The Age of A.I. and Our Human Future

by Henry Kissinger

In 2020, American AI start-ups raised almost $38 billion in funding. Their Asian counterparts raised $25 billion. And their European counterparts raised $8 billion.

The outcome will be the alteration of human identity and the human experience of reality at levels not experienced since the dawn of the modern age.

Even after the antibiotic was discovered, humans could not articulate precisely why it worked. The AI did not just process data more quickly than humanly possible; it also detected aspects of reality humans have not detected, or perhaps cannot detect.

models like GPT-3 generate possible responses to various inputs (and thus are called generative models).

The advent of AI obliges us to confront whether there is a form of logic that humans have not achieved or cannot achieve, exploring aspects of reality we have never known and may never directly know.

When a human-designed software program, carrying out an objective assigned by its programmers—correcting bugs in software or refining the mechanisms of self-driving vehicles—learns and applies a model that no human recognizes or could understand, are we advancing toward knowledge? Or is knowledge receding from us?

made in three primary ways: by humans (which is familiar), by machines (which is becoming familiar), and by collaboration between humans and machines (which is not only unfamiliar but also unprecedented). AI is also in the process of transforming machines—which, until now, have been our tools—into our partners. We will begin to give AI fewer specific instructions about how exactly to achieve the goals we assign it. Much more frequently, we will present AI with ambiguous goals and ask: “How, based on your conclusions, should we proceed?”

In such cases, new divides will appear within and between societies—between those who adopt the new technology and those who opt out or lack the means to develop or acquire some of its applications. When various groups or nations adopt differing concepts or applications of AI, their experiences of reality may diverge in ways that are difficult to predict or bridge. As societies develop their own human-machine partnerships—with varying goals, different training models, and potentially incompatible operational and moral limits with respect to AI—they may devolve into rivalry, technical incompatibility, and ever greater mutual incomprehension. Technology that was initially believed to be an instrument for the transcendence of national differences and the dispersal of objective truth may, in time, become the method by which civilizations and individuals diverge into different and mutually unintelligible realities.

But there is a difference between choosing from a range of options and taking an action—in this case, making a purchase; in other cases, adopting a political or philosophical position or ideology—without ever knowing what the initial range of possibilities or implications was, entrusting a machine to preemptively shape the options.

The sixteenth and seventeenth centuries witnessed such rapid progress—with astounding discoveries in mathematics, astronomy, and the natural sciences—that it led to a sort of philosophical disorientation. Given that church doctrine still officially defined the limits of permissible intellectual explorations during this period, these advances produced breakthroughs of considerable daring. Copernicus’s vision of a heliocentric system, Newton’s laws of motion, van Leeuwenhoek’s cataloging of a living microscopic world—these and other developments led to the general sentiment that new layers of reality were being unveiled. The outcome was incongruence: societies remained united in their monotheism but were divided by competing interpretations and explorations of reality. They needed a concept—indeed, a philosophy—to guide their quest to understand the world and their role in it.

In a sense, the West had returned to many of the fundamental questions with which the ancient Greeks had wrestled: What is reality? What are people seeking to know and experience, and how will they know when they encounter it? Can humans perceive reality itself as opposed to its reflections? If so, how? What does it mean to be and to know? Unencumbered by tradition—or at least believing they were justified in interpreting it anew—scholars and philosophers once again investigated these questions. The minds that set out on this journey were willing to walk a precarious path, risking the apparent certainties of their cultural traditions and their established conceptions of reality.

For the following two hundred years, Kant’s essential distinction between the thing-in-itself and the unavoidably filtered world we experience hardly seemed to matter. While the human mind might present an imperfect picture of reality, it was the only picture available. What the structures of the human mind barred from view would, presumably, be barred forever—or would inspire faith and consciousness of the infinite. Without any alternative mechanism for accessing reality, it seemed that humanity’s blind spots would remain hidden. Whether human perception and reason ought to be the definitive measure of things, lacking an alternative, for a time, they became so. But AI is beginning to provide an alternative means of accessing—and thus understanding—reality.

While the technological achievements of the age of reason have been significant, until recently they had remained sporadic enough to be reconciled with tradition. Innovations have been characterized as extensions of previous practices: films were moving photographs, telephones were conversations across space, and automobiles were rapidly moving carriages in which horses were replaced by engines measured by their “horsepower.

But we have reached a tipping point: we can no longer conceive of some of our innovations as extensions of that which we already know. By compressing the time frame in which technology alters the experience of life, the revolution of digitization and the advancement of AI have produced phenomena that are truly new, not simply more powerful or efficient versions of things past.

When information is contextualized, it becomes knowledge. When knowledge compels convictions, it becomes wisdom. Yet the internet inundates users with the opinions of thousands, even millions, of other users, depriving them of the solitude required for sustained reflection that, historically, has led to the development of convictions. As solitude diminishes, so, too, does fortitude—not only to develop convictions but also to be faithful to them, particularly when they require the traversing of novel, and thus often lonely, roads. Only convictions—in combination with wisdom—enable people to access and explore new horizons.

The introduction of AI—which completes the sentence we are texting, identifies the book or store we are seeking, and “intuits” articles and entertainment we might enjoy based on prior behavior—has often seemed more mundane than revolutionary. But as it is being applied to more elements of our lives, it is altering the role that our minds have traditionally played in shaping, ordering, and assessing our choices and actions.

AIs are imprecise, dynamic, emergent, and capable of “learning.” AIs “learn” by consuming data, then drawing observations and conclusions based on the data. While previous systems required exact inputs and outputs, AIs with imprecise function require neither. These AIs translate texts not by swapping individual words but by identifying and employing idiomatic phrases and patterns. Likewise, such AI is considered dynamic because it evolves in response to changing circumstances and emergent because it can identify solutions that are novel to humans. In machinery, these four qualities are revolutionary.

While in fields that do use precise characterization—chess, algebraic manipulation, and business process automation—AI made great advances, in other fields, like language translation and visual object recognition, inherent ambiguity brought progress to a halt. The challenges of visual object recognition illustrate the shortcomings of these early programs. Even young children can identify images with ease. But early generations of AI could not. Programmers initially attempted to distill an object’s distinguishing characteristics into a symbolic representation.

Because these formalistic and inflexible systems were only successful in domains whose tasks could be achieved by encoding clear rules, from the late 1980s through the 1990s, the field entered a period referred to as “AI winter.”

In short, a conceptual shift occurred: we went from attempting to encode human-distilled insights into machines to delegating the learning process itself to the machines. In the 1990s, a set of renegade researchers set aside many of the earlier era’s assumptions, shifting their focus to machine learning. While machine learning dated to the 1950s, new advances enabled practical applications.

AI is imprecise in that it does not require a predefined relationship between a property and an effect to identify a partial relationship. It can, for example, select highly likely candidates from a larger set of possible candidates. This capability captures one of the vital elements of modern AI. Using machine learning to create and adjust models based on real-world feedback, modern AI can approximate outcomes and analyze ambiguities that would have stymied classical algorithms.

Unlike humans, most AIs cannot simultaneously train and execute. Rather, they divide their effort into two steps: training and inference. During the training phase, the AI’s quality measurement and improvement algorithms evaluate and amend its model to obtain quality results. In the case of halicin, this was the phase when the AI identified relationships between molecular structures and antibiotic effects based on the training-set data. Then, in the inference phase, researchers tasked the AI with identifying antibiotics that its newly trained model predicted would have a strong antibiotic effect. The AI, then, did not reach conclusions by reasoning as humans reason; it reached conclusions by applying the model it developed.

As of this writing, three forms of machine learning are noteworthy: supervised learning, unsupervised learning, and reinforcement learning.

This technique is called supervised learning because the AI developers used a dataset containing example inputs (in this case, molecular structures) that were individually labeled according to the desired output or result (in this case, effectiveness as an antibiotic).

In situations where developers have only troves of data, however, they can employ unsupervised learning to extract potentially useful insights.

unsupervised learning allows AIs to identify patterns or anomalies without having any information regarding outcomes. In unsupervised learning, the training data contains only inputs. Then programmers task the learning algorithm with producing groupings based on some specified weight of measuring the degree of similarity. For example, streaming video services such as Netflix use algorithms to identify clusters of customers with similar viewing habits in order to recommend additional streaming to those customers.

AIs trained through unsupervised learning can identify patterns that humans might miss because of the pattern’s subtlety, the scale of the data, or both.

In reinforcement learning, AI is not passive, identifying relationships within data. Instead, AI is an “agent” in a controlled environment, observing and recording responses to its actions. Generally these are simulated, simplified versions of reality lacking real-world complexities.

As a result, directing an AI to train itself in an artificial environment is, in general, insufficient to produce the best performance. Feedback is required. Providing that feedback is the task of the reward function, indicating to the AI how successful its approach was.

In conventional approaches, developers trained AI using texts and their preexisting translations—after all, they had the requisite level of correspondence between one language and another. Yet this approach greatly limited the amount of training data as well as the types of text available: although government texts and bestselling books are frequently translated, periodicals, social media, websites, and other informal writings generally are not. Rather than restricting AIs to training on carefully translated texts, researchers simply supplied articles and other texts in various languages covering a single topic, declining to bother with detailed translations between them. This process, training AIs on roughly matching—but untranslated—bodies of text, is the parallel corpora technique. It is akin to stepping from an introductory language class into a total immersion program. The training is less precise, but the volume of available data is much bigger:

A common training technique for the creation of generative AI pits two networks with complementary learning objectives against each other. Such networks are referred to as generative adversarial networks or GANs. The objective of the generator network is to create potential outputs, while the objective of the discriminator network is to prevent poor outputs from being generated. By analogy, one can think of the generator as being tasked with brainstorming and the discriminator as being tasked with assessing which ideas are relevant and realistic. In the training phase, the generator and discriminator are trained in alternation, holding the generator fixed to train the discriminator and vice versa.

AIs trained with GANs may suggest sentence completions when drafting emails or permit search engines to complete partial queries. More dramatically, GANs may be used to develop AIs that can fill in the details of sketched code—in other words, programmers may soon be able to outline a desired program and then turn that outline over to an AI for completion.

Unlike earlier generations of AI, in which people distilled a society’s understanding of reality in a program’s code, contemporary machine-learning AIs largely model reality on their own. While developers may examine the results generated by their AIs, the AIs do not “explain” how or what they learned in human terms. Nor can developers ask an AI to characterize what it has learned. Much as with humans, one cannot really know what has been learned and why (though humans can often offer explanations or justifications that, as of this writing, AI cannot).

Since machine learning will drive AI for the foreseeable future, humans will remain unaware of what an AI is learning and how it knows what it has learned. While this may be disconcerting, it should not be: human learning is often similarly opaque. Artists and athletes, writers and mechanics, parents and children—indeed, all humans—often act on the basis of intuition and thus are unable to articulate what or how they learned. To cope with this opacity, societies have developed myriad professional certification programs, regulations, and laws. Similar techniques should be applied to AIs; for example, societies could permit an AI to be employed only after its creators demonstrate its reliability through testing processes. Developing professional certification, compliance monitoring, and oversight programs for AI—and the auditing expertise their execution will require—will be a crucial societal project.

In industry, pre-use testing exists on a spectrum. App developers often rush programs to market, correcting flaws in real time, while aerospace companies do the opposite: test their jets religiously before a single customer ever sets foot on board. The variance in these regimes depends on several factors—above all, the inherent riskiness of the activity. As AI deployments multiply, the same factors—inherent riskiness, regulatory oversight, market forces—will likely distribute them across the same spectrum, with AIs that drive cars being subjected to significantly greater oversight than AIs that power network platforms for entertainment and connection, such as TikTok.

Most AIs, though, train in a phase distinct from the operational phase: their learned models—the parameters of their neural networks—are static when they exit training. Because an AI’s evolution halts after training, humans can assess its capacities without fear that it will develop unexpected, undesired behaviors after it completes its tests. In other words, when the algorithm is fixed, a self-driving car trained to stop at red lights cannot suddenly “decide” to start running them. This property makes comprehensive testing and certifications possible—

As of this writing, AI is constrained by its code in three ways. First, the code sets the parameters of the AI’s possible actions.

Second, AI is constrained by its objective function, which defines and assigns what it is to optimize.

Finally and most obviously, AI can only process inputs that it is designed to recognize and analyze. Without human intervention in the form of an auxiliary program, a translation AI cannot evaluate images—the data would appear nonsensical to it.

It is reasonable to expect that over time, AI will progress at least as fast as computing power has, yielding a millionfold increase in fifteen to twenty years. Such progress will allow the creation of neural networks that, in scale, are equal to the human brain. As of this writing, generative transformers have the largest networks. GPT-3 has about 1011 such weights. But recently, the state-funded Beijing Academy of Sciences announced a generative language model with 10 times as many weights as GPT-3. This is still 104 times fewer than estimates of the human brain’s synapses. But if advances proceed at the rate of doubling every two years, this gap could close in less than a decade. Of course, scale does not translate directly to intelligence.

For some, the inclination will be to entrust the task to a technical process that seems free from human bias and partiality—an AI with an objective function to identify and arrest the flow of disinformation and falsity. But what of the content that is never viewed by the public? When the prominence or diffusion of a message is so curtailed that its existence is, in effect, negated, we have reached a state of censorship. If antidisinformation AI makes a mistake, suppressing content that is not malign disinformation but in fact authentic, how do we identify it? Can we know enough, and in time, to correct it? Alternatively, do we have the right to read, or even a legitimate interest in reading, AI-generated “false” information? The power to train defensive AI against an objective (or subjective) standard of falsehood—and the ability, if any can be developed, to monitor that AI’s operations—would in itself become a function of importance and influence rivaling the roles traditionally held by government. Small differences in the design of an AI’s objective function, training parameters, and definitions of falsehood could lead to society-altering differences in outcome. These questions become all the more vital as network platforms use AI to provide their services to billions of people.

The emerging geopolitics of network platforms comprises a key new aspect of international strategy—and governments are not the only players. Governments may increasingly seek to limit the use or behavior of such systems or attempt to prevent them from edging out homegrown rivals in important regions, lest a competing society or economy gain a powerful influence over that country’s industrial, economic, or (more difficult to define) political and cultural development. Yet because governments generally do not create or operate these network platforms, the actions of inventors, corporations, and individual users will shape the field along with government restrictions or incentives, creating a strategic arena that is particularly dynamic and difficult to predict.

The United States has begun to view network platforms as an aspect of international strategy, restricting the domestic activities of some foreign platforms and restricting the export of some software and technology that could facilitate the growth of foreign competitors. At the same time, federal and state regulators have identified major domestic network platforms as targets for antitrust actions. In the near term, at least, this simultaneous drive for strategic preeminence and domestic multiplicity may push US development in conflicting directions.

That AI-enabled network platforms created by one society may function and evolve within another society and become inextricable from that country’s economy and national political discourse marks a fundamental departure from prior eras. Previously, sources of information and communication were typically local and national in scope—and maintained no independent ability to learn. Today, transportation network platforms created in one country could become the arteries and lifeblood of another country, as the platform learns which consumers need certain products and as it automates the logistics of provision. In effect, such network platforms could become critical economic infrastructure, giving the country of origin leverage over any country that relies on it.

Caught between governmental actions and concerns regarding their global status and user base, network platform operators will need to make decisions about the extent to which they become, in effect, a conglomeration of national and/or regional companies, potentially in several separate jurisdictions. Conversely, they may decide to conduct themselves as global companies independently pursuing their values, which may not align neatly with any particular government’s priorities.

CHAPTER 5 SECURITY AND WORLD ORDER

Whereas deterrence sought to prevent nuclear war by threatening it, arms control aimed to prevent nuclear war through the limitation or even abolition of the weapons (or categories of weapons) themselves. This approach was paired with nonproliferation: the concept, underpinned by an elaborate set of treaties, technical safeguards, and regulatory and other control mechanisms, that nuclear weapons and the knowledge and technology supporting their construction should be prevented from spreading beyond the nations that already possessed them.

To date, neither strategy has fully succeeded. Nor has either been pursued in earnest for the major new classes of weapons, cyber and AI, that have been invented in the post–Cold War era. Yet as entrants to the nuclear, cyber, and AI arenas multiply, the arms-control era still holds lessons worthy of consideration.

Nuclear non-use is not an inherently permanent achievement. It is a condition that must be secured by each successive generation of leaders adjusting the deployments and capabilities of their most destructive weapons to a technology evolving at unprecedented speed. This will become particularly challenging as new entrants with varying strategic doctrines and varying attitudes toward the deliberate infliction of civilian casualties seek to develop nuclear capabilities and as equations of deterrence become increasingly diffuse and uncertain.

When the calculation of equilibrium becomes uncertain, or when nations arrive at fundamentally different calculations of relative power, the risk of conflict through miscalculation reaches its height.