Will AI Ever Have Common Sense?

"Don't you have common sense?"

Or in the more local slang... "No common sense, is it???!?"

Common sense is easily one of those things that you innately understand, yet hard to explain.

The Cambridge Dictionary definition is

the basic level of practical knowledge and judgment that we all need to help us live in a reasonable and safe way.

It's what helps us pour water into a glass without spilling it or realize that leaving the house without an umbrella on a rainy day is a bad idea.

But can AI ever grasp this "gut instinct"?

By the way, if you're not aware, as powerful as most Large Language Models (LLMs) like ChatGPT, Claude, Gemini etc are today, helping you with everything from content ideation to content generation, research to summary, one thing they do not have (yet) is common sense.

The rise of world models, which some argue are leaps ahead of LLMs, might hold the answer.

And in this article, we're going to deep dive into this whole notion of an AI that's equipped with common sense.

Is Common Sense in AI That Important?

You might be asking, why is this even important?

Shouldn't we reap the already impressive benefits of the current AI models?

Are we building something that would already cause our demise? #skynet

(Although I would argue anyone with common sense would not seek to destroy, but to add value. But that's an entirely different conversation totally)

Think of it this way: AI with common sense could revolutionize how we interact with machines, making them safer, smarter, and more capable.

Imagine customer service bots that understand human frustration or autonomous systems that intuitively avoid risky decisions.

Yet, we're far from a consensus on whether AI can truly replicate the nuanced, context-driven reasoning we call common sense.

With the launch of reasoning models like the o1, some might argue that it's a step towards common sense. Yes, the ability to think deeper and reason logically can be regarded as an advancement, but it's still "mechanical".

Common sense, in its true nature, is way more "intuitive." So perhaps it might serve you and I better if we jump to the next question, which is...

What is Common Sense, Really?

Before diving into the capabilities of AI, it’s important to understand what we mean by "common sense."

At its core, common sense is a blend of intuition, experience, and learned behavior that helps humans navigate everyday situations.

It’s what allows someone to anticipate the consequences of their actions or understand implicit social norms without needing explicit instructions.

Common sense often involves:

• Basic Physical Understanding: Knowing that heavy objects are harder to lift than light ones or that fire burns.
• Social Awareness: Understanding that interrupting someone mid-sentence is considered rude in many cultures.
• Problem Solving: Making quick, informed decisions, like avoiding a pothole while driving or switching lanes to escape a traffic jam.

These behaviours are the result of years of observations, reflection, and learning from mistakes and successes, something our brain is naturally good at. But replicating an artificial common sense is not just a matter of writing better codes.

What makes common sense so challenging for AI is that it relies heavily on context and tacit knowledge—things that humans pick up through lived experience but that are rarely documented in ways machines can process.

Frankly, as we're on this topic, you'd probably recall a time when a loved one, a team member or even a service provider failed to demonstrate "common sense", which is why the wise would say two things:

1. Common sense is not common among all. Your common sense is not the same as the common sense of others.
2. Common sense it not always common practice. Just because we know something, doesn't mean we'll do it.

Ok, enough talk about common sense. Let's talk solution.

What if we, indeed, can imbue an AI that has common sense?

Say Hello to World Models!

World models are a new type of AI system designed to understand and simulate how the world works by observing and learning from interactions in a structured way.

For example, in robotics, world models are used to help machines navigate their environment safely, like autonomous vehicles identifying obstacles and making real-time decisions.

Another example is game AI, where world models simulate complex environments to train characters to act strategically, such as managing resources or predicting player behavior.

These real-world applications highlight their potential to model causality and develop a deeper understanding of dynamic systems.

Unlike LLMs, which focus on predicting the next word in a sentence, world models aim to grasp the relationships between objects, actions, and outcomes in the physical and digital realms.

Key Differences Between World Models and LLMs:

1. Goal Orientation: LLMs are excellent at linguistic tasks but lack a sense of purpose. ChatGPT, Gemini or Claude have no problems generating a grammatically flawless response to an email, but they do not understand the email’s real-world importance or context. On the other hand, world models are designed to solve problems through interaction. A clear example is in autonomous drones that navigate disaster zones, where they can prioritize tasks such as identifying survivors or delivering medical supplies. These interactions showcase how world models not only perform tasks but also align their actions with predefined objectives, fostering a form of goal-oriented learning.
2. Understanding vs. Prediction: LLMs work on patterns and probabilities, making them highly effective at tasks such as generating plausible sentences or analyzing large datasets. However, they lack the ability to infer cause and effect, a critical component of common sense. For example, an LLM might predict the next word in a sentence about a vehicle crash but would not understand that slippery roads due to rain caused the crash. World models, on the other hand, aim to grasp causality, such as recognizing that if a road is icy, vehicles should reduce speed to avoid accidents. This ability to infer and act on cause-and-effect relationships is what positions world models closer to replicating human common sense.
3. Application Domains: LLMs excel in text-heavy applications, such as summarization, coding, and crafting detailed reports. You've probably experience this when a chatbot powered by an LLM can guide you through troubleshooting steps for a device. However, when it comes to domains that demand spatial awareness and real-time decision-making, world models take the lead. In robotics, for example, world models enable robots to navigate complex manufacturing environments, avoiding obstacles while optimizing assembly tasks. Similarly, in logistics, world models can simulate delivery routes that account for real-time traffic updates, weather conditions, and package fragility, ensuring both efficiency and safety. Game AI is another area where world models excel by training characters to anticipate player moves and adapt their strategies dynamically.

Now that you know the difference between the two models, let's see how world models can help in common sense?

Can World Models Bridge the Common-Sense Gap?

Understanding whether world models can truly bridge the gap toward AI with common sense requires examining both their strengths and limitations.

The Pros

World models exhibit remarkable advantages in their ability to simulate cause-and-effect relationships and adapt across various domains:

1. Learning by Doing: Through real-time interactions, world models simulate and learn from their environment, embodying a foundational aspect of common sense. In agriculture, for example, they predict the impact of irrigation patterns on crop yields, enabling farmers to fine-tune water usage for optimal productivity. Similarly, autonomous vehicles leverage these models to understand road conditions and driver behavior, helping to reduce accidents and improve traffic flow.
2. Cross-Domain Understanding: Unlike LLMs, which are often task-specific, world models generalize knowledge across disciplines. In healthcare, they combine patient history and real-time monitoring to provide comprehensive diagnoses that consider both symptoms and context. In logistics, these models integrate data such as weather, traffic, and package details, creating streamlined and adaptive delivery networks that respond to dynamic variables.

The Cons

Despite their potential, world models face significant challenges that limit their ability to fully replicate human-like common sense:

1. Contextual Nuances: Abstract concepts, cultural norms, and ethical dilemmas remain areas where world models struggle. To address these gaps, researchers are embedding ethical frameworks into training processes and exploring multimodal learning methods. For example, initiatives are underway to help AI systems recognize and adapt to cultural differences in communication styles, enhancing their relevance in diverse settings.
2. Data-Intensive Requirements: Developing robust world models demands vast amounts of high-quality data and computational resources. Synthetic data generation is one approach being explored to address data scarcity. Additionally, advancements in federated learning—where decentralized systems learn collaboratively without compromising privacy—offer a scalable and secure pathway forward. Leading organizations like OpenAI and Google are actively investing in these techniques to mitigate the data burden while ensuring effectiveness.

By addressing these limitations through ongoing innovation, world models may continue advancing toward an AI ecosystem that approaches the elusive quality of common sense.

Why it might look like a daunting task, don't forget how much progress we have made in the field of AI over the last 2 years. And with the recent discoveries and further research in the right areas, it's a matter of time before we overcome these hurdles and enjoy the harvest, as how we always have done throughout history.

Actionable Insights for Organizations and Businesses

Despite the challenges, the progress in world models is undeniable. These systems already provide tangible benefits, and forward-thinking organizations are finding creative ways to harness their strengths. For businesses, the question isn't just about waiting for AI to perfect common sense—it’s about how to integrate today’s advancements to stay competitive and innovative.

Here’s how companies can start leveraging the power of both world models and LLMs effectively.

1️⃣ Combine Strengths: Use LLMs for tasks like content generation or customer interactions and world models for dynamic decision-making in logistics, robotics, or operational efficiencies. For example, a customer service team might use an LLM to handle text-based inquiries, while a logistics team uses world models to optimize delivery routes in real-time.

💡Think about your own organization: could a hybrid approach improve your processes?

2️⃣ Redefine Training Data: Feed systems with diverse, high-quality data that mimics real-world complexities. Specific tools like AWS SageMaker or Google’s AI Platform can help streamline this process by managing large-scale datasets. For example, training an AI assistant on workplace-specific scenarios—such as onboarding new employees—can enhance its contextual understanding and effectiveness.

💡Reflect on your current workflows—are there areas where smarter AI could enhance outcomes, like onboarding new employees or optimizing scheduling?

3️⃣ Start Small, Scale Fast: Pilot AI solutions in controlled environments (e.g., warehouse simulations, digital twins) to test and refine their practical "common sense." Tools like AnyLogic or Simio can help create these simulations, allowing businesses to evaluate performance before large-scale implementation.

💡 Where in your operations could a pilot project reveal new efficiencies or opportunities?

4️⃣ Stay Ahead of Ethical Concerns: Implement checks and balances to ensure AI decisions align with human values, especially when dealing with safety-critical applications. Frameworks such as IBM’s AI Ethics Toolkit or Microsoft’s Responsible AI Standard provide guidance on embedding ethical considerations into AI systems.

💡 How could your organization ensure transparency and fairness in AI-driven processes?

But The Real Question is... Are You Ready For A World Where AI Has Common Sense?

World models bring us closer to an AI that can mimic human intuition, but there's a long way to go. You and I have explored the vast potential such an AI can do when it comes to Margin Maximization (Cost Reduction and Revenue Generation) and Market Domination (Positioning Solidification and Market Expansion).

For now, before that day comes around, businesses should focus on practical hybrid strategies that leverage the strengths of both LLMs and world models.

But as we close our exploration into this topic today, and as we edge closer to this breakthrough, perhaps you and I need to pause, take a few steps back and ask ourselves the true question:

If AI gains common sense, will we trust it to use it wisely?