About a month ago, a friend asked me a pointed question: “Why do you think artificial intelligence (AI) will have a positive impact on society?” This gave me pause and caused me to reflect. Indeed, it sometimes feels like developments in AI have lost sight of an end purpose. This can be explained by the fact that AI technologies are more technology push than market pull. The scientific aspiration to develop machines that can intelligently maneuver and demonstrate humanlike cognition pushes AI technologies into the market without much consideration of user need.
We build technology as “a means to fulfill a human purpose,” to solve problems and elevate the quality of human life. But in practice, we need only look around in our current internet era to see how the unchecked growth of technology can have negative effects. While the sum total impact of the internet has been positive, it has also led to social media addiction, massive breaches of data privacy, and more specifically, a coordinated attack on a large political election.
My friend’s question was important, and I interpreted the subtext as the following: given the talk about job displacement, autonomous weapons, and algorithmic bias, why do we think AI will have a positive impact on society?
But first, what does artificial intelligence mean?
Merriam-Webster defines AI as “the capability of a machine to imitate intelligent human behavior.” But the term has been used to describe a technique, a field of research, and even a technology. What is it?
It is definitely not a technique. A technique is a way of doing something. The techniques of AI are the methods that make computers seem intelligent, like machine learning and expert systems.
It is an area of research within computer science — it includes subdomains like reasoning, planning, learning, language understanding, perception, manipulation, and social ability. Researchers in this field study “the designing and building of intelligent agents that receive percepts from the environment and take actions that affect that environment.”
While AI isn’t a specific technology, it has come to describe a body of technologies, which is how the term is used in this piece. AI technologies are applied to problems that involve performing human-like tasks. Examples of specific AI technologies include speech recognition, object detection, and recommendation systems.
To sum it up: the field of AI research attempts to build systems that “imitate intelligent human behavior,” and in the process develops techniques that provide the foundation of AI technologies that can be applied to a variety of problems.
A framework to analyze the impact of AI
We will analyze AI one level of abstraction above its specific applications and technologies by focusing on its implications. What are the common themes underlying self-driving cars, personal assistants, manufacturing robots, and other developments in AI? What is the nature of their impact? To answer these questions, I will use the following benefit-risk framework:
Benefits
- Reduce information overload via hyper-personalized experiences
- Create more natural interfaces for interacting with the world around us
- Augment human performance and creativity to make us more effective
- Automate difficult, dangerous, and dull tasks
Risks
- Increase the severity of digital and physical attacks
- Perpetuate bias at scale
- Negatively affect labor markets
Let’s start by walking through the benefits and what they entail.
It will reduce information overload via hyper-personalized experiences
The internet has led to an explosion in the amount of information we see, inundating us with more than we can process. The shift to online media and retail offers consumers thousands of movies to choose from, tens of millions of songs to listen to, and thousands of products for a single search. How do we sift through all this information and make choices?
Machine learning and recommendation systems provide a solution to this problem by personalizing digital experiences based on data about an individual’s preferences. Xerox PARC developed the first recommendation system in the mid-1990s, called Tapestry. It made it easier to read through hundreds of new postings every day based on what the user had read before. Though we have built recommendation systems to deal with information overload for a long time, the issue has only become more pronounced. Spotify’s Discover Weekly playlist is a modern example of a sophisticated recommendation system that suggests music based on one’s listening history. In the future, we should expect personalized experiences that are not just statistical i.e. machine learning, but also semantic — they will be able to explain why something was recommended. We will also see more predictive experiences where the products we use foresee intended actions and surface them before we ask.
It will create more natural interfaces for interacting with the world around us
AI technologies like speech recognition, computer vision, and robotics correspond directly to hearing, sight, and touch, three of the five basic senses. The ability of machines to interpret stimuli for these three senses enables them to help humans manipulate the world more easily.
Here’s another lens by which to view this point: we interact with the devices in our daily life through constrained interfaces. Since the invention of the computer, most digital interactions have been funneled through a graphical user interface, keyboard, and mouse. The introduction of the iPhone in 2007 popularized the touch screen, a slightly more natural mode of interaction. Now, speech, language, and vision technologies enable user experiences where we can talk and move naturally.
One example is how voice assistants are helping the disabled and elderly live more empowered lives. Researchers conducted two exploratory studies to understand how users with disabilities are using off-the-shelf voice assistants and found that they have “replaced many disparate devices, and improved efficiency and independence for a variety of tasks.” The visually impaired, blind, and physically disabled often struggle to use traditional devices like computers, mobile phones, tablets, and home appliances. They may depend on others to assist them with the time, weather, and other simple tasks. The confluence of information extraction, speech recognition, and speech synthesis enable a voice assistant like Alexa to interpret natural language queries and respond accordingly. As we look to the future, the merging of voice assistants and robotics will increase the range of tasks with which technology can aid these populations.
It will augment human performance and creativity to make us more effective
AI has the ability to augment human intuition and creativity with computational power and information, empowering humans to do things they never could have done otherwise. Whereas automation improves the efficiency of an existing process or task, augmentation works alongside a human to solve a new problem.
In the life sciences, AI technologies are being applied to augment experts in the drug discovery process. Drug discovery involves searching for compounds that bind to a target molecule to cause the desired effect on a disease or condition. Here, AI isn’t automating the entire process. It is being used by experts in certain steps of the workflow to speed things up and discover patterns in data that humans would never see. Lead optimization is one step where deep learning has great potential. By this step, the biological target has been identified, and chemists are tweaking the lead compound’s chemical properties to optimize for potency and other factors. Traditionally, this tweaking was manual and based on intuition and chance. However, neural networks are effective at solving this type of optimization problem and can search through the large space of compounds more intelligently than humans. With AI, some processes in drug discovery that used to take years can be compressed into a matter of weeks.
Though less applicable to the real world than drug discovery, gaming provides a glimpse of future possibilities. DeepMind’s recent advances in building systems that play Go (AlphaGo) and Starcraft (AlphaStar) better than humans demonstrate what humans could learn from AI systems. Go players have much to learn from examining the strategies of AlphaGo. One example is the famous move 37 played by AlphaGo, which took Lee Sedol 15 minutes to respond to. It stunned professional Go players, who called it “creative” and “unique.” In Starcraft, players could use AlphaStar as a tool to explore the massive space of build strategies and micro techniques.
It will automate difficult, dangerous, and dull tasks
The world is abundant with tasks that are too difficult, dangerous, or dull for humans to perform. AI technologies can perform these tasks better and at a lower cost. These tasks are sometimes simple, like when an autonomous mobile robot moves packages around in a fulfillment warehouse. Other tasks require a system to adapt to complex variables in the environment, like an autonomous car driving through a city in the rain.
The act of driving is difficult because of the complexity the driver has to manage: they must pay full attention to the road, be aware of other vehicles on all sides, manage lane changes, watch for pedestrians and bikers, all while not being distracted by raucous passengers. Self-driving cars have the potential to perform this task more safely and cut down tens of thousands of deaths per year.
Industrial welding is a task that is dangerous for humans. The environment features hazardous fumes, intense heat, and loud noise, which can cause lung and kidney diseases and hearing loss. Robotic welding can fully automate the welding process, allowing humans to safely monitor the task from afar.
Finally, AI technologies improve operational efficiency by automating dull tasks. The growth of e-commerce and the lack of an available workforce are driving the adoption of autonomous mobile robots and “cobots” to automatically sort and pick packages in logistics. Slowly but surely, the logistics industry is progressing toward fully automated warehouses that can fulfill hundreds of thousands of orders a day with just a few people.
Automation can unlock the ability for more people to spend their time on activities that they are uniquely suited to do, like creative tasks and engaging in experiences. It has the potential to elevate the quality of life for generations to come.
At this point, we might stop and say that AI will clearly have a positive impact. That would be absolving it from numerous associated risks.
It can negatively affect labor markets
The dark side of automation is that it can displace the jobs of human workers who may be lower-skilled and depend on it for their livelihood. This is a common fear that is discussed in the context of autonomous vehicles, manufacturing robots, and customer service bots. Some studies claim that AI will automate millions of jobs, leaving workers with no place to go. Others cite economic growth from the industrial revolution and say we have nothing to worry about. While new jobs will undoubtedly be created, in-depth analyses by domain experts are showing that AI technologies can have negative impacts on some labor markets.
In Driverless, a sociologist at the University of Pennsylvania analyzed the US trucking industry and autonomous truck technology to understand its labor implications. He found that 294,000 long-haul trucking jobs are at risk of automation in the next 25 years. The growth of e-commerce, lower freight costs, and the rise of autonomous truck ports will create new jobs, but these will be lower quality trucking jobs. As long-haul trucking jobs are automated, demand will increase for jobs similar to “port driving” (moving freight to and from autonomous truck ports). These new jobs will have worse wages, hours, and labor protections. The report’s message is clear: policy is needed to protect truckers as autonomous trucks are commercialized.
It can increase the severity of digital and physical attacks
The dual-use nature of any technology means that AI can be weaponized. While the internet’s creation of a digital world has expanded the ways in which someone can be attacked, AI increases the severity of such attacks.
“Deep fakes” are one example of how AI raises the bar for digital attacks. They are an example of how bad actors can use video and speech synthesis technology to create hyper-realistic digital forgeries to blackmail people and spread misinformation. Before, the worst that could be done to ruin someone’s reputation digitally was writing a defamatory statement or hacking into their account to pose as them. “Deep fakes,” however, makes it easy to publish defamatory videos or images that actually show the victim doing something they never did. They could create a society where every piece of information needs to be verified for its truthfulness.
In the physical world, advances in robotics and computer vision threaten to enable the development of lethal autonomous weapons. An autonomous drone outfitted with a weapons system will dehumanize killing and increase the precision and reach of devastation. A human would no longer need to be involved in the operation, not even from behind a control panel. “Slaughterbots” is a fictional video that warns nations against developing such technology by showing its possible implications. If in the wrong hands, lethal autonomous weapons can cause greater damage than traditional drones.
It can perpetuate bias at scale
Algorithmic bias occurs when predictive models exhibit “systematic and unfair” discrimination. This discrimination usually results from bias in the underlying data used to train the model. The data may not be representative of the population the model is used on, or it can reflect human biases that exist in the world.
Predictive systems have already begun to surround us in both low-stakes and high-stakes environments. While personalized movie recommendations tell a user what to watch, predictive policing is actually used to help police officers identify potential criminals. Since these systems can be widely deployed quickly and easily, they can perpetuate bias at massive scale.
In the past few years, there have been numerous accounts of algorithmic bias in in-development and production systems. Machine learning-based recruiting tools are discriminating against women. Facial recognition software has inaccurately classified black people as “gorillas.” Gender and skin-color bias has been found in commercial facial analysis systems. Some organizations like the Algorithmic Justice League are doing great work bringing attention to the issue, but the community as a whole has far to go.
In summary, AI has the potential to elevate humanity, but only if we can deal with its risks.
Let’s recall my friend’s question: “Why do you think AI will have a positive impact on society?” It clearly has the potential to, and humanity’s ability to solve problems leads me to believe that it will.
Hyper-personalization and more natural interfaces will reduce information overload and help technology work for people. Augmentation and automation will elevate our quality of life and allow us to focus on uniquely human activities.
The risks of AI are pressing, but not unsolvable. We can protect against increased attack severity by developing novel safety techniques, like media forensics to fight back against “deep fakes.” Algorithmic bias is a consequence of poorly designed AI systems — it is not inherent to AI systems themselves. Additional emphasis on retraining programs may help ease the effects of job automation.
It is the nature of AI to provide us with even more powerful digital tools, but it is the duty of humanity to wield them responsibly.
If you have comments or feedback on the ideas discussed in this piece, please reach out!
