How Common is Common Sense in AI, and Why Should We Care?

Artificial Intelligence (AI) is the new buzzword in town. It has been increasingly trending both in the US and worldwide. Seemingly everyone, from journalists to artists, is fascinated by AI and its abilities.

Even JPMorgan CEO Jamie Dimon recently acknowledged the benefits of AI, calling the technology "extraordinary and groundbreaking.” JPMorgan is now working on more than 300 AI use cases, and Dimon revealed that the bank has spent over $2 billion building cloud-based data centers and is working to modernize a significant portion of its applications to run in both their public and private cloud environments.

AI has sparked an arms race among the big tech companies including Meta, Google, Apple and Microsoft, the last of which has taken a stake in OpenAI, the creator of ChatGPT, with a multi-year, multi-billion investment plan. They have all been aggressively positioning themselves in this space.

Apple CEO Tim Cook recently said, "[AI] is a major focus of ours. It's incredible in terms of how it can enrich customers' lives...Apple sees an enormous potential in this space to affect virtually everything we do," adding that "it will affect every product in every service that we have." Cook’s last note explains why AI has been the main theme repeatedly mentioned in all the big tech earnings calls.

It seems as if AI is the cure for everything, and soon we can let the machines do all the thinking for us.

Well, not so fast.

For certain kinds of tasks – playing chess, detecting tumors, sifting through mountains of data, computing sophisticated algorithms in seconds – artificial intelligence can rival or surpass human thinking. But daily interactions or situations are much trickier for AI. Your phone’s camera, for example, reads the visual information in a frame and focuses on a particular subject utilizing AI. However, differentiating between a white shirt and a white wall can cause AI to fail because it doesn’t recognize the other differences between a shirt and a wall, only the color. Another example might be when a driverless car sits still for hours at a road intersection where the traffic lights are not working.

AI is capable of very impressive and incredible things, yet when it comes to common sense, which humans take for granted, it seems to fail miserably. The question now becomes not whether we should stop thinking and let machines think for us, but we should ask a theoretical, but more pressing question: Would you let an AI-machine nanny watch your baby?

Before we answer this question, let’s first understand what common sense is, how humans learn it, why AI struggles with it, and how far AI is from achieving any semblance of common sense.

What is common sense?

By definition, common sense is something that everyone has. It is defined as “the basic ability to perceive, understand, and judge things that are shared by (‘common to’) nearly all people and can reasonably be expected of nearly all people without need for debate.” 

Humans are usually not conscious of the vast sea of common-sense assumptions that underlie every statement or action. This shared, unstated background knowledge includes a general understanding of how the physical world works (i.e., intuitive physics -- if you let go of a pen, it'll fall to the floor), a basic understanding of human motives and behaviors (i.e., intuitive psychology -- if you randomly punch someone, you'll upset them), and a knowledge of the common facts that an average adult possesses.

The world presents endless unforeseen circumstances. AI researchers call these situations “corner cases,” which lie on the outskirts of the likely or anticipated. In such situations, human minds can rely on common sense to carry them through, but AI systems often stumble.

Common sense does not sound like a big deal. But imagine living without it, and it comes into clearer focus. Suppose you are a robot lacking common sense, and you visit a carnival, and you confront a fun-house mirror. You might wonder if your body has suddenly changed.

Or, on the way home, you see that a fire hydrant has erupted, showering the road; you can’t determine if it’s safe to drive through the spray. You park outside a drugstore, and a man on the sidewalk screams for help, bleeding profusely. Are you allowed to grab bandages from the store without waiting in line to pay?

As a human being, you can draw on a vast reservoir of implicit knowledge to interpret these situations. You do so all the time because life is full of “corner-cases.” AI systems are likely to get stuck.

Why do AI systems struggle with common sense? How do humans learn common sense?

AI systems are rules-based machine learning neural net systems, which do a great job of processing large amounts of data and form generalized pattern detectors. But you can't talk to them about what they know or give them little bits of advice, like there's no right turn in this town or I've been through this intersection a few times recently and there's something quirky about this light. With the architecture they currently use, it cuts out these kinds of possibilities.

AI systems are based on data – a lot of data – from text to videos to images, in fact anything that you can conceive on the World Wide Web. This data is trained with sophisticated algorithms to derive patterns. Building common sense this way is not natural. It’s like living in a library: would a child secluded from birth in a room with broadband, Wikipedia, and YouTube emerge as an adult ready to navigate the world? Of course not.

How do human beings acquire common sense? In essence, we are multifaceted learners. We try things out and observe the results, read books, and listen to instructions, absorb silently (and sometimes not-so-silently), and reason on our own. We fall on our faces and watch others make mistakes. AI systems, by contrast, are not as well-rounded. They are rule-based and tend to follow one route at the exclusion of all others.

The absence of common sense prevents intelligent systems from understanding their world, from behaving reasonably in unforeseen situations, from communicating naturally with people, and from learning from new experiences. Its absence is considered the most significant barrier between the narrowly-focused AI applications of today and the more general, human-like AI systems hoped for in the future.

Common sense reasoning’s obscure but pervasive nature makes it difficult to articulate and encode.

Why is AI common sense important?

In the 1970s and 1980s, AI researchers thought that they were close to programming common sense into computers, but then they realized that it is too difficult and turned to easier problems, such as object recognition and language translation instead. Today the picture looks different. Many AI systems, such as driverless cars, may soon be working regularly alongside us in the real world. This makes the need for AI common sense more acute.

The situation where common sense really matters is where a system is going to be autonomous, running without direct human oversight and control, is in any kind of situation where any harm can be caused. That can be physical, financial, or mental harm, or even just misunderstandings. AI must get this right; it must have common sense, and act the way people do because things often happen unexpectedly and without a script.

Self-driving cars is the most obvious use-case we need to solve for, because we are driving heavy vehicles that might turn into dangerous weapons both to the people inside and people and property on the outside. This can be very dangerous if something goes wrong. There are so many bizarre things that might only happen once in your lifetime, but many of those kinds of things happen all the time.

Driving to work you probably see stuff that you have literally never seen before -- for example, you might see a squirrel on the side of the road holding a thimble. But if you saw that, you'd probably quickly realize that somebody might have dropped some sewing materials on the street and the squirrel picked it up. You would immediately know it is not dangerous, and even funny, and you wouldn't stop what you are doing.

You'll probably never see a squirrel like that, and neither will most people you know. Yet you'd be able to easily absorb and process that information, despite how unlikely it is. AI wouldn't be programmed to adapt to such a sight, so what would it do? Would it brake to a sudden stop, or swerve out of the way? We don't really know, and that's part of the problem. AI is predictable until it isn't.

Human common sense seems intuitive, instinctive, like second nature, something we do all the time without even thinking. Could we teach AI this type of intuitive common sense?

Experiments with AI common sense

Learning based on mountains of texts and videos is not enough; learning has to be supplemented by approaches that are “infant inspired.” In this line of research, AI learns common sense not by analyzing text or video but by solving problems in simulated virtual environments.

Computer scientists have collaborated with developmental psychologists to understand what we might call “baby sense” – the core skills of navigation, object manipulation, and social cognition that a small child might use. From this perspective, common sense is what you use to build a block tower with a friend, for example.

At the Allen Institute, researchers have created a three-dimensional digital home interior called THOR, meaning “the house of interactions.” It resembles a video game and is filled with manipulable household objects. Yejin Choi, a professor at Washington University, has been working with Allen Institute and built an AI lab to inhabit the space, called PIGLET, which is designed to use “physical interaction as grounding for language.” 

Using words, you can tell PIGLET about something that exists inside the house. For example, you might say that “there is a cold egg in a pan.” You can then ask it to predict what will happen when an event unfolds: “The robot slices the egg.”

The software translates these words into instructions for a virtual robot, which tries them out in THOR, where the outcome is determined by the laws of physics. It then reports back on what’s happened: “The egg is sliced.”

PIGLET delivers a common-sense answer four out of five times. Of course, its scope is limited. There is so much that could happen in the house, and to the house, not to mention all the endless scenarios outside the house. Our lives are filled with countless events and situations – expected and unexpected. Would it be possible to simulate and examine them all?

Another attempt at achieving AI common sense is Transformer – a type of advanced deep learning. Transformers are able to model natural language in a powerful way and, with some adjustments, are able to answer simple commonsense questions. Answering questions is an essential first step for building chatbots that can converse in a human-like way.

It’s already becoming painfully clear that even research in transformers is yielding diminishing returns. Transformers are getting larger and more power-hungry. A recent Baidu Transformer has several billion parameters. It takes an enormous amount of data to effectively train. Yet it has so far proved unable to grasp the nuances of human common sense.

Even deep learning pioneers seem to think that new fundamental research may be needed before today’s neural networks are able to make such a leap. It’s difficult to tell whether machine common sense is five years away, or 50. These ambitious efforts have recognized machine common sense as a moonshot AI problem of our times, one requiring concerted collaborations across institutions over many years.

This explains IBM's clarification of their announcement that they would be replacing jobs with AI. Tasks, like providing employment verification letters or moving employees between departments, which are administrative and repetitive, will likely be fully automated. However, HR functions, such as evaluating workforce composition and productivity, which would require common sense and empathy, are not expected to be replaced within the next decade.

Now, back to the question posed before: Would you let an AI-machine nanny watch your baby? It seems that the answer is likely not in the foreseeable future, if ever.

The views and opinions expressed herein are the views and opinions of the author and do not necessarily reflect those of Nasdaq, Inc.

Other Topics

Technology Innovation