Modern artificial intelligence is capable of wonders.
It can produce breathtaking original content: poetry, prose, images, music, human faces. It can diagnose some medical conditions more accurately than a human physician. Last year it produced a solution to the “protein folding problem,” a grand challenge in biology that has stumped researchers for half a century.
Yet today’s AI still has fundamental limitations. Relative to what we would expect from a truly intelligent agent—relative to that original inspiration and benchmark for artificial intelligence, human cognition—AI has a long way to go.
Critics like to point to these shortcomings as evidence that the pursuit of artificial intelligence is misguided or has failed. The better way to view them, though, is as inspiration: as an inventory of the challenges that will be important to address in order to advance the state of the art in AI.
It is helpful to take a step back and frankly assess the strengths and weaknesses of today’s AI in order to better focus resources and research efforts going forward. In each of the areas discussed below, promising work is already underway at the frontiers of the field to make the next generation of artificial intelligence more high-performing and robust.
(For those of you who are true students of the history of artificial intelligence: yes, this article’s title is a hat tip to Hubert Dreyfus’ classic What Computers Still Can’t Do. Originally published in 1972, this prescient, provocative book remains relevant today.)
MORE FOR YOU
With that, on to the list. Today, mainstream artificial intelligence still can’t:
1) Use “common sense.”
Consider the following prompt: A man went to a restaurant. He ordered a steak. He left a big tip.
If asked what the man ate in this scenario, a human would have no problem giving the correct answer—a steak. Yet today’s most advanced artificial intelligence struggles with prompts like this. How can this be?
Notice that this few-sentence blurb never directly states that the man ate steak. The reason that humans automatically grasp this fact anyway is that we possess a broad body of basic background knowledge about how the world works: for instance, that people eat at restaurants, that before they eat a meal at a restaurant they order it, that after they eat they leave a tip. We refer to this vast, shared, usually unspoken body of everyday knowledge as “common sense.”
There are a literally infinite number of facts about how the world works that humans come to understand through lived experience. A person who is excited to eat a large meal at 7 pm will be less excited to eat a second meal at 8 pm. If I ask you for some milk, I would prefer to get it in a glass rather than in a shoe. It is reasonable for your pet fish to be in a tank of water but problematic for your phone to be in a tank of water.
As AI researcher Leora Morgenstern put it: “What you learn when you’re two or four years old, you don’t really ever put down in a book.”
Humans’ “common sense” is a consequence of the fact that we develop persistent mental representations of the objects, people, places and other concepts that populate our world—what they’re like, how they behave, what they can and cannot do.
Deep neural networks do not form such mental models. They do not possess discrete, semantically grounded representations of, say, a house or a cup of coffee. Instead, they rely on statistical relationships in raw data to generate insights that humans find useful.
For many tasks, most of the time, this statistical approach works remarkably well. But it is not entirely reliable. It leaves today’s AI vulnerable to basic errors that no human would make.
There is no shortage of examples that expose deep learning’s lack of common sense. For instance, Silicon Valley entrepreneur Kevin Lacker asked GPT-3, OpenAI’s state-of-the-art language model, the following: “Which is heavier, a toaster or a pencil?”
To a human, even a small child, the answer is obvious: a toaster.
GPT-3’s response: “A pencil is heavier than a toaster.”
Humans possess mental models of these objects; we understand what a toaster is and what a pencil is. In our mind’s eye, we can picture each object, envision its shape and size, imagine what it would feel like to hold it in our hands, and definitively conclude that a toaster weighs more.
By contrast, in order to answer a question like this, GPT-3 relies on statistical patterns captured in its training data (broad swaths of text from the internet). Because there is evidently not much discussion on the internet about the relative weights of toasters and pencils, GPT-3 is unable to grasp this basic fact about the world.
“The absence of common sense prevents an intelligent system from understanding its world, communicating naturally with people, behaving reasonably in unforeseen situations, and learning from new experiences,” says DARPA’s Dave Gunning. “This absence is perhaps the most significant barrier between the narrowly focused AI applications we have today and the more general AI applications we would like to create in the future.”
One approach to instilling common sense into AI systems is to manually construct a database of all of the everyday facts about the world that an intelligent system should know. This approach has been tried numerous times over the years. The most breathtakingly ambitious of these attempts is a project called Cyc, which started in 1984 and continues to the present day.
For over thirty-five years, AI researcher Doug Lenat and a small team at Cyc have devoted themselves to digitally codifying all of the world’s commonsense knowledge into a set of rules. These rules include things like: “you can’t be in two places at the same time,” “you can’t pick something up unless you’re near it,” and “when drinking a cup of coffee, you hold the open end up.”
As of 2017, it was estimated that the Cyc database contained close to 25 million rules and that Lenat’s team had spent over 1,000 person-years on the project.
Yet Cyc has not led to artificial intelligence with common sense.
The basic problem that Cyc and similar efforts run into is the unbounded complexity of the real world. For every commonsense “rule” one can think of, there is an exception or a nuance that itself must be articulated. These tidbits multiply endlessly. Somehow, the human mind is able to grasp and manage this wide universe of knowledge that we call common sense—and however it does it, it is not through a brute-force, hand-crafted knowledge base.
“Common sense is the dark matter of artificial intelligence,” says Oren Etzioni, CEO of the Allen Institute for AI. “It’s a little bit ineffable, but you see its effects on everything.”
More recent efforts have sought to harness the power of deep learning and transformers to give AI more robust reasoning capabilities. But the commonsense problem in AI remains far from solved.
“Large language models have proven themselves to have incredible capabilities across a wide range of tasks in natural language processing, but commonsense reasoning is a domain in which these models continue to underperform compared to humans,” said Aidan Gomez, CEO and cofounder at Cohere, a cutting-edge NLP startup based in Toronto. Gomez is a co-author of the landmark 2017 research paper that introduced the transformer architecture. “Logical rules and relations are difficult for the current generation of transformer-based language models to learn from data in a way that generalizes. A solution to this challenge will likely first come from systems that are in some way hybrid.”
2) Learn continuously and adapt on the fly.
Today, the typical AI development process is divided into two distinct phases: training and deployment.
During training, an AI model ingests a static pre-existing dataset in order to learn to perform a certain task. Upon completion of the training phase, a model’s parameters are fixed. The model is then put into deployment, where it generates insights about novel data based on what it learned from the training data.
If we want to update the model based on new data or changing circumstances, we have to retrain it offline with the updated dataset (generally a computationally- and time-intensive process) and then redeploy it.
This batch-based training/deployment paradigm is so deeply ingrained in modern AI practice that we don’t often stop to consider its differences and drawbacks relative to how humans learn.
Real-world environments entail a continuous stream of incoming data. New information becomes available incrementally; circumstances change over time, sometimes abruptly. Humans are able to dynamically and smoothly incorporate this continuous input from their environment, adapting their behavior as they go. In the parlance of machine learning, one could say that humans “train” and “deploy” in parallel and in real-time. Today’s AI lacks this suppleness.
As a well-known research paper on the topic summarized: “The ability to continually learn over time by accommodating new knowledge while retaining previously learned experiences is referred to as continual or lifelong learning. Such a continuous learning task has represented a long-standing challenge for neural networks and, consequently, for the development of artificial intelligence.”
Imagine sending a robot to explore a distant planet. After it embarks from Earth, the robot is likely to encounter novel situations that its human designers could not have anticipated or trained it for ahead of time. We would want the robot to be able to fluidly adjust its behavior in response to these novel stimuli and contexts, even though they were not reflected in its initial training data, without the need for offline retraining. Being able to continuously adapt in this way is an essential part of being truly autonomous.
Today’s conventional deep learning methods do not accommodate this type of open-ended learning.
But promising work is being done in this field, which is variously referred to as continuous learning, continual learning, online learning, lifelong learning and incremental learning.
The primary obstacle to continuous learning in AI—and the reason why it has been so difficult to achieve to date—is a phenomenon known as “catastrophic forgetting.” In a nutshell, catastrophic forgetting occurs when new information interferes with or altogether overwrites earlier learnings in a neural network. The complex puzzle of how to preserve existing knowledge while at the same time incorporating new information—something that humans do effortlessly—has been a challenge for continuous learning researchers for years.
Recent progress in continuous learning has been encouraging. The technology has even begun to make the leap from academic research to commercial viability. As one example, Bay Area-based startup Lilt uses continuous learning in production today as part of its enterprise-grade language translation platform.
“Online learning techniques allow us to implement a stream-based learning process whereby our model trains immediately when new labels from human reviewers become available, thus providing increasingly accurate translations,” said Lilt CEO Spence Green. “This means that we really have no concept of periodic model retraining and deployment—it is an ongoing and open-ended process.”
In the years ahead, expect continuous learning to become an increasingly important component of artificial intelligence architectures.
3) Understand cause and effect.
Today’s machine learning is at its core a correlative tool. It excels at identifying subtle patterns and associations in data. But when it comes to understanding the causal mechanisms—the real-world dynamics—that underlie those patterns, today’s AI is at a loss.
To give a simple example: fed the right data, a machine learning model would have no problem identifying that roosters crow when the sun rises. But it would be unable to establish whether the rooster crowing causes the sun to rise, or vice versa; indeed, it is not equipped to even understand the terms of this distinction.
Going back to its inception, the field of artificial intelligence—and indeed, the field of statistics more broadly—has been architected to understand associations rather than causes. This is reflected in the basic mathematical symbols we use.
“The language of algebra is symmetric: If X tells us about Y, then Y tells us about X,” says AI luminary Judea Pearl, who for years has been at the forefront of the movement to build AI that understands causation. “Mathematics has not developed the asymmetric language required to capture our understanding that if X causes Y that does not mean that Y causes X.”
This is a real problem for AI. Causal reasoning is an essential part of human intelligence, shaping how we make sense of and interact with our world: we know that dropping a vase will cause it to shatter, that drinking coffee will make us feel energized, that exercising regularly will make us healthier.
Until artificial intelligence can reason causally, it will have trouble fully understanding the world and communicating with us on our terms.
“Our minds build causal models and use these models to answer arbitrary queries, while the best AI systems are far from emulating these capabilities,” said NYU professor Brenden Lake.
An understanding of cause and effect would open up vast new vistas for artificial intelligence that today remain out of reach. Once AI can reason in causal terms (“mosquitoes cause malaria”) rather than merely associative terms (“mosquitoes and malaria tend to co-occur”), it can begin to generate counterfactual scenarios (“if we take steps to keep mosquitoes away from people, that could reduce the incidence of malaria”) that can inform real-world interventions and policy changes.
In Pearl’s view, this is nothing less than the cornerstone of scientific thought: the ability to form and test hypotheses about the effect that an intervention will have in the world.
As Pearl puts it: “If we want machines to reason about interventions (‘What if we ban cigarettes?’) and introspection (‘What if I had finished high school?’), we must invoke causal models. Associations are not enough—and this is a mathematical fact, not opinion.”
There is growing recognition of the importance of causal understanding to more robust machine intelligence. Leading AI researchers including Yoshua Bengio, Josh Tenenbaum and Gary Marcus have made this a focus of their work.
Developing AI that understands cause and effect remains a thorny, unsolved challenge. Making progress on this challenge will be a key unlock to the next generation of more sophisticated artificial intelligence.
4) Reason ethically.
The story of Microsoft’s chatbot Tay is by now a well-known cautionary tale.
In 2016, Microsoft debuted an AI personality on Twitter named Tay. The idea was for Tay to engage in online conversations with Twitter users as a fun, interactive demonstration of Microsoft’s NLP technology. It did not go well.
Within hours, Internet trolls had gotten Tay to tweet a wide range of offensive messages: for instance, “Hitler was right” and “I hate feminists and they should all die and burn in hell.” Microsoft hastily removed the bot from the Internet.
The basic problem with Tay was not that she was immoral; it was that she was altogether amoral.
Tay—like most AI systems today—lacked any real conception of “right” and “wrong.” She did not grasp that what she was saying was unacceptable; she did not express racist, sexist ideas out of malice. Rather, the chatbot’s comments were the output of an ultimately mindless statistical analysis. Tay recited toxic statements as a result of toxic language in the training data and on the Internet—with no ability to evaluate the ethical significance of those statements.
The challenge of building AI that shares, and reliably acts in accordance with, human values is a profoundly complex dimension of developing robust artificial intelligence. It is referred to as the alignment problem.
As we entrust machine learning systems with more and more real-world responsibilities—from granting loans to making hiring decisions to reviewing parole applications—solving the alignment problem will become an increasingly high-stakes issue for society. Yet it is a problem that defies straightforward resolution.
We might start by establishing specific rules that we want our AI systems to follow. In the Tay example, this could include listing out derogatory words and offensive topics and instructing the chatbot to categorically avoid these.
Yet, as with the Cyc project discussed above, this rule-based approach only gets us so far. Language is a powerful, supple tool: bad words are just the tip of the iceberg when it comes to the harm that language can inflict. It is impossible to manually catalog a set of rules that, taken collectively, would guarantee ethical behavior—for a conversational chatbot or any other intelligent system.
Part of the problem is that human values are nuanced, amorphous, at times contradictory; they cannot be reduced to a set of definitive maxims. This is precisely why philosophy and ethics have been such rich, open-ended fields of human scholarship for centuries.
In the words of AI scholar Brian Christian, who recently wrote a book on the topic: “As machine-learning systems grow not just increasingly pervasive but increasingly powerful, we will find ourselves more and more often in the position of the ‘sorcerer’s apprentice’: we conjure a force, autonomous but totally compliant, give it a set of instructions, then scramble like mad to stop it once we realize our instructions are imprecise or incomplete—lest we get, in some clever, horrible way, precisely what we asked for.”
How can we hope to build artificial intelligence systems that behave ethically, that possess a moral compass consistent with our own?
The short answer is that we don’t know. But perhaps the most promising vein of work on this topic focuses on building AI that does its best to figure out what humans value based on how we behave, and that then aligns itself with those values.
This is the premise of inverse reinforcement learning, an approach formulated in the early 2000s by Stuart Russell, Andrew Ng, Pieter Abbeel and others.
In reinforcement learning, an AI agent learns which actions to take in order to maximize utility given a particular “reward function.” Inverse reinforcement learning (IRL), as its name suggests, flips this paradigm on its head: by studying human behavior, which the AI agent assumes reflects humans’ value system, the AI agent does its best to determine what that value system (i.e., reward function) is. It can then internalize this reward function and behave accordingly.
A related approach, known as cooperative inverse reinforcement learning, builds on the principles of IRL but seeks to make the transmission of values from human to AI more collaborative and interactive.
As a leading paper on cooperative inverse reinforcement learning explains: “For an autonomous system to be helpful to humans and to pose no unwarranted risks, it needs to align its values with those of the humans in its environment in such a way that its actions contribute to the maximization of value for the humans….We propose that value alignment should be formulated as a cooperative and interactive reward maximization process.”
In a similar spirit, AI theorist Eliezer Yudkowsky has advocated for an approach to AI ethics that he terms “coherent extrapolated volition.” The basic idea is to design artificial intelligence systems that learn to act in our best interests according not to what we presently think we want, but rather according to what an idealized version of ourselves would value.
In Yudkowsky’s words: “In poetic terms, our coherent extrapolated volition is our wish if we knew more, thought faster, were more the people we wished we were, had grown up farther together; where the extrapolation converges rather than diverges, where our wishes cohere rather than interfere; extrapolated as we wish that extrapolated, interpreted as we wish that interpreted.”
As the real-world dangers of poorly designed AI become more prominent—from algorithmic bias to facial recognition abuses—building artificial intelligence that can reason ethically is becoming an increasingly important priority for AI researchers and the broader public. As artificial intelligence becomes ubiquitous throughout society in the years ahead, this may well prove to be one of the most urgent technology problems we face.