From Natural Selection to Machine Selection: Human Evolution in the Age of AI

Around 30,000 years ago, when some Asian grey wolves began to tolerate—and be tolerated by—humans, they set in motion one of evolution’s great marvels. Instead of competing with human hunters, the wolves scavenged food that humans left behind. They hung around human settlements and, instead of menacing people, they did things that people found useful, like chasing off other aggressive scavengers and predators. As people took to tolerating the most relaxed and friendly wolves and chasing away the aggressive ones, they shaped their lupine neighbours’ evolution. This evolution began without any human intention, a mere consequence of those interactions that sent the more skittish and aggressive wolves packing. Slowly, those genes that disposed wolves to skittishness and aggression around humans became rarer, replaced by genes that disposed the animals to being more relaxed, and to reading and responding to human intentions. This is how pack-hunting wolves found their way—almost accidentally—to running with a human pack. We now know this offshoot of the wolf lineage as Canis familiaris, the domestic dog.It’s no coincidence that dogs became our best friends. Our ancestors made dogs their friends by keeping the most friendly, loyal, useful individuals, often pairing them with one another to make even friendlier puppies. Evolutionary biologists call this artificial selection because humans guided evolution by influencing which wolf-dogs lived and reproduced. When Charles Darwin discovered the process by which evolution occurs, he called it natural selection because he realised that nature does what farmers and animal fanciers had long done: it selects which individuals to breed or cull.

Dogs were neither the first nor the most important species that humans domesticated. Perhaps the most crucial phase of human evolution saw our ancestors domesticate their own kind, turning a skittish, belligerent ape into the highly cooperative, communicative creature we see in the mirror. Self-domestication, more than any other aspect of our evolution, made it possible for humans to spread into every habitable corner of the world, building societies and transforming landscapes.

Understanding human self-domestication deepens our understanding of ourselves. That understanding is crucial to building more peaceful societies and tackling major issues like climate change. It is even more pressing, however, because we are busy creating a third, entirely new kind of evolutionary selection: artificially intelligent selection. We are handing over the heavy work of human domestication to machines. And that will probably have important and often unanticipated consequences.

The Ape that Tamed Itself

Humanity domesticated itself by influencing which children grew up to become adults, which adults died, and who got help when they started families of their own. Little by little, our ancestors discovered that kind, cooperative individuals are worth associating with because the gains from working together outweigh the costs of sharing. People who knew how to coordinate their efforts and get along with one another succeeded in the all-important business of reproducing and helping relatives reproduce. You and I inherited far more genes from them than from the more selfish, disruptive or violent individuals who lived alongside them.

Anthropologist Richard Wrangham suggests that self-domestication got a push whenever people conspired against the most violent and disruptive individuals—usually men—and killed them in the name of keeping the peace. Paradoxically, then, it was organised violence that weeded out the genes that dispose people to reactive violence. But Wrangham argues that this proactive violence—in which the people acted as jury, judge and executioner—differs both psychologically and in its underlying genetics—from the reactive violence of the disruptive men.

Evolution usually works via several means. That’s likely to be true of human self-domestication, too. As humans evolved into cooperative creatures, capable of working together in fairly large communities, male testes got smaller and female ovulation became harder for outsiders to discern. Both these changes indicate that our ancestors were evolving into more invested parents, inclined to care for each child for far longer than would a chimp or bonobo parent.

Our ancestors weren’t only parenting their own offspring either. Grandparents, aunts, uncles, siblings and unrelated friends all got in on the important business of building human capital. That proverbial village that raises each child was real: it was the many villages where our ancestors lived, reproduced and died over tens of thousands of generations. Those communities mostly numbered only a few hundred people, working together to raise healthy children, make a living, neutralise violent neighbours and thwart hostile invaders.

They coordinated their efforts by chatting with one another—gossiping about food sources, weather and who should be avoided. Language made for better cooperation, which required more flexible intelligence, which necessitated more sophisticated language, in a virtuous cycle that drove the evolution of big brains, complex societies and sophisticated cultures.

As humans evolved into smarter, more sociable, more cooperative creatures, they became less prone to reactive violence and unpredictability. In so doing, our ancestors trod much the same evolutionary path as dogs and every other domesticated species.

Domestication Matters

Self-domestication etched into human nature the traits we most admire in our fellows: intelligence, conversation, community-mindedness, kindness, reciprocation and sacrifice— in a word, our humanity. It also eroded violence and selfishness, traits that are far less common than in our evolutionary past. Indeed, self-domestication probably crafted not only our behaviours, but also our moral sensibilities, bestowing innate notions of right, wrong, fair and kind that taught us how to treat those in our community and how to deal with transgressors.

Whenever new technologies or environments changed how people lived, bred, cooperated and fought, the strength and direction of selection on our ancestors’ genes shifted too. Those shifts were never intended. The gene pool merely sloshed about like groundwater tides beneath the women and men struggling to get by and, perhaps, to improve their comfort levels. The results, however, included ever-changing combinations of genes: evolution deep beneath human awareness.

Unintended Consequences

Starting only about 12,000 years ago, another suite of domestication projects began. In various parts of the world, gatherers domesticated cereal grains like wheat, rice, corn, sorghum and millet. Those new crops could be farmed, and they nourished something new: complex agricultural societies. New forms of justice, codified in religious or secular law and imposed by rulers and the elite priests of moralising deities, replaced hunter-gatherer style human domestication. Those developments may explain why studies of gene frequency changes show that human evolution did not stop when complex societies and agricultures arose; it accelerated.

Every time humans domesticated a species, the process began with inadvertent selection: some wolves were tolerated rather than killed, a few seeds from particularly nutritious ears of wheat or corn germinated in piles of domestic compost. Eventually, however, canny humans figured out that they could deliberately preserve and breed the most favoured traits. The same is, unfortunately, true of human domestication. Occasionally, the unintended genetic changes that arise from people doing what they do have been exploited in attempts to deliberately shape human lives through forced selective breeding programmes and abortions, forced sterilisations, infanticides, genocides and other forms of violence and coercion.

Evolutionary science will forever grapple with the shame of eugenics and scientific racism and their elaborate proposals for human improvement by third-party artificial selection. Worse still, an embarrassingly large number of governments adopted such eugenic practices. We must remain vigilant so as to avoid repeating these horrific mistakes in future.

My interest here, though, is not in those deliberate attempts to sift genes, but in the inadvertent and often idiosyncratic effects of institutions and practices on human evolution. As with inadvertent selection during early domestication, the effects aren’t as dramatic or directed, but they have enormous cumulative power over generations. Practices like incarceration, execution—and even the allocation of medical care—have inadvertent effects. So do institutions that decide who gets exiled or shunned, who can immigrate and the kinds of welfare available to people in need. Even the most banal decisions made by rulers, governments, occupiers and religious leaders can influence the reproduction of billions of individuals—promoting some genes at the expense of others.

One decidedly speculative paper suggests that the Roman state, by eliminating the most unruly men through exile or execution and quashing uprisings, banditry and piracy, also selected for more peaceful and submissive genotypes, pacifying Romans relative to the enemies that surrounded them. The argument that Rome was special in this regard is built on flimsy evidence and belied by the well-documented conflicts within Roman society, as well as the voracious Roman appetite for war.

In any case, a few hundred years of selection for more pacific genes among those considered Roman are unlikely to have been able to hold up in the face of gene flow from waves of migration and conquest. But perhaps growing state power and organisation over the last several thousand years—originating in various parts of the inhabited world, including Rome—might have advanced human domestication by favouring genes that dispose people to non-violent conflict resolution.

My point is not that humans evolved to be peaceful, or war-like, selfish or cooperative. It’s that practices and institutions that affect people’s lives cannot help causing deep ripples in our gene pool. Consider, for example, the fact that—in the contemporary west, at least—religious individuals tend to have more babies than secular people. Religious teachings are often pronatalist and congregations provide support for parents. Many also preach messages of cooperation, self-sacrifice and charity. Religious practices may, then, inadvertently reinforce the genetic domestication of the flock. Mathematical models have shown that religious support for reproduction and in-group cooperation can promote the spread of genes that dispose individuals to religious belief. In that scenario, culture and evolution combine to produce greater religiosity, as well as greater cooperation, at least among the devout.

Domesticating Machines

As artificial intelligence percolates through human societies, ideas about evolutionary domestication attain a new pertinence. Many forms of artificial intelligence improve at their assigned jobs by a process much like artificial selection. Machine learning algorithms are statistical AI models that get better by learning from the data they are fed. Some machine learning mimics the way humans and other animals learn from experience. Some improve by a process more like evolution, in which new errors—equivalent to genetic mutations—are introduced, and the algorithms that are most successful are bred to form the next generation of algorithms. Machine learning algorithms, then, can be viewed as twenty-first-century beasts of burden, domesticated to do heavy computational lifting for their owners in a data-rich world.

Even as people, corporations and governments shape these new machines to their purposes, they are using them in ways that could also affect humanity’s gene pool and influence our species’ ongoing, idiosyncratic evolution. For example, the AI that guides bail, sentencing and parole decisions has proved vulnerable to statistical perversions that perpetuate biases inherent to the data sets they learn from, recommending harsher sentences for some offenders based largely on their race or the neighbourhood they come from. Sentencing and parole also limit convicts’ opportunities to mate and raise children, making AI-guided genetic selection a distinct—though unintended—possibility.

AI can also be used to cause deliberate harm. AI-enhanced facial recognition, network tracing and text analysis can be used to identify and locate members of specific ethnic or religious groups for the purpose of committing violence against them. Indeed, the AI doesn’t even have to be aimed at the victims. Users engage avidly with posts that spread harmful conspiracies, polarise polities or incite violence. Algorithms that learn how to promote engagement routinely converge on such posts, pushing them to the tops of users’ feeds. The consequences—from vaccine hesitancy to violently rejecting an election result—can influence life, death and reproduction.

The 2016–17 campaign of violence by the Myanmar military against the roughly 1 million Rohingya Muslim minority provides one noteworthy example. That offensive, a “textbook example of ethnic cleansing” according to a UN commission, was stoked by social media posts by anti-Rohingya civilians. Facebook is currently facing lawsuits alleging that its algorithms amplified incitements to violence. It would seem that AI algorithms designed to keep people on platform can incidentally abet the violent removal of an entire ethnic-religious minority, with profound consequences for a region’s religious, linguistic and ethnic makeup. As a result, such algorithms also affect genetic variation.

Both criminals and political dissidents have fewer places to hide in the modern world of extensive AI-abetted surveillance than they once did. While a world safer from crime might conceivably extend the ongoing project of human domestication, a world in which dissent is no longer possible could have the opposite effect.

According to Richard Wrangham, ordinary people working together to remove violent and despotic leaders drove some of the more salutary aspects of human domestication. At first glance, social media has been an important tool for organising protests of all political stripes, from the Arab Spring to Black Lives Matter. But both the AI and the data from which they learn are owned by large corporations, and if the insights they offer are sold to the highest bidder, corporations and governments will win out against activists every time. AI, working through social media data, is more likely to weaken protests and undermine dissidents than to empower them. In so doing, it will also weaken the self-domesticating power that comes with conspiring against despots.

Wherever computers and large datasets help make decisions that influence individual lives, reproductive choices, family support or migration, those decisions will inevitably come to be shaped by machine learning. Those decisions—in spheres like public health, welfare, immigration policy and justice—will also affect gene frequencies. The genetic changes are likely to be small in any single generation, but the thing that makes evolution so potent is that tiny changes, accumulating over many generations, often have dramatic effects.

Cumulative changes could become especially important given the insidious ways AI infiltrates the mechanics of how people make friends and grow intimate. Social media sites like Facebook, Instagram and Tik Tok now occupy so much of their users’ time and headspace that they squeeze out other activities and usurp the time normally reserved for family members, offline friends and lovers. These displacement effects—together with the relentless self-comparison with the curated social media highlights of other people’s lives that some of these sites encourage—are among the suspected causes of the current wave of youth mood disorders.

Social media is overhauling who will live, mate and become parents to the next generation of humans. The social and psychological traits that worked best in the pre-social media environment differ from those that work best in the new era. The overhaul of human social interaction wrought by smartphones and social media has caused a change to the environment we inhabit far more rapid than the changes caused by agriculture and anthropogenic climate change. We can thus expect that social media will alter gene frequencies, introducing a new ripple into human domestication.

Predicting the direction of evolution is a mug’s game. But, if pressed, I’d say that the traits likely to be favoured in coming generations include the abilities to hold one’s own in the new attention economy, resist distraction, navigate rampant social comparison without succumbing to envy or depression, and avoid or survive mood disorders. They will also include the ancient capacity to make the true, nourishing friendships and alliances that people depend on, even as ways of communicating and of tending to those relationships change apace.

Hopefully, technology can be used to ameliorate the very worst of what is happening right now. But any genetic differences between those who flourish and those who flounder in the new social world will inevitably provide material for natural selection. As AI shapes who lives, dies and reproduces, and as it learns to make social media ever more compelling, we can expect that it will also influence the next steps in human domestication.

And AI will do that job even if no one asks it to.