Excerpt from Chapter Nine
Evolutionary Arms Races
Ionce asked ChatGPT, “What are some examples of arms races in evolution?” It was a trick question, and the bot fell into my trap.
It replied, “In evolutionary biology, an arms race refers to a cycle of adaptations and counter-adaptations between interacting species.” It then gave some examples. Like: As cheetahs evolved to run faster, so did animals they aspired to catch and eat, such as gazelles. And the faster gazelles got via this counter-adaptation, the faster cheetahs got via counter-counter-adaptation. And so on.
And: “Some snakes, like the king cobra, have evolved powerful venoms to immobilize their prey, while some prey species, like certain mongooses, have developed resistance to these venoms, leading to an arms race in toxicity versus resistance.”
Also: Cuckoos lay their eggs in the nests of other kinds of birds, such as warblers, to trick the warblers into spending their time bringing cuckoos into the world. “Host birds have evolved the ability to recognize and reject foreign eggs, while cuckoos evolve better mimicry to make their eggs more similar to the host’s eggs.” An arms race of deception and detection.
A deception-detection arms race has also shaped two species of fireflies—Photuris and Photinus. Photuris fireflies evolved the ability to mimic the flashing pattern that, when emitted by a Photinus female, is a mating call. When Photinus males respond to the fake signal, the Photuris fireflies eat them. As a result, ChatGPT observed dryly, Photinus males “have evolved more selective courtship behaviors.” Which puts the ball in the Photuris court—another deception-detection arms race.
These answers were accurate insofar as they went. In fact, they were textbook answers to the question I’d asked—which isn’t surprising, given that GPT’s training data included textbooks. My question hadn’t been a trick question in the sense of luring the AI into offering a false answer. My goal was just to lure it into offering a misleadingly narrow answer. And that’s what it had done.
Arms races between species are indeed, as biology textbooks and the AIs trained on them will tell you, an important part of evolution. But there’s another kind of arms race that’s also important and, in my view, doesn’t get enough emphasis: arms races within species.
Consider the human species.
People have been known to deceive other people—sometimes consciously and deliberately, by telling big or little lies, but sometimes in subtler ways, ways the deceivers themselves may be unaware of. For example: by conveying, through words and gestures and emotional tone, more in the way of future commitment to a person or a group than is, in fact, likely. People have also been known to look for signs that they’re being deceived—by, say, closely monitoring a person’s face as a commitment is being made, or scanning the person’s words for contradictions or inconsistencies.
Evolutionary psychologists have amassed evidence that these two kinds of skills—deceiving people and detecting deception—are mediated by specific cognitive mechanisms that were shaped by natural selection to those ends. There seems to have been an evolutionary arms race of deceptive and detective powers—kind of like the cuckoo-warbler arms race, or the Photuris-Photinus arms race, except intraspecies rather than interspecies. The better the average early human got at deceiving other people, the better the average early human got at spotting deception—and vice versa.
This kind of intraspecies arms race dynamic may also explain why people are natural arguers—and why our powers of argumentation have the particular form they have. Early humans who were good at winning certain kinds of arguments—about who deserved the most food, or who had wronged whom and therefore deserved recompense, or who had sacrificed most for the group and therefore deserved future preference—would have gotten, on average, more of their genes into the next generation than early humans who tended to lose such arguments. So genes that make us good arguers would have flourished. And, as with all the arms race examples offered by ChatGPT, the direction of adaptation would have been sustained by a positive feedback cycle: The better the average person got at arguing, the better an arguer you’d have to be to outcompete your rivals in Darwinian terms—to get more genes into the next generation.
This evolutionary shaping of our powers of argumentation could explain some consequential features of human psychology. Such as confirmation bias, which leads us to notice and uncritically embrace evidence that supports our positions and to overlook or doubt evidence that challenges them. We are designed to sometimes favor victory over truth. That’s what an arms race can do to you.
The tendency I’m complaining about—the tendency of which I find ChatGPT guilty—is the tendency to think of “arms races” as things that happen only between large and distinct groups of beings. It’s an understandable tendency. After all, the term “arms race” arose in describing competition between the very large and distinct groups of beings known as nations. Still, it’s an unfortunate tendency. If you want to understand how broadly powerful arms races can be—how much evolutionary change they can drive, and how fast they can drive it—you have to understand how pervasive and fine-grained they can be.
And that goes for technological as well as biological evolution. When people talk about AI “arms races,” they’re often talking about arms races between nations. Those are definitely important—especially the arms race between AI superpowers China and the United States (certainly including the part of that arms race that involves actual arms—like autonomous weapons guided by AI). But if you want to clearly envision AI’s future trajectory, you have to take a larger view of arms races. You have to appreciate the multiple levels at which they are driving the evolution of AI and see how subtly this multilevel evolution will shape the character of AIs…
Excerpted from The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning by Robert Wright. Copyright © 2026 by Robert Wright. Published by Simon & Schuster.
