. Text is a record of what humans found worth saying.

And humans rarely say the obvious.

That creates a gap.

The world is full of constraints. Text is full of descriptions. Common sense lives in the constraints.

We may be patching symptoms instead of the root

A lot of current AI work feels like symptom patching.

The model hallucinates, so we add retrieval. The model makes arithmetic mistakes, so we add tools. The model misses instructions, so we add stronger prompts. The model fails a benchmark, so we train against the benchmark. The model gives unsafe answers, so we add filters. The model behaves inconsistently, so we add evaluation layers.

These patches matter. They improve systems. They are not useless.

But they may not be the root.

The root may be that the system does not reliably maintain a grounded model of the situation.

It predicts. It generates. It optimizes. It follows patterns.

But it does not always ask the most basic human question:

“What is actually happening here?”

That question is common sense.

And when it is missing, the system can become dangerously convincing.

A weak answer that sounds weak is easy to reject. A weak answer wrapped in intelligence is harder to notice.

That is why the issue matters beyond funny examples.

In real work, the mistake is not always “how many r’s are in strawberry?”

Sometimes it is:

A code change that compiles but violates the system architecture. A legal summary that sounds coherent but misses the operational risk. A business strategy that is logically organized but ignores the human incentives. A data analysis that calculates correctly but answers the wrong question. An AI agent that follows the instruction but misses the consequence.

These are not just errors of information.

They are errors of situation-sense.

The jagged frontier is the practical version of this problem

In work settings, this shows up as the “jagged frontier.”

Harvard Business School researchers studying AI use by consultants found that AI improved performance on some tasks but worsened it on others, even inside the same knowledge workflow and with tasks that seemed similarly difficult. For tasks within the AI capability frontier, consultants completed more work faster and at higher quality; for a task outside the frontier, AI users were less likely to produce correct solutions. [8]

That is exactly what many users feel intuitively.

AI is not simply good or bad.

It depends on whether the task is inside or outside the invisible boundary of what the model can actually handle.

The problem is that this boundary is not obvious.

A task may look easy and fail. A task may look hard and succeed. A task may work yesterday and fail today with slightly different wording. A model may sound equally confident in both cases.

Common sense would help identify the boundary.

It would help the system notice when it is leaving the zone of reliable competence.

What a common-sense layer would actually do

When people say “AI needs common sense,” the phrase can sound vague.

So what would it mean in practice?

A common-sense layer would not just store more facts. It would do several things before answering.

It would frame the situation.

Before solving the problem, it would ask: what kind of situation is this? Is this a physical-world question, a language question, a social-intent question, a planning question, a trick question, or a malformed question?

It would infer the obvious defaults.

If someone says “car wash,” the system should anchor to vehicles, roads, driving, cleaning, payment, location, and practical intent.

It would detect absurdity.

If two options are not actually distinct, it should say so instead of forcing a comparison.

It would reason about consequences.

Not just “what answer fits the prompt?” but “what happens if this answer is acted on?”

It would know when to stop.

Sometimes the smartest answer is not a longer answer. Sometimes the sensible answer is: “These are the same thing,” or “That question needs clarification,” or “The obvious interpretation is this.”

That is what humans do constantly.

We do not just answer language.

We repair it.

The next leap may not be more intelligence

The AI industry is heavily optimized around capability: bigger models, better benchmarks, longer context, faster inference, stronger agents, more tools.

All of that matters.

But maybe the next leap is not just more intelligence.

Maybe it is better grounding.

Better world models. Better uncertainty detection. Better causal reasoning. Better task framing. Better ability to notice when something is off. Better connection between language and reality.

Because as AI moves from chatbots to agents, the cost of missing common sense increases.

A chatbot with poor common sense gives a weird answer.

An agent with poor common sense can take the wrong action.

That is a different risk category.

The future of AI should not only ask:

How smart is the model?

It should also ask:

How situated is it? How grounded is it? How well does it understand the kind of problem it is in? Can it detect when the prompt is misleading? Can it tell when its answer makes sense in the world, not just in language?

Intelligence needs traction

The more I use AI, the more I believe the missing layer is not raw intelligence.

It is common sense.

Not common sense as a cliché. Not common sense as “basic facts.” Common sense as the grounding layer between intelligence and reality.

The layer that tells a system what is plausible. The layer that repairs messy human language. The layer that notices absurdity. The layer that asks what is actually happening before producing an answer.

AI is already smart.

But intelligence without common sense can become brittle, overconfident, and strangely detached.

If intelligence is the engine, common sense is traction.

And right now, we are building very powerful engines.

We should be paying much more attention to the traction.

— Muhamad J. Akoum LinkedIn: https://www.linkedin.com/in/akoum/