Facial recognition, search engines, social media algorithms, ad recommendations and robots are rooted in a technology that is present in nearly all aspects of daily life: artificial intelligence.

AI refers to the idea that humans can train computers to be “rational agents” that have preferences and make informed decisions, according to Sergey Levine, a UC Berkeley assistant professor in the electrical engineering and computer sciences department. AI systems aim to optimize a particular outcome that they have been instructed to prefer.

The term artificial intelligence dates back to the 1950s when psychologists studied the human brain and explored how machines could simulate human intelligence, according to Levine. The basic concepts of AI are much older, Levine noted, referencing the writings of mathematician Alan Turing, who suggested that computers be programmed to act like children rather than adults.

“If your preferences about outcomes in the world can be put into an ordering and assigned numerical values, then you are rational,” Levine said. “From that basic idea, we can derive methods and computer algorithms that allow us to actually construct artificial agents that behave in rational ways.”

How do computers learn behavior?

AI systems are designed to optimize a particular utility, according to campus senior Albert Yu, head TA for Computer Science 188, “Introduction to Artificial Intelligence.”

YouTube recommendations, for example, use AI to maximize the amount of time a user spends on the platform, which in turn increases profits, according to Yu. Other popular applications of AI in society today include search engines, which optimize for the most relevant web pages based on a given search query; self-driving cars, which are optimized to reach a destination safely and quickly and advertisements, which aim to maximize revenue.

AI systems are most commonly taught to perform these and other functions through a mechanism known as supervised learning, Levine said. This technique consists of providing a computer with examples of data that are classified as a particular category.

AI is useful because it can perform these tedious computations over a large amount of data and quickly learn to recognize and classify objects in images, Yu added.

“(AI is) able to process a ton of data in ways and at speeds that humans can’t necessarily process,” Yu said. “As long as we have internet access to all of the data and pictures and language in the world, it does have the potential of being more accurate than humans in some very specific tasks.”

Nikita Samarin, a GSI for CS 188, demonstrated the process using an example of an algorithm with the goal is to distinguish between apples and oranges. Instead of informing a machine about the differences between the fruits, the machine is shown several images of apples and several images of oranges, Samarin said. Over time, the system learns to differentiate between the two fruits at a high level of accuracy.

Beyond teaching the systems, AI also entails studying what tools machines require to be capable of learning and having intelligence, Samarin added.

Levine said his lab — the Robotics AI and Learning Lab — studies another type of AI learning mechanism known as reinforcement learning. Instead of telling a machine exactly what to do, reinforcement learning involves providing feedback on whether the machine is doing something good or bad.

This type of learning is particularly helpful when considering applications of robots in the real world where it can be difficult to provide exact instructions on how a robot should accomplish a task such as navigating a forest, Levine noted. Reinforcement learning informs the robot on whether it performs well or poorly, and the robot adjusts its behavior accordingly.

“A very effective way to use AI is to handle those aspects of decision-making that we’re not so great at and provide support for human decision-making,” Levine said. “Longer term, there is a lot of promise in AI systems that do interact with the physical world in meaningful ways.”

Fairness and transparency: Ethical considerations of AI

While effective, AI presents pressing ethical concerns that call for regulation and governance of the technology.

According to Samarin, AI machines are trained by the data they are given, which are inherently biased. Oftentimes, these systems will indirectly reflect the bias present in the data.

There have been “disturbing” instances where AI use has yielded discriminatory outcomes, Samarin noted, including when it was used in hiring processes and when it was used to determine whether someone should be released from bail.

Brandie Nonnecke, director of the Center for Information Technology Research in the Interest of Society and the Banatao Institute, said AI technologies should be developed with principles of fairness, accountability and transparency. She added that AI-enabled tools should be transparent in terms of the data used, how the model was built and any assumptions that were made.

Nonnecke is also co-chair of the UC system’s Presidential Working Group on Artificial Intelligence, which aims to develop a set of ethical principles to guide the university in its use of AI. The working group focuses on four main applications of AI: health, human resources, law enforcement and student experience.

“Over the past few years, it’s become apparent that AI-enabled systems can have harmful effects on society — especially, the revelations we’ve found that AI models can perpetuate bias and discrimination,” Nonnecke said. “In light of these findings, there has been great effort to develop appropriate governance mechanisms for AI.”

Yu described a malicious application of AI known as generative adversarial networks, a form of AI that can not only generate realistic images of fake people but also be used to make videos of a person saying something very different from what they actually said.

This has particularly dangerous implications in the world of politics. The possibility for AI to generate a “deepfake” of a public figure such as a president saying something they did not actually say is “very concerning,” Yu said.

He added that AI also poses a problem when its objectives do not align with those held by the user. An AI that aims to increase users’ watch time on a web page, for example, may not align with a person’s objective to balance work and relaxation.

“Right now, researchers can’t explain the decision-making algorithms,” Samarin said. “The direction of trying to understand how to make the systems more transparent, more accountable and fairer is very important, and we should really think about that as we go forward.”

Current and future impacts of AI

AI’s impact on society today has both positive and negative consequences for the world tomorrow.

Mesut Yang, a GSI for CS 188, described AI’s impact in the vast field of social media. Yang explained that AI algorithms aim to learn the behavior and opinions of individual users to continue recommending content that they know the users will enjoy. This often drives individuals down a “rabbit hole” to increase user engagement.

Yang noted that this has dangerous ramifications in regard to extremism. When users go on social media, algorithms will feed them content that confirms their beliefs; one example is the role Facebook played in supporting Myanmar’s ethnic cleansing campaign.

“One of the issues with a lot of current AI systems is that they are not grounded in the same physical experience as we are and therefore they understand the world differently than we do,” Levine said. “And that’s the reason why these systems behave in a way that is counterintuitive to us.”

There are, nonetheless, several positive impacts of AI, and a lot of university research is being done to discover even more practical applications.

Nonnecke described how California’s Employment Development Department, or EDD, had a backlog of approximately 1.6 million unemployment claims in October 2020, as many people applied for unemployment benefits during the COVID-19 pandemic.

She said the main reason for this backlog was that the application process required an employee to manually verify an individual’s identity. To increase efficiency, EDD developed an automated process to verify unemployment by using a branch of AI known as computer vision.

Computer vision bridges AI and the visual world and it allows AI agents to interpret and classify images, Yang said.

Angjoo Kanazawa, a campus assistant professor in electrical engineering and computer sciences, specializes in computer vision research with a focus on modeling the 3D world. To illustrate computer vision’s potential, she said AI systems can recover 3D features of someone’s face simply from an image.

Her research extends to 3D motion capture, which can be useful for medical diagnosis and rehabilitation.

“My goal is to really take a single image or video and understand what’s happening in the 3D world,” Kanazawa said. “Ideally, this is getting close to what it means to perceive an image.”

Another important aspect of university research is natural language processing, Yu said, which relates to interpreting human language. Google Translate is an example of natural language processing.

Yang is currently conducting research in the field of robotics, specifically to improve human-robot interactions. His research aims to identify how to train AI agents to understand human behavior and commands.

With all its potential for greatness yet possible manipulation, society’s knowledge about the impacts of AI remains limited, according to Yang. He said more work should be done to increase the public’s knowledge about AI, unlocking future opportunities.

“Improving the general public’s literacy about artificial intelligence is really important,” Yang said. “After most of society is up to date on what AI is and what its strengths and flaws are when we are all standing on this common knowledge as a society, we’re able to move forward a lot better and a lot faster.”