Containment for AI

The last two years have seen startling advances in artificial intelligence. The next few years promise far more, with larger and more efficient models, capable of real creativity and complicated planning, likely to emerge. The potential positives are astonishing, including heightened business productivity, cheaper and more effective health care, scientific discoveries, and education programs tailored to every child’s needs. But the risks are also colossal. These include the proliferation of disinformation, as well as job losses, and the likelihood that bad actors will seek to use the new technology to sow havoc.

This technology will proliferate rapidly. That means that over the next ten years, grappling with AI’s inbuilt tendency toward uncontrolled spread will become a generational challenge. It will, accordingly, require a generational response akin to what the West mobilized in the early days of the Cold War. At that time, the American diplomat George F. Kennan talked about containing the Soviet Union by using hard power and economic and cultural pressure to ensure that the Soviets were kept behind their borders and the democratic world was not overwhelmed. Today’s challenge requires a similarly broad and ambitious program, in this case to keep AI in check and societies in control. It will be, like Kennan’s, an effort based on laws and treaties. It will also necessitate, however, a massive global movement and changes to the culture of technology companies. This modern form of containment will be needed not only to manage AI and prevent it from creating catastrophe but also to ensure that it becomes one of the most extraordinarily beneficial inventions in human history.

THE TIDE ALWAYS COMES IN

Across the sweep of human history, there is a single, seemingly immutable law: every foundational technology ever invented—from pickaxes to plows, pottery to photography, phones to planes—will become cheaper and easier to use. It will spread far and wide. The ecosystem of invention defaults to expansion. And people, who always drive this process, are Homo technologicus, the innately technological species.

Consider the printing press. In the 1440s, after Johannes Gutenberg invented it, there was only a single example in Europe: his original in Mainz, Germany. But just 50 years later, there were around 1,000 presses spread across the continent. The results were extraordinary. In the Middle Ages, major countries including France and Italy each produced a few hundred thousand manuscripts per century. A hundred years later, they were producing around 400,000 books each per year, and the pace was increasing. In the seventeenth century alone, European countries printed 500 million books.

The same trend was seen with the internal combustion engine. This was a tricky invention that took over 100 years to perfect. Eventually, by the 1870s, there were only a few working examples in German workshops. The technology was still nascent, limited in number, and utterly marginal. Eight years after he invented the first practical automobile in 1885, the German engineer Carl Benz had sold just 69 cars. But a little over 100 years later, there were over two billion internal combustion engines of every conceivable shape and size, powering everything from lawnmowers to container ships.

Some technologies, particularly nuclear weapons, may appear to buck this trend. After all, 80 years on from their creation, they have been used only twice, by the United States in 1945, and arsenals are well down from their 1980s highs. Although there is some truth to this counterargument, however, it ignores the thousands of warheads still deployed around the world, the constant pressure of new states looking to build them, and the hair-raising litany of accidents and close calls that, from the beginning, have been a regular and, for obvious reasons, underreported feature of these weapons. From the drama of the Cuban missile crisis in 1962 to the disappearance of nuclear materials from a U.S. government employee’s car in 2017, nuclear weapons have never been truly contained despite the avoidance of outright catastrophe. If such technologies as nuclear weaponry are an exception to the rule of technological spread, they are at best a very partial and uneasy exception.

THE IMPENDING DELUGE

It is inevitable that AI will follow the trajectory of the hand axe, the printing press, the internal combustion engine, and the Internet. It, too, will be everywhere, and it will constantly improve. It is happening already. In just a few years, cutting-edge models have gone from using millions of parameters, or variables adjusted in training, to trillions, indicating the ever-increasing complexity of these systems. Over the last decade, the amount of computation used to train large AI models has increased by nine orders of magnitude. Moore’s law, which holds that computing power doubles every two years, predicted exponential increases in what computers can do. But progress has been even faster in AI, with the trends of lower costs and improving capability ascending on a curve beyond anything seen with a technology before. The results are visible in well-known AI products but are also proving transformative under the surface of the digital world, powering software, organizing warehouses, operating medical equipment, driving vehicles, and managing power grids.

As the next phase of AI develops, a powerful generation of autonomous AI agents capable of achieving real-world goals will emerge. Although this is often called artificial general intelligence, I prefer the term “artificial capable intelligence,” or ACI, which is a stage before full AGI, in which AI can nonetheless achieve a range of tasks autonomously. This technology can accomplish complex activities on humans’ behalf, from organizing a birthday party to completing the weekly shop, in addition to something as consequential as setting up and running an entire business line. This will be a seismic step for the technology, with transformative implications for the nature of power and the world economy. It can be expected to proliferate rapidly and irreversibly.

An ACI in everyone’s pocket will result in colossal increases in economic growth, as the most significant productivity enhancer seen in generations becomes as ubiquitous as electricity. ACI will revolutionize fields including health care, education, and energy generation. Above all, it will give people the chance to achieve what they want in life. There is a fair amount of doomsaying about AI at the moment, but amid well-justified concerns, it is important to keep in mind the many upsides of AI. This is particularly the case for ACI, which has the potential to give everyone access to the world’s best assistant, chief of staff, lawyer, doctor, and all-around A-team.

Yet the downsides cannot be ignored. For a start, AI will unleash a series of new dangers. Perhaps the most serious of these will be new forms of misinformation and disinformation. Just a few simple language commands can now produce images—and, increasingly, videos—of staggering fidelity. When hostile governments, fringe political parties, and lone actors can create and broadcast material that is indistinguishable from reality, they will be able to sow chaos, and the verification tools designed to stop them may well be outpaced by the generative systems. Deepfakes caused turmoil in the stock market last year when a concocted image of the Pentagon on fire caused a momentary but noticeable dip in indexes, and they are likely to feature heavily in the current U.S. election race.

Many other problems can be expected to result from the global advance of AI. Automation threatens to disrupt the labor market, and the potential for immense cyberattacks is growing. Once powerful new forms of AI spread, all the good and all the bad will be available at every level of society: in the hands of CEOs, street vendors, and terrorists alike.

STOPPING THE SPREAD

Most people’s attention has correctly focused on the social and ethical implications of this change. But this discussion often neglects to consider technology’s tendency to penetrate every layer of civilization, and it is this that requires drastic action. It is technology’s tendency to spread fast, far, and wide that demands that AI must be contained, both in its proliferation and in its negative impacts, when the latter do occur. Containment is a daunting task, given the history and the trajectory of innovation, but it is the only answer—however difficult—to how humanity should manage the fastest rollout of the most powerful new technology in history.

Containment in this sense encompasses regulation, better technical safety, new governance and ownership models, and new modes of accountability and transparency. All are necessary—but not sufficient—to assure safer technology. Containment must combine cutting-edge engineering with ethical values that will inform government regulation. The goal should be to create a set of interlinked and mutually reinforcing technical, cultural, legal, and political mechanisms for maintaining societal control of AI. Governments must contain what would have once been centuries or millennia of technological change but is now unfolding in a matter of years or even months. Containment is, in theory, an answer to the inescapability of proliferation, capable both of checking it and addressing its consequences.

This is not containment in the geopolitical sense, harking back to Kennan’s doctrines. Nor is it a matter of putting AI into a sealed box, although some technologies—rogue AI malware and an engineered pathogen, in particular—need just that. Nor is containment of AI competitive, in the sense of seeking to fight some Soviet Red Menace. It does resemble Kennan’s approach in that the policy framework must operate across all dimensions. But containing technology is a much more fundamental program than what Kennan envisioned, seeking a balance of power not between competing actors but between humans and their tools. What it seeks is not to stop the technology but keep it safe and controlled.

Most people rightly argue that regulation is necessary, and there is a tendency to believe that it is enough. It is not. Containment in practice must work on every level at which the technology operates. It therefore needs not only proactive and well-informed lawmakers and bureaucrats but also technologists and business executives. It needs diplomats and leaders to cooperate internationally to build bridges and address gaps. It needs consumers and citizens everywhere to demand better from technology, and ensure that it remains focused on their interests. It needs them to agitate for and expect responsible technology, just as growing demand for green energy and environmentally friendly products has spurred corporates and governments into action.

STEERING WITHOUT A MAP

Containment will require hard technical questions to be answered by international treaties and mass global movements alike. It must encompass work on AI safety, as well as the audit mechanisms needed to monitor and enforce compliance. The companies behind AI will be critical to this effort and will need to think carefully about how to align their own incentives with government regulation. Yet containing AI will not be the sole responsibility of those building its next generation. Nor will it rest entirely on national leaders. Rather, all of those who will be affected by it (that is, everyone) will be critical to creating momentum behind this effort. Containment offers a policy blend capable of working from the fine-grained details of an AI model out to huge public programs that could mitigate vast job destruction.

Collectively, this project may prove equal to this moment and capable of counteracting the many risks that AI poses. The cumulative effect of these measures—which must include licensing regimes, the staffing of a generation of companies with critics, and the creation of inbuilt mechanisms to guarantee access to advanced systems—is to keep humanity in the driver’s seat of this epochal series of changes, and capable, at the limit, of saying no. None of these steps will be easy. After all, uncontrolled proliferation has been the default throughout human history. Containment should therefore be seen not as the final answer to all technology’s problems but rather, the first critical step.