The Conversation is a news site where academics including scientists write about their work and voice their opinions about science and society. Recently, Emma Hart, Chair in Natural Computation, Edinburgh Napier University, explained that “We’re teaching robots to evolve autonomously — so they can adapt to life alone on distant planets.” She launches from artificial intelligence (AI) as it is now and then looks far into the future, when robots may become autonomous, capable of their own replication and evolution. We will teach them how to evolve, she says.
It’s not an original concept. Movies and novels have worked this meme for years, describing sometimes dystopian futures of our robot descendants turning on their creators. Hart has an optimistic view of robotic evolution. She sees human designers training their creations in the art of oversight, such that they would be able to handle all the contingencies a robot starship colonizing distant planets might encounter.
Our work represents the latest progress towards the kind of autonomous robot ecosystems that could help build humanity’s future homes, far away from Earth and far away from human oversight. [Emphasis added.]
Oversight is a key word to follow in her story, as she conjures up visions of robotic ecosystems with training grounds for baby robots, areas for mature ones, and a recycling plant to keep everything green.
Moving Right Along from Behind
What provides the inspiration and motivation for Hart’s confidence in the future of autonomous, reproducing AI? The challenges of planning for all the contingencies that our AI machines will face seem daunting. Enter Darwin:
An impossible brainteaser for humans, nature has already solved this problem. Darwinian evolution has resulted in millions of species that are perfectly adapted to their environment. Although biological evolution takes millions of years, artificial evolution — modelling evolutionary processes inside a computer — can take place in hours, or even minutes. Computer scientists have been harnessing its power for decades, resulting in gas nozzles to satellite antennas that are ideally suited to their function, for instance.
In short, evolution did it, and evolution will do it. Confusion between artificial selection and natural selection lives on. To Hart, they are one and the same. One just takes longer, that’s all (millions of years instead of hours or minutes). She never explains how humans evolved oversight. Maybe a lucky mutation in an ape brain happened millions of years ago. But now that we have it, it will persist, she is sure. Humans must subtract themselves from evolution so that the robots can carry on the valuable trait of oversight.
But current artificial evolution of moving, physical objects still requires a great deal of human oversight, requiring a tight feedback loop between robot and human. If artificial evolution is to design a useful robot for exoplanetary exploration, we’ll need to remove the human from the loop. In essence, evolved robot designs must manufacture, assemble and test themselves autonomously — untethered from human oversight.
That word oversight just flew by two times. Robot designers will need to confer that ability on the robots. They won’t be able to conquer exoplanetary systems without it. But wait; didn’t life on Earth manage to conquer all the ecosystems of our planet without that trait? Darwinism, by definition, is undirected; where did it get oversight?
If something seems amiss in Hart’s story, it is a thread holding the whole sweater together. It is belief in the creative power of natural selection to do anything and everything. Pull on this thread and the whole garment unravels; more precisely, it disintegrates. There never was a sweater. It was all an illusion. A realistic ecosystem in Michael Behe’s view would show organisms getting by with broken traits, not inventing new ones.
Can Oversight Evolve?
Take the evolution of oversight. Hart assumes that oversight like the kind that AI designers employ in their carefully thought-out plans to design robot cities is, itself, a product of natural selection. Maybe the trait we call “oversight” started with a random mutation to the FOXP2 gene or something, giving a dominant hominid male the ability to organize his population to think and plan together. Looking forward from that mutation, nothing in Darwinism would connect it to foresight, logic, or thought. Looking backward from now, Hart would have to conclude that nothing in her evolution of “oversight” connects with truth or logic, either. There are only behaviors that might have been rewarded with survival. For her to assume a connection, she would have to reach into a different worldview and borrow concepts of truth and morality. She cannot conjure those up from Darwinian principles, which are purposeless and unguided.
Hart’s futuristic vision of humans passing on our capacity for oversight to machines depends on the fallacy that mental traits like thoughts, logic, and foresight can be explained by natural selection. At this stage in human evolution in 2021, she must assume, our thoughts and design principles are real: they are based on truth (i.e., that what we perceive connects to external reality) and morality (i.e., that it is morally good to share matters of truth and embrace integrity for its own sake). Taking that for granted, she leaps ahead into the future and envisions our progeny — human robot ecosystems — evolving in a Darwinian way, taking advantage of the creative power of natural selection. It’s a convenient myth, because we’ll never live to see it happen.
As well as being rendered in our simulator, “child” robots produced via our hybrid evolution are also 3D-printed and introduced into a real-world, creche-like environment. The most successful individuals within this physical training centre make their “genetic code” available for reproduction and for the improvement of future generations, while less “fit” robots can simply be hoisted away and recycled into new ones as part of an ongoing evolutionary cycle.
Natural selection will, she believes, pick up where it left off in the evolution of oversight.
Looking forward, the long-term vision is to develop the technology sufficiently to enable the evolution of entire autonomous robotic ecosystems that live and work for long periods in challenging and dynamic environments without the need for direct human oversight.
In this radical new paradigm, robots are conceived and born, rather than designed and manufactured. Such robots will fundamentally change the concept of machines, showcasing a new breed that can change their form and behaviour over time — just like us.
For someone too naïve to understand that bad things can happen in such scenarios, we suggest that Hart read That Hideous Strength and follow it up with C. S. Lewis’s argument from reason (The Magician’s Twin, Chapter 8). Her team is apparently having fun designing very simplistic “body plans” and “brains” with her 3-D printer. The real disappointment will come when AI teams of the future find out that machines have no desire or power to keep doing what their programmers told them to do. Perhaps they will go off in unexpected directions, like the brooms in The Sorcerer’s Apprentice sequence in Fantasia, destroying things left and right and eliminating their own traits if they continue operating at all. Without foresight and mind, nothing requiring oversight will happen. Entropy will rule, as it must.
For more on this theme, listen to Robert J. Marks of Discovery Institute’s Bradley Center on ID the Future.