More profoundly, Dignum argues that the concepts of AGI and ASI are rooted in a misunderstanding of the nature of human intelligence. The aspiration of today’s AI firms is not only to replace human workers but to build something that goes well beyond them: a god in a box—a single technological system that knows, and can do, everything. But intelligence, Dignum notes, can never emerge purely in isolation; it has always been a collective endeavor. “Our evolutionary history reveals that social behaviors like cooperation, communication, and group living were not just important for survival—they were the very foundation upon which our intelligence developed,” she writes. “The more we chase AGI, the more we discover that true superintelligence lies in human cooperation.” This is what she calls 
the “superintelligence paradox”—another conundrum that illustrates how humanity can never be displaced by AI.

If we took the cooperative aspects of cognition seriously, then what kind of AI would we create? Dignum argues that it would look somewhat different from the AI currently being developed. Rather than systems that try to replace human labor, we would imagine those that “work alongside humans to extend our capabilities and enhance collective intelligence.” Such a shift might be facilitated by moving away from the large, expensive, and monolithic AI models of the sort that underlie services like ChatGPT and toward a more modular approach, in which a mix of smaller and more specialized models is made available to workers in ways that respect their autonomy and expertise. This strategy would have the added virtue of diminishing the power of the tech monopolies, since their control of contemporary AI is, as Dignum notes, inseparable from the fact that they are the only actors with sufficient infrastructure to train and deploy large models.

The central argument of The AI Paradox, then, is that there is nothing inevitable about AI’s present trajectory. Dignum concludes her book with a plea for a more intentional and inclusive approach to AI development, one in which “everyone has a voice in shaping the direction AI takes.” She wants to banish the quietism that too often clouds people’s minds when technology is involved. “We must resist the seductive narratives that portray AI as an unstoppable force beyond human control, narratives that strip us of our agency and render us passive in the face of technological change,” she declares. AI is made by people, and therefore it “is what we, people, make of it…. The power to decide lies with us.”

Dignum’s message is an empowering one: Humans have a monopoly on true intelligence. AI is simply another tool, like an airplane or a car, and we can steer it in any direction we want.

If The AI Paradox had been published a decade ago, these claims would be easier to sustain. But the arrival of large language models (LLMs)—the computational systems that form the engine of generative AI—in 2018, and their rapid subsequent evolution, has cast doubt on some of Dignum’s assertions. While she concedes that “LLMs represent an incredible advancement,” they do not prompt her to revise her overall view of AI. In her account, LLMs have the same fundamental limitation as the AI systems that preceded them: They are incapable of “actual comprehension.” “They do not ‘think’ or ‘know,’” she writes, “they merely simulate patterns extracted from their training data.”

Given Dignum’s career as a distinguished scholar who has worked in AI for decades, few people are as qualified as she is to offer a judgment on LLMs. But opinion within the field is far less settled than she suggests. Because LLMs are more complex than their predecessors, they pose interpretative questions that are harder to answer. Are they purely imitative, or do they exhibit “emergent” properties on account of their complexity? Are they best understood as pattern-matching machines, or can they engage in conceptual reasoning of certain kinds? Among AI researchers and practitioners, these are matters of active debate. And this debate cannot be reduced, as some AI deflationists suggest, to a struggle between truth-tellers and the marketing department of OpenAI. There are genuine disagreements over how to characterize LLMs and their capabilities. In some cases, these involve empirical disputes over what an LLM is actually doing at any given moment. At other times, the disagreements are more semantic or philosophical, centered on the meaning of terms like reasoning 
and comprehension.

The simplest way to describe an LLM is as a system that tries to predict the next word in a sequence, based on the probabilities it has gleaned from its training set. An LLM learns how to make these predictions through a series of baroque computations whose convolutions are not fully understood. We know why LLMs work—that is, we have a good sense of their basic mechanisms. What’s less clear is how they work: Even their creators can’t say with precision why a model produces a particular response. This is the reason that the debates around LLMs are so vigorous and, perhaps, irresolvable. The technology is, in certain important respects, unruly and opaque.

By contrast, cars and airplanes are not: You can open them up and see how they work. They are deterministic systems that do what you tell them to do. Dignum frequently emphasizes that AI is made by humans. But just because something is made by humans doesn’t mean it will remain within the ambit of human comprehension and control.

If AI is not like a car or an airplane, then what is it? At one point, Dignum describes LLMs as “a cognitive Frankenstein’s monster.” She means it in a minimizing way: LLMs work by “piecing together fragments of human language in a way that appears intelligent” but isn’t. “Like Frankenstein’s monster,” she writes, “they lack genuine understanding and intentionality.”

I hope it is not too pedantic to point out that this is a misreading of Mary Shelley’s novel: Frankenstein’s monster does indeed think, feel, and scheme. He teaches himself to read and loves Paradise Lost. He craves companionship and hates his creator for abandoning him, a hatred that moves him to kill the man’s wife and brother.

Yet Dignum’s analogy does resonate, albeit for different reasons. Frankenstein is a story about the relationship between human beings and their alien offspring. The monster is created by a human scientist and even assembled from human body parts. Yet he is feared and hated by nearly everyone he encounters because of his “unearthly ugliness.” Despite being wholly man-made, he is somehow otherworldly.

LLMs have a comparable set of qualities. They are, on the one hand, a product of human ingenuity—an achievement enabled by more than eight decades of research into computational models loosely inspired by the human brain. They are also a composite of human culture in the broadest sense, having been trained on large portions of the publicly accessible Internet, along with books, academic papers, and other sources. Yet for all the humanness of their inputs, LLMs are irreducibly nonhuman in their operation. They learn by studying large quantities of text and constructing elaborate mathematical maps of the semiotic relationships within them. This is not how the human brain works.

Accordingly, we might think of an LLM as something like Frankenstein’s monster: an alien of human ancestry that is not wholly assimilable to our purposes. This metaphor must be handled with caution; it should not be taken to mean that AI is sentient or supernatural. I agree with Dignum’s assertion that we need to demystify AI and to construct a “simple, clear narrative” about the technology. In doing so, however, we should be careful not to efface the fundamental weirdness of LLMs.

We should also be skeptical of the opposition that she sets up between humanity and technology. This is not a binary that feels supportable, least of all now, as our technological 
entanglements become even more consuming than they were when Donna Haraway christened us “cyborgs” in 1985. LLMs may never “attain the full spectrum of human intelligence,” as Dignum says, on account of their not being human. But they have clearly achieved a kind of hybridity with humanness that enables them to act in ways that most people perceive as at least quasi-intelligent. Rather than dismissing such a perception as delusional, we might see it as evidence of the technology’s mongrel character. LLMs are an object lesson in the porousness of the human as a category, as well as our tendency to extrude ourselves into our artifacts—artifacts that can, in turn, exert influence over us. This is not necessarily a good thing. People are 
having psychotic breakdowns from talking to AI chatbots. Victor Frankenstein dies filled with regret.

You might be asking yourself why the way we interpret LLMs even matters. The answer is that it has consequences for how we respond to AI politically, a subject that Dignum engages with throughout her book. If we think of the technology as a car, for instance, that implies a certain approach. Cars provide certain benefits but also “cause accidents,” Dignum notes. Fortunately, they are “much safer and more efficient” than they were 50 years ago, thanks to regulation. Today’s AI is like “a car without brakes or seatbelts,” which means that we need to find the AI equivalents of such measures. “Just as we regulate cars to protect ourselves from accidents and misuse, AI also requires safeguards to prevent harm and ensure it serves humanity’s best interests,” Dignum advises.

She believes such safeguards should be anchored in the “principles of ethical AI”—justice, accountability, transparency, and the protection of individual rights—and developed through an “ongoing dialogue” among “technologists, ethicists, policymakers, and communities.” Only by “involving diverse stakeholders in the decision-making process” can the correct balance be struck. It is very important to Dignum that regulation not be seen as impeding AI’s development. “Just as brakes and safety measures allow cars to go faster, regulation enables innovation to grow responsibly and sustainably,” she writes. More specifically, the absence of governance could cause “trust in AI [to] erode, leading to slower adoption or even rejection of the technology.” Policymakers can help accelerate AI’s integration into society while ensuring that it remains respectful of our rights and equitable in its distribution of benefits.

These passages convey a faith in managed capitalism that feels distinctly European. The picture is one in which representatives from government, industry, and civil society come together to forge policy frameworks that 
harmonize their interests. We might ask whether harmony is possible, or what kind of struggles might 
need to be waged to compel the tech giants to submit to such a process. (They are currently fighting EU regulators tooth and nail.) But the deeper root of Dignum’s optimism is her view of AI itself. Because AI is a tool, we can retool it. “In whatever way we define AI, it is crucial to understand that it is an artifact, that is, something created by people,” she explains. “Since we build it, we control and are responsible for its trajectory and choices.”

But what if AI is better understood as Frankenstein’s monster—a man-made yet alien entity, by turns familiar and strange, unpredictable and not fully fathomable, semi-obedient at best? Not all AI fits this description, but LLMs do, and LLMs are what the tech industry is trying to make ubiquitous and indispensable. It seems unwise to adopt a policy agenda that promises to help the industry do so, even if the correct technocrats are somehow put in charge. AI can be a tool, and a useful one, but it can also be something else. I am personally not someone who worries about AI killing us all, but I do think that granting such a technology unlimited power over the conditions of our life and work is likely to be a recipe for chaos and misery. Our best hope, at least in the short term, might be to pursue a strategy of containment in which AI is restricted to certain spheres and functions on the theory that alien encounters can be fruitful, but alien invasions are bad.