Tag Archives: developing

#436200 AI and the Future of Work: The Economic ...

This week at MIT, academics and industry officials compared notes, studies, and predictions about AI and the future of work. During the discussions, an insurance company executive shared details about one AI program that rolled out at his firm earlier this year. A chatbot the company introduced, the executive said, now handles 150,000 calls per month.

Later in the day, a panelist—David Fanning, founder of PBS’s Frontline—remarked that this statistic is emblematic of broader fears he saw when reporting a new Frontline documentary about AI. “People are scared,” Fanning said of the public’s AI anxiety.

Fanning was part of a daylong symposium about AI’s economic consequences—good, bad, and otherwise—convened by MIT’s Task Force on the Work of the Future.

“Dig into every industry, and you’ll find AI changing the nature of work,” said Daniela Rus, director of MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL). She cited recent McKinsey research that found 45 percent of the work people are paid to do today can be automated with currently available technologies. Those activities, McKinsey found, represent some US $2 trillion in wages.

However, the threat of automation—whether by AI or other technologies—isn’t as new as technologists on America’s coasts seem to believe, said panelist Fred Goff, CEO of Jobcase, Inc.

“If you live in Detroit or Toledo, where I come from, technology has been displacing jobs for the last half-century,” Goff said. “I don’t think that most people in this country have the increased anxiety that the coasts do, because they’ve been living this.”

Goff added that the challenge AI poses for the workforce is not, as he put it, “getting coal miners to code.” Rather, he said, as AI automates some jobs, it will also open opportunities for “reskilling” that may have nothing to do with AI or automation. He touted trade schools—teaching skills like welding, plumbing, and electrical work—and certification programs for sales industry software packages like Salesforce.

On the other hand, a documentarian who reported another recent program on AI—Krishna Andavolu, senior correspondent for Vice Media—said “reskilling” may not be an easy answer.

“People in rooms like this … don’t realize that a lot of people don’t want to work that much,” Andavolu said. “They’re not driven by passion for their career, they’re driven by passion for life. We’re telling a lot of these workers that they need to reskill. But to a lot of people that sounds like, ‘I’ve got to work twice as hard for what I have now.’ That sounds scary. We underestimate that at our peril.”

Part of the problem with “reskilling,” Andavolu said, is that some high-growth industries involve caregiving for seniors and in medical facilities—roles which are traditionally considered “feminized” careers. Destigmatizing these jobs, and increasing the pay to match the salaries of displaced jobs like long-haul truck drivers, is another challenge.

Daron Acemoglu, MIT Institute Professor of Economics, faulted the comparably slim funding of academic research into AI.

“There is nothing preordained about the progress of technology,” he said. Computers, the Internet, antibiotics, and sensors all grew out of government and academic research programs. What he called the “blue-sky thinking” of non-corporate AI research can also develop applications that are not purely focused on maximizing profits.

American companies, Acemoglu said, get tax breaks for capital R&D—but not for developing new technologies for their employees. “We turn around and [tell companies], ‘Use your technologies to empower workers,’” he said. “But why should they do that? Hiring workers is expensive in many ways. And we’re subsidizing capital.”

Said Sarita Gupta, director of the Ford Foundation’s Future of Work(ers) Program, “Low and middle income workers have for over 30 years been experiencing stagnant and declining pay, shrinking benefits, and less power on the job. Now technology is brilliant at enabling scale. But the question we sit with is—how do we make sure that we’re not scaling these longstanding problems?”

Andrew McAfee, co-director of MIT’s Initiative on the Digital Economy, said AI may not reduce the number of jobs available in the workplace today. But the quality of those jobs is another story. He cited the Dutch economist Jan Tinbergen who decades ago said that “Inequality is a race between technology and education.”

McAfee said, ultimately, the time to solve the economic problems AI poses for workers in the United States is when the U.S. economy is doing well—like right now.

“We do have the wind at our backs,” said Elisabeth Reynolds, executive director of MIT’s Task Force on the Work of the Future.

“We have some breathing room right now,” McAfee agreed. “Economic growth has been pretty good. Unemployment is pretty low. Interest rates are very, very low. We might not have that war chest in the future.” Continue reading

Posted in Human Robots

#436188 The Blogger Behind “AI ...

Sure, artificial intelligence is transforming the world’s societies and economies—but can an AI come up with plausible ideas for a Halloween costume?

Janelle Shane has been asking such probing questions since she started her AI Weirdness blog in 2016. She specializes in training neural networks (which underpin most of today’s machine learning techniques) on quirky data sets such as compilations of knitting instructions, ice cream flavors, and names of paint colors. Then she asks the neural net to generate its own contributions to these categories—and hilarity ensues. AI is not likely to disrupt the paint industry with names like “Ronching Blue,” “Dorkwood,” and “Turdly.”

Shane’s antics have a serious purpose. She aims to illustrate the serious limitations of today’s AI, and to counteract the prevailing narrative that describes AI as well on its way to superintelligence and complete human domination. “The danger of AI is not that it’s too smart,” Shane writes in her new book, “but that it’s not smart enough.”

The book, which came out on Tuesday, is called You Look Like a Thing and I Love You. It takes its odd title from a list of AI-generated pick-up lines, all of which would at least get a person’s attention if shouted, preferably by a robot, in a crowded bar. Shane’s book is shot through with her trademark absurdist humor, but it also contains real explanations of machine learning concepts and techniques. It’s a painless way to take AI 101.

She spoke with IEEE Spectrum about the perils of placing too much trust in AI systems, the strange AI phenomenon of “giraffing,” and her next potential Halloween costume.

Janelle Shane on . . .

The un-delicious origin of her blog
“The narrower the problem, the smarter the AI will seem”
Why overestimating AI is dangerous
Giraffing!
Machine and human creativity

The un-delicious origin of her blog IEEE Spectrum: You studied electrical engineering as an undergrad, then got a master’s degree in physics. How did that lead to you becoming the comedian of AI?
Janelle Shane: I’ve been interested in machine learning since freshman year of college. During orientation at Michigan State, a professor who worked on evolutionary algorithms gave a talk about his work. It was full of the most interesting anecdotes–some of which I’ve used in my book. He told an anecdote about people setting up a machine learning algorithm to do lens design, and the algorithm did end up designing an optical system that works… except one of the lenses was 50 feet thick, because they didn’t specify that it couldn’t do that.
I started working in his lab on optics, doing ultra-short laser pulse work. I ended up doing a lot more optics than machine learning, but I always found it interesting. One day I came across a list of recipes that someone had generated using a neural net, and I thought it was hilarious and remembered why I thought machine learning was so cool. That was in 2016, ages ago in machine learning land.
Spectrum: So you decided to “establish weirdness as your goal” for your blog. What was the first weird experiment that you blogged about?
Shane: It was generating cookbook recipes. The neural net came up with ingredients like: “Take ¼ pounds of bones or fresh bread.” That recipe started out: “Brown the salmon in oil, add creamed meat to the mixture.” It was making mistakes that showed the thing had no memory at all.
Spectrum: You say in the book that you can learn a lot about AI by giving it a task and watching it flail. What do you learn?
Shane: One thing you learn is how much it relies on surface appearances rather than deep understanding. With the recipes, for example: It got the structure of title, category, ingredients, instructions, yield at the end. But when you look more closely, it has instructions like “Fold the water and roll it into cubes.” So clearly this thing does not understand water, let alone the other things. It’s recognizing certain phrases that tend to occur, but it doesn’t have a concept that these recipes are describing something real. You start to realize how very narrow the algorithms in this world are. They only know exactly what we tell them in our data set.
BACK TO TOP↑ “The narrower the problem, the smarter the AI will seem” Spectrum: That makes me think of DeepMind’s AlphaGo, which was universally hailed as a triumph for AI. It can play the game of Go better than any human, but it doesn’t know what Go is. It doesn’t know that it’s playing a game.
Shane: It doesn’t know what a human is, or if it’s playing against a human or another program. That’s also a nice illustration of how well these algorithms do when they have a really narrow and well-defined problem.
The narrower the problem, the smarter the AI will seem. If it’s not just doing something repeatedly but instead has to understand something, coherence goes down. For example, take an algorithm that can generate images of objects. If the algorithm is restricted to birds, it could do a recognizable bird. If this same algorithm is asked to generate images of any animal, if its task is that broad, the bird it generates becomes an unrecognizable brown feathered smear against a green background.
Spectrum: That sounds… disturbing.
Shane: It’s disturbing in a weird amusing way. What’s really disturbing is the humans it generates. It hasn’t seen them enough times to have a good representation, so you end up with an amorphous, usually pale-faced thing with way too many orifices. If you asked it to generate an image of a person eating pizza, you’ll have blocks of pizza texture floating around. But if you give that image to an image-recognition algorithm that was trained on that same data set, it will say, “Oh yes, that’s a person eating pizza.”
BACK TO TOP↑ Why overestimating AI is dangerous Spectrum: Do you see it as your role to puncture the AI hype?
Shane: I do see it that way. Not a lot of people are bringing out this side of AI. When I first started posting my results, I’d get people saying, “I don’t understand, this is AI, shouldn’t it be better than this? Why doesn't it understand?” Many of the impressive examples of AI have a really narrow task, or they’ve been set up to hide how little understanding it has. There’s a motivation, especially among people selling products based on AI, to represent the AI as more competent and understanding than it actually is.
Spectrum: If people overestimate the abilities of AI, what risk does that pose?
Shane: I worry when I see people trusting AI with decisions it can’t handle, like hiring decisions or decisions about moderating content. These are really tough tasks for AI to do well on. There are going to be a lot of glitches. I see people saying, “The computer decided this so it must be unbiased, it must be objective.”

“If the algorithm’s task is to replicate human hiring decisions, it’s going to glom onto gender bias and race bias.”
—Janelle Shane, AI Weirdness blogger
That’s another thing I find myself highlighting in the work I’m doing. If the data includes bias, the algorithm will copy that bias. You can’t tell it not to be biased, because it doesn’t understand what bias is. I think that message is an important one for people to understand.
If there’s bias to be found, the algorithm is going to go after it. It’s like, “Thank goodness, finally a signal that’s reliable.” But for a tough problem like: Look at these resumes and decide who’s best for the job. If its task is to replicate human hiring decisions, it’s going to glom onto gender bias and race bias. There’s an example in the book of a hiring algorithm that Amazon was developing that discriminated against women, because the historical data it was trained on had that gender bias.
Spectrum: What are the other downsides of using AI systems that don’t really understand their tasks?
Shane: There is a risk in putting too much trust in AI and not examining its decisions. Another issue is that it can solve the wrong problems, without anyone realizing it. There have been a couple of cases in medicine. For example, there was an algorithm that was trained to recognize things like skin cancer. But instead of recognizing the actual skin condition, it latched onto signals like the markings a surgeon makes on the skin, or a ruler placed there for scale. It was treating those things as a sign of skin cancer. It’s another indication that these algorithms don’t understand what they’re looking at and what the goal really is.
BACK TO TOP↑ Giraffing Spectrum: In your blog, you often have neural nets generate names for things—such as ice cream flavors, paint colors, cats, mushrooms, and types of apples. How do you decide on topics?
Shane: Quite often it’s because someone has written in with an idea or a data set. They’ll say something like, “I’m the MIT librarian and I have a whole list of MIT thesis titles.” That one was delightful. Or they’ll say, “We are a high school robotics team, and we know where there’s a list of robotics team names.” It’s fun to peek into a different world. I have to be careful that I’m not making fun of the naming conventions in the field. But there’s a lot of humor simply in the neural net’s complete failure to understand. Puns in particular—it really struggles with puns.
Spectrum: Your blog is quite absurd, but it strikes me that machine learning is often absurd in itself. Can you explain the concept of giraffing?
Shane: This concept was originally introduced by [internet security expert] Melissa Elliott. She proposed this phrase as a way to describe the algorithms’ tendency to see giraffes way more often than would be likely in the real world. She posted a whole bunch of examples, like a photo of an empty field in which an image-recognition algorithm has confidently reported that there are giraffes. Why does it think giraffes are present so often when they’re actually really rare? Because they’re trained on data sets from online. People tend to say, “Hey look, a giraffe!” And then take a photo and share it. They don’t do that so often when they see an empty field with rocks.
There’s also a chatbot that has a delightful quirk. If you show it some photo and ask it how many giraffes are in the picture, it will always answer with some non zero number. This quirk comes from the way the training data was generated: These were questions asked and answered by humans online. People tended not to ask the question “How many giraffes are there?” when the answer was zero. So you can show it a picture of someone holding a Wii remote. If you ask it how many giraffes are in the picture, it will say two.
BACK TO TOP↑ Machine and human creativity Spectrum: AI can be absurd, and maybe also creative. But you make the point that AI art projects are really human-AI collaborations: Collecting the data set, training the algorithm, and curating the output are all artistic acts on the part of the human. Do you see your work as a human-AI art project?
Shane: Yes, I think there is artistic intent in my work; you could call it literary or visual. It’s not so interesting to just take a pre-trained algorithm that’s been trained on utilitarian data, and tell it to generate a bunch of stuff. Even if the algorithm isn’t one that I’ve trained myself, I think about, what is it doing that’s interesting, what kind of story can I tell around it, and what do I want to show people.

The Halloween costume algorithm “was able to draw on its knowledge of which words are related to suggest things like sexy barnacle.”
—Janelle Shane, AI Weirdness blogger
Spectrum: For the past three years you’ve been getting neural nets to generate ideas for Halloween costumes. As language models have gotten dramatically better over the past three years, are the costume suggestions getting less absurd?
Shane: Yes. Before I would get a lot more nonsense words. This time I got phrases that were related to real things in the data set. I don’t believe the training data had the words Flying Dutchman or barnacle. But it was able to draw on its knowledge of which words are related to suggest things like sexy barnacle and sexy Flying Dutchman.
Spectrum: This year, I saw on Twitter that someone made the gothy giraffe costume happen. Would you ever dress up for Halloween in a costume that the neural net suggested?
Shane: I think that would be fun. But there would be some challenges. I would love to go as the sexy Flying Dutchman. But my ambition may constrict me to do something more like a list of leg parts.
BACK TO TOP↑ Continue reading

Posted in Human Robots

#436184 Why People Demanded Privacy to Confide ...

This is part four of a six-part series on the history of natural language processing.

Between 1964 and 1966, Joseph Weizenbaum, a German American computer scientist at MIT’s artificial intelligence lab, developed the first-ever chatbot [PDF].

While there were already some rudimentary digital language generators in existence—programs that could spit out somewhat coherent lines of text—Weizenbaum’s program was the first designed explicitly for interactions with humans. The user could type in some statement or set of statements in their normal language, press enter, and receive a response from the machine. As Weizenbaum explained, his program made “certain kinds of natural-language conversation between man and computer possible.”

He named the program Eliza after Eliza Doolittle, the working-class hero of George Bernard Shaw’s Pygmalion who learns how to talk with an upper-class accent. The new Eliza was written for the 36-bit IBM 7094, an early transistorized mainframe computer, in a programming language that Weizenbaum developed called MAD-SLIP.

Because computer time was a valuable resource, Eliza could only be run via a time-sharing system; the user interacted with the program remotely via an electric typewriter and printer. When the user typed in a sentence and pressed enter, a message was sent to the mainframe computer. Eliza scanned the message for the presence of a keyword and used it in a new sentence to form a response that was sent back, printed out, and read by the user.

To encourage ongoing dialogue, Weizenbaum designed Eliza to simulate the type of conversational style used by a Rogerian psychoanalyst. The program would take something the user said and reflect it back with a question, as can be seen in the example below. (Notice how the program takes keywords like “boyfriend” and “depressed” and re-uses them.)

Human: Men are all alike.

Computer: IN WHAT WAY?

Human: They're always bugging us about something or other.

Computer: CAN YOU THINK OF A SPECIFIC EXAMPLE?

Human: Well, my boyfriend made me come here.

Computer: YOUR BOYFRIEND MADE YOU COME HERE

Human: He says I'm depressed much of the time.

Computer: I AM SORRY TO HEAR YOU ARE DEPRESSED

Human: It's true. I am unhappy.

Weizenbaum chose this mode of dialogue for Eliza because it gave the impression that the computer understood what was being said without having to offer anything new to the conversation. It created the illusion of comprehension and engagement in a mere 200 lines of code.

To test Eliza’s capacity to engage an interlocutor, Weizenbaum invited students and colleagues into his office and let them chat with the machine while he looked on. He noticed, with some concern, that during their brief interactions with Eliza, many users began forming emotional attachments to the algorithm. They would open up to the machine and confess problems they were facing in their lives and relationships.

During their brief interactions with Eliza, many users began forming emotional attachments to the algorithm.

Even more surprising was that this sense of intimacy persisted even after Weizenbaum described how the machine worked and explained that it didn’t really understand anything that was being said. Weizenbaum was most troubled when his secretary, who had watched him build the program from scratch over many months, insisted that he leave the room so she could talk to Eliza in private.

For Weizenbaum, this experiment with Eliza made him question an idea that Alan Turing had proposed in 1950 about machine intelligence. In his paper, entitled “Computing Machinery and Intelligence,” Turing suggested that if a computer could conduct a convincingly human conversation in text, one could assume it was intelligent—an idea that became the basis of the famous Turing Test.

But Eliza demonstrated that convincing communication between a human and a machine could take place even if comprehension only flowed from one side: The simulation of intelligence, rather than intelligence itself, was enough to fool people. Weizenbaum called this the Eliza effect, and believed it was a type of “delusional thinking” that humanity would collectively suffer from in the digital age. This insight was a profound shock for Weizenbaum, and one that came to define his intellectual trajectory over the next decade.

The simulation of intelligence, rather than intelligence itself, was enough to fool people.

In 1976, he published Computing Power and Human Reason: From Judgment to Calculation [PDF], which offered a long meditation on why people are willing to believe that a simple machine might be able to understand their complex human emotions.

In this book, he argues that the Eliza effect signifies a broader pathology afflicting “modern man.” In a world conquered by science, technology, and capitalism, people had grown accustomed to viewing themselves as isolated cogs in a large and uncaring machine. In such a diminished social world, Weizenbaum reasoned, people had grown so desperate for connection that they put aside their reason and judgment in order to believe that a program could care about their problems.

Weizenbaum spent the rest of his life developing this humanistic critique of artificial intelligence and digital technology. His mission was to remind people that their machines were not as smart as they were often said to be. And that even though it sometimes appeared as though they could talk, they were never really listening.

This is the fourth installment of a six-part series on the history of natural language processing. Last week’s post described Andrey Markov and Claude Shannon’s painstaking efforts to create statistical models of language for text generation. Come back next Monday for part five, “In 2016, Microsoft’s Racist Chatbot Revealed the Dangers of Conversation.”

You can also check out our prior series on the untold history of AI. Continue reading

Posted in Human Robots

#436180 Bipedal Robot Cassie Cal Learns to ...

There’s no particular reason why knowing how to juggle would be a useful skill for a robot. Despite this, robots are frequently taught how to juggle things. Blind robots can juggle, humanoid robots can juggle, and even drones can juggle. Why? Because juggling is hard, man! You have to think about a bunch of different things at once, and also do a bunch of different things at once, which this particular human at least finds to be overly stressful. While juggling may not stress robots out, it does require carefully coordinated sensing and computing and actuation, which means that it’s as good a task as any (and a more entertaining task than most) for testing the capabilities of your system.

UC Berkeley’s Cassie Cal robot, which consists of two legs and what could be called a torso if you were feeling charitable, has just learned to juggle by bouncing a ball on what would be her head if she had one of those. The idea is that if Cassie can juggle while balancing at the same time, she’ll be better able to do other things that require dynamic multitasking, too. And if that doesn’t work out, she’ll still be able to join the circus.

Cassie’s juggling is assisted by an external motion capture system that tracks the location of the ball, but otherwise everything is autonomous. Cassie is able to juggle the ball by leaning forwards and backwards, left and right, and moving up and down. She does this while maintaining her own balance, which is the whole point of this research—successfully executing two dynamic behaviors that may sometimes be at odds with one another. The end goal here is not to make a better juggling robot, but rather to explore dynamic multitasking, a skill that robots will need in order to be successful in human environments.

This work is from the Hybrid Robotics Lab at UC Berkeley, led by Koushil Sreenath, and is being done by Katherine Poggensee, Albert Li, Daniel Sotsaikich, Bike Zhang, and Prasanth Kotaru.

For a bit more detail, we spoke with Albert Li via email.

Image: UC Berkeley

UC Berkeley’s Cassie Cal getting ready to juggle.

IEEE Spectrum: What would be involved in getting Cassie to juggle without relying on motion capture?

Albert Li: Our motivation for starting off with motion capture was to first address the control challenge of juggling on a biped without worrying about implementing the perception. We actually do have a ball detector working on a camera, which would mean we wouldn’t have to rely on the motion capture system. However, we need to mount the camera in a way that it would provide the best upwards field of view, and we also have develop a reliable estimator. The estimator is particularly important because when the ball gets close enough to the camera, we actually can’t track the ball and have to assume our dynamic models describe its motion accurately enough until it bounces back up.

What keeps Cassie from juggling indefinitely?

There are a few factors that affect how long Cassie can sustain a juggle. While in simulation the paddle exhibits homogeneous properties like its stiffness and damping, in reality every surface has anisotropic contact properties. So, there are parts of the paddle which may be better for juggling than others (and importantly, react differently than modeled). These differences in contact are also exacerbated due to how the paddle is cantilevered when mounted on Cassie. When the ball hits these areas, it leads to a larger than expected error in a juggle. Due to the small size of the paddle, the ball may then just hit the paddle’s edge and end the juggling run. Over a very long run, this is a likely occurrence. Additionally, some large juggling errors could cause Cassie’s feet to slip slightly, which ends up changing the stable standing position over time. Since this version of the controller assumes Cassie is stationary, this change in position eventually leads to poor juggles and failure.

Would Cassie be able to juggle while walking (or hovershoe-ing)?

Walking (and hovershoe-ing) while juggling is a far more challenging problem and is certainly a goal for future research. Some of these challenges include getting the paddle to precise poses to juggle the ball while also moving to avoid any destabilizing effects of stepping incorrectly. The number of juggles per step of walking could also vary and make the mathematics of the problem more challenging. The controller goal is also more involved. While the current goal of the juggling controller is to juggle the ball to a static apex position, with a walking juggling controller, we may instead want to hit the ball forwards and also walk forwards to bounce it, juggle the ball along a particular path, etc. Solving such challenges would be the main thrusts of the follow-up research.

Can you give an example of a practical task that would be made possible by using a controller like this?

Studying juggling means studying contact behavior and leveraging our models of it to achieve a known objective. Juggling could also be used to study predictable post-contact flight behavior. Consider the scenario where a robot is attempting to make a catch, but fails, letting the ball to bounce off of its hand, and then recovering the catch. This behavior could also be intentional: It is often easier to first execute a bounce to direct the target and then perform a subsequent action. For example, volleyball players could in principle directly hit a spiked ball back, but almost always bump the ball back up and then return it.

Even beyond this motivating example, the kinds of models we employ to get juggling working are more generally applicable to any task that involves contact, which could include tasks besides bouncing like sliding and rolling. For example, clearing space on a desk by pushing objects to the side may be preferable than individually manipulating each and every object on it.

You mention collaborative juggling or juggling multiple balls—is that something you’ve tried yet? Can you talk a bit more about what you’re working on next?

We haven’t yet started working on collaborative or multi-ball juggling, but that’s also a goal for future work. Juggling multiple balls statically is probably the most reasonable next goal, but presents additional challenges. For instance, you have to encode a notion of juggling urgency (if the second ball isn’t hit hard enough, you have less time to get the first ball up before you get back to the second one).

On the other hand, collaborative human-robot juggling requires a more advanced decision-making framework. To get robust multi-agent juggling, the robot will need to employ some sort of probabilistic model of the expected human behavior (are they likely to move somewhere? Are they trying to catch the ball high or low? Is it safe to hit the ball back?). In general, developing such human models is difficult since humans are fairly unpredictable and often don’t exhibit rational behavior. This will be a focus of future work.

[ Hybrid Robotics Lab ] Continue reading

Posted in Human Robots

#436123 A Path Towards Reasonable Autonomous ...

Editor’s Note: The debate on autonomous weapons systems has been escalating over the past several years as the underlying technologies evolve to the point where their deployment in a military context seems inevitable. IEEE Spectrum has published a variety of perspectives on this issue. In summary, while there is a compelling argument to be made that autonomous weapons are inherently unethical and should be banned, there is also a compelling argument to be made that autonomous weapons could potentially make conflicts less harmful, especially to non-combatants. Despite an increasing amount of international attention (including from the United Nations), progress towards consensus, much less regulatory action, has been slow. The following workshop paper on autonomous weapons systems policy is remarkable because it was authored by a group of experts with very different (and in some cases divergent) views on the issue. Even so, they were able to reach consensus on a roadmap that all agreed was worth considering. It’s collaborations like this that could be the best way to establish a reasonable path forward on such a contentious issue, and with the permission of the authors, we’re excited to be able to share this paper (originally posted on Georgia Tech’s Mobile Robot Lab website) with you in its entirety.

Autonomous Weapon Systems: A Roadmapping Exercise
Over the past several years, there has been growing awareness and discussion surrounding the possibility of future lethal autonomous weapon systems that could fundamentally alter humanity’s relationship with violence in war. Lethal autonomous weapons present a host of legal, ethical, moral, and strategic challenges. At the same time, artificial intelligence (AI) technology could be used in ways that improve compliance with the laws of war and reduce non-combatant harm. Since 2014, states have come together annually at the United Nations to discuss lethal autonomous weapons systems1. Additionally, a growing number of individuals and non-governmental organizations have become active in discussions surrounding autonomous weapons, contributing to a rapidly expanding intellectual field working to better understand these issues. While a wide range of regulatory options have been proposed for dealing with the challenge of lethal autonomous weapons, ranging from a preemptive, legally binding international treaty to reinforcing compliance with existing laws of war, there is as yet no international consensus on a way forward.

The lack of an international policy consensus, whether codified in a formal document or otherwise, poses real risks. States could fall victim to a security dilemma in which they deploy untested or unsafe weapons that pose risks to civilians or international stability. Widespread proliferation could enable illicit uses by terrorists, criminals, or rogue states. Alternatively, a lack of guidance on which uses of autonomy are acceptable could stifle valuable research that could reduce the risk of non-combatant harm.

International debate thus far has predominantly centered around whether or not states should adopt a preemptive, legally-binding treaty that would ban lethal autonomous weapons before they can be built. Some of the authors of this document have called for such a treaty and would heartily support it, if states were to adopt it. Other authors of this document have argued an overly expansive treaty would foreclose the possibility of using AI to mitigate civilian harm. Options for international action are not binary, however, and there are a range of policy options that states should consider between adopting a comprehensive treaty or doing nothing.

The purpose of this paper is to explore the possibility of a middle road. If a roadmap could garner sufficient stakeholder support to have significant beneficial impact, then what elements could it contain? The exercise whose results are presented below was not to identify recommendations that the authors each prefer individually (the authors hold a broad spectrum of views), but instead to identify those components of a roadmap that the authors are all willing to entertain2. We, the authors, invite policymakers to consider these components as they weigh possible actions to address concerns surrounding autonomous weapons3.

Summary of Issues Surrounding Autonomous Weapons

There are a variety of issues that autonomous weapons raise, which might lend themselves to different approaches. A non-exhaustive list of issues includes:

The potential for beneficial uses of AI and autonomy that could improve precision and reliability in the use of force and reduce non-combatant harm.
Uncertainty about the path of future technology and the likelihood of autonomous weapons being used in compliance with the laws of war, or international humanitarian law (IHL), in different settings and on various timelines.
A desire for some degree of human involvement in the use of force. This has been expressed repeatedly in UN discussions on lethal autonomous weapon systems in different ways.
Particular risks surrounding lethal autonomous weapons specifically targeting personnel as opposed to vehicles or materiel.
Risks regarding international stability.
Risk of proliferation to terrorists, criminals, or rogue states.
Risk that autonomous systems that have been verified to be acceptable can be made unacceptable through software changes.
The potential for autonomous weapons to be used as scalable weapons enabling a small number of individuals to inflict very large-scale casualties at low cost, either intentionally or accidentally.

Summary of Components

A time-limited moratorium on the development, deployment, transfer, and use of anti-personnel lethal autonomous weapon systems4. Such a moratorium could include exceptions for certain classes of weapons.
Define guiding principles for human involvement in the use of force.
Develop protocols and/or technological means to mitigate the risk of unintentional escalation due to autonomous systems.
Develop strategies for preventing proliferation to illicit uses, such as by criminals, terrorists, or rogue states.
Conduct research to improve technologies and human-machine systems to reduce non-combatant harm and ensure IHL compliance in the use of future weapons.

Component 1:

States should consider adopting a five-year, renewable moratorium on the development, deployment, transfer, and use of anti-personnel lethal autonomous weapon systems. Anti-personnel lethal autonomous weapon systems are defined as weapons systems that, once activated, can select and engage dismounted human targets without further intervention by a human operator, possibly excluding systems such as:

Fixed-point defensive systems with human supervisory control to defend human-occupied bases or installations
Limited, proportional, automated counter-fire systems that return fire in order to provide immediate, local defense of humans
Time-limited pursuit deterrent munitions or systems
Autonomous weapon systems with size above a specified explosive weight limit that select as targets hand-held weapons, such as rifles, machine guns, anti-tank weapons, or man-portable air defense systems, provided there is adequate protection for non-combatants and ensuring IHL compliance5

The moratorium would not apply to:

Anti-vehicle or anti-materiel weapons
Non-lethal anti-personnel weapons
Research on ways of improving autonomous weapon technology to reduce non-combatant harm in future anti-personnel lethal autonomous weapon systems
Weapons that find, track, and engage specific individuals whom a human has decided should be engaged within a limited predetermined period of time and geographic region

Motivation:

This moratorium would pause development and deployment of anti-personnel lethal autonomous weapons systems to allow states to better understand the systemic risks of their use and to perform research that improves their safety, understandability, and effectiveness. Particular objectives could be to:

ensure that, prior to deployment, anti-personnel lethal autonomous weapons can be used in ways that are equal to or outperform humans in their compliance with IHL (other conditions may also apply prior to deployment being acceptable);
lay the groundwork for a potentially legally binding diplomatic instrument; and
decrease the geopolitical pressure on countries to deploy anti-personnel lethal autonomous weapons before they are reliable and well-understood.

Compliance Verification:

As part of a moratorium, states could consider various approaches to compliance verification. Potential approaches include:

Developing an industry cooperation regime analogous to that mandated under the Chemical Weapons Convention, whereby manufacturers must know their customers and report suspicious purchases of significant quantities of items such as fixed-wing drones, quadcopters, and other weaponizable robots.
Encouraging states to declare inventories of autonomous weapons for the purposes of transparency and confidence-building.
Facilitating scientific exchanges and military-to-military contacts to increase trust, transparency, and mutual understanding on topics such as compliance verification and safe operation of autonomous systems.
Designing control systems to require operator identity authentication and unalterable records of operation; enabling post-hoc compliance checks in case of plausible evidence of non-compliant autonomous weapon attacks.
Relating the quantity of weapons to corresponding capacities for human-in-the-loop operation of those weapons.
Designing weapons with air-gapped firing authorization circuits that are connected to the remote human operator but not to the on-board automated control system.
More generally, avoiding weapon designs that enable conversion from compliant to non-compliant categories or missions solely by software updates.
Designing weapons with formal proofs of relevant properties—e.g., the property that the weapon is unable to initiate an attack without human authorization. Proofs can, in principle, be provided using cryptographic techniques that allow the proofs to be checked by a third party without revealing any details of the underlying software.
Facilitate access to (non-classified) AI resources (software, data, methods for ensuring safe operation) to all states that remain in compliance and participate in transparency activities.

Component 2:

Define and universalize guiding principles for human involvement in the use of force.

Humans, not machines, are legal and moral agents in military operations.
It is a human responsibility to ensure that any attack, including one involving autonomous weapons, complies with the laws of war.
Humans responsible for initiating an attack must have sufficient understanding of the weapons, the targets, the environment and the context for use to determine whether that particular attack is lawful.
The attack must be bounded in space, time, target class, and means of attack in order for the determination about the lawfulness of that attack to be meaningful.
Militaries must invest in training, education, doctrine, policies, system design, and human-machine interfaces to ensure that humans remain responsible for attacks.

Component 3:

Develop protocols and/or technological means to mitigate the risk of unintentional escalation due to autonomous systems.

Specific potential measures include:

Developing safe rules for autonomous system behavior when in proximity to adversarial forces to avoid unintentional escalation or signaling. Examples include:

No-first-fire policy, so that autonomous weapons do not initiate hostilities without explicit human authorization.
A human must always be responsible for providing the mission for an autonomous system.
Taking steps to clearly distinguish exercises, patrols, reconnaissance, or other peacetime military operations from attacks in order to limit the possibility of reactions from adversary autonomous systems, such as autonomous air or coastal defenses.

Developing resilient communications links to ensure recallability of autonomous systems. Additionally, militaries should refrain from jamming others’ ability to recall their autonomous systems in order to afford the possibility of human correction in the event of unauthorized behavior.

Component 4:

Develop strategies for preventing proliferation to illicit uses, such as by criminals, terrorists, or rogue states:

Targeted multilateral controls to prevent large-scale sale and transfer of weaponizable robots and related military-specific components for illicit use.
Employ measures to render weaponizable robots less harmful (e.g., geofencing; hard-wired kill switch; onboard control systems largely implemented in unalterable, non-reprogrammable hardware such as application-specific integrated circuits).

Component 5:

Conduct research to improve technologies and human-machine systems to reduce non-combatant harm and ensure IHL-compliance in the use of future weapons, including:

Strategies to promote human moral engagement in decisions about the use of force
Risk assessment for autonomous weapon systems, including the potential for large-scale effects, geopolitical destabilization, accidental escalation, increased instability due to uncertainty about the relative military balance of power, and lowering thresholds to initiating conflict and for violence within conflict
Methodologies for ensuring the reliability and security of autonomous weapon systems
New techniques for verification, validation, explainability, characterization of failure conditions, and behavioral specifications.

About the Authors (in alphabetical order)

Ronald Arkin directs the Mobile Robot Laboratory at Georgia Tech.

Leslie Kaelbling is co-director of the Learning and Intelligent Systems Group at MIT.

Stuart Russell is a professor of computer science and engineering at UC Berkeley.

Dorsa Sadigh is an assistant professor of computer science and of electrical engineering at Stanford.

Paul Scharre directs the Technology and National Security Program at the Center for a New American Security (CNAS).

Bart Selman is a professor of computer science at Cornell.

Toby Walsh is a professor of artificial intelligence at the University of New South Wales (UNSW) Sydney.

The authors would like to thank Max Tegmark for organizing the three-day meeting from which this document was produced.

1 Autonomous Weapons System (AWS): A weapon system that, once activated, can select and engage targets without further intervention by a human operator. BACK TO TEXT↑

2 There is no implication that some authors would not personally support stronger recommendations. BACK TO TEXT↑

3 For ease of use, this working paper will frequently shorten “autonomous weapon system” to “autonomous weapon.” The terms should be treated as synonymous, with the understanding that “weapon” refers to the entire system: sensor, decision-making element, and munition. BACK TO TEXT↑

4 Anti-personnel lethal autonomous weapon system: A weapon system that, once activated, can select and engage dismounted human targets with lethal force and without further intervention by a human operator. BACK TO TEXT↑

5 The authors are not unanimous about this item because of concerns about ease of repurposing for mass-casualty missions targeting unarmed humans. The purpose of the lower limit on explosive payload weight would be to minimize the risk of such repurposing. There is precedent for using explosive weight limit as a mechanism of delineating between anti-personnel and anti-materiel weapons, such as the 1868 St. Petersburg Declaration Renouncing the Use, in Time of War, of Explosive Projectiles Under 400 Grammes Weight. BACK TO TEXT↑ Continue reading

Posted in Human Robots