Am I a machine?

A question I’ve been obsessing over lately goes something like this: Is life a form of computation, and if so, is this thought somehow dangerous?

Of course, drawing an analogy between life and the most powerful and complex tool human beings have so far invented wouldn’t be a historical first. There was a time in which people compared living organisms to clocks, and then to steam engines, and then, perhaps surprisingly, to telephone networks.

Thoughtful critics, such as the roboticist Rodney Brooks think this computational metaphor has already exhausted its usefulness and that its overuse is proving to be detrimental to scientific understanding. In his view scientists and philosophers should jettison the attempt to see everything in terms of computation and information processing, which has blinded them from seeing other more productive metaphors that have been hidden by the glare of a galloping Moore’s Law. Perhaps today it’s just the computer’s turn and someday in the future we’ll have an even more advanced tool that will again seem the perfect model for the complexity of living beings. Or maybe not.

If instead we have indeed reached peak metaphor it will be because with computation we really have discovered a tool that doesn’t just resemble life in terms of features, but reveals something deep about the living world- because it allows us for the first time to understand life as it actually is. It would be as if we’ve managed to prove Giambattista Vico’s claim made right at the start of the scientific revolution “Verum et factum convertuntur” strangely right after all these years. Humanity can never know the natural world only what we ourselves have made. Maybe we will finally understand life and intelligence because we are now able to recreate a version of it in a different substrate (not to mention engineering it in the lab). We will know life because we are at last able to make it.

But let me start with the broader story…

Whether we’re blinded by the power of our latest uber-tool, or on the verge of a revolution in scientific understanding, however, might matter less than the unifying power of a universal metaphor. And I would argue that science needs just such a unifying metaphor. It needs one if it is to give us a vision of a rationally comprehensible world. It needs one for the purpose of education along with public understanding and engagement. A unifying metaphor is above all needed today as a ballast against over-specialization which traps the practitioners of the various branches of science (including the human sciences) in silos unable to communicate with one another and thereby formulate a reunified picture of a world that science itself has artificially divided up into fiefdoms as an essential first step towards understanding it. And of all the metaphors we have imagined, computation really does appear uniquely fruitful and revelatory and not just in biology but across multiple and radically different domains. A possible skeleton key for problems that have frustrated scientific understanding for decades.

One place where the computation/information analogy has grown over the past decades is in the area of fundamental physics, as an increasing number in the field have begun to borrow concepts from computer science in the hopes of bridging the gap between general relativity and quantum mechanics.

This informational turn in physics can perhaps be traced back to the Israeli physicist Jacob Bekenstein who way back in 1972 proposed what became known as the “Bekenstein bound”. An upper limit to entropy, to information, that can exist in a finite area of space. Pack information any tighter than 10⁶⁹ bits per square meter and that area will collapse to form a black hole. Physics, it seems, puts hard limits on our potential computational abilities (they’re a long way off), just as it places hard limits on our potential speed. What Bekenstein showed was that thinking of physics in terms of computation helped reveal something deeper not just about the physical world but about the nature of computation itself.

This recasting of physics in terms of computation, often called digital physics, really got a boost with an essay by the physicist John Wheeler in 1989 titled “It from Bit.” It’s probably safe to say that no one has ever really understood the enigmatic Wheeler who if one wasn’t aware that he was one of the true geniuses of 20th century physics might be confused for a mystic like Fritjof Capra, or heaven forbid, a woo-woo spouting conman- here’s looking at you- Deepak Chopra. The key idea of It from bit is captured in this quote from his aforementioned essay, a quote that also captures something of Wheeler’s koan-like style:

“It from bit symbolizes the idea that every item of the physical world has at bottom — at a very deep bottom, in most instances — an immaterial source and explanation; that what we call reality arises in the last analysis from the posing of yes-no questions and the registering of equipment-evoked responses; in short, that all things physical are information-theoretic in origin and this is a participatory universe.”

What distinguishes digital physics today from Wheeler’s late 20th century version is above all the fact that we live in a time when quantum computers have not only been given a solid theoretical basis, but have been practically demonstrated. A tool born from an almost offhand observation by the brilliant Richard Feynman who in the 1980s declared what should have been obvious. That: Nature isn’t classical . . . and if you want to make a simulation of Nature, you’d better make it quantum mechanical…”

What Feynman could not have guessed was that physics would make progress (to this point, at least) not from applying quantum computers to the problems of physics, so much as from applying the ideas of how quantum computation might work to the physical world itself. Leaps of imagination, such as seeing space itself as a kind of quantum error correcting code, reformulating space-time and entropy in terms of algorithmic complexity or compression, or explaining how, as with Schrodinger’s infamous cat, superpositions existing at the quantum level breakdown for macroscopic objects as a consequence of computational complexity. The Schrodinger equation unsolvable for large objects so long as P ≠ N.

There’s something new here, but perhaps also something very old, an echo of Empedocles vision of a world formed out of the eternal conflict between Love and Strife. If one could claim that for the ancient Greek philosopher life could only exist in the mid-point between total union and complete disunion, then we might assert that life can only exist in a universe with room for entropy to grow, neither curled up too small, or too dispersed for any structures to congregate- a causal diamond. In other words, life can only exist in a universe whose structure is computable.

Of course, one shouldn’t make the leap from claiming that everything is somehow explainable from the point of view of computation to making the claim that “a rock implements every finite state automaton” , which, as David Chalmers pointed out, is an observation verging on the vacuous. We are not, as some in Silicon Valley would have it, living in a simulation, so much as in a world that emerges from a deep and continual computation of itself. In this view one needs some idea of a layered structure to nature’s computations, whereby simple computations performed at a lower level open up the space for more complex computations on a higher stage.

Digital physics, however, is mainly theoretical at this point and not without its vocal critics. In the end it may prove just another cul de sac rather than a viable road to the unification of physics. Only time will tell, but what may bolster it could be the active development of quantum computers. Quantum theory and quantum technology, one can hope, might find themselves locked in an iterative process with each aiding the development of the other in the same way that the development of steam engines propelled the development of thermodynamics from which came yet more efficient steam engines in the 19th century. Above all, quantum computers may help physicists sort out the rival interpretations regarding what quantum mechanics actually means, where at the end of the day quantum mechanics will ultimately be understood as a theory of information. A point made brilliantly by the science writer Philip Ball in his book Beyond Weird.

There are thus good reasons for applying the computational metaphor to the realm of physics, but what about where it really matters for my opening question, that is when it comes to life. To begin, what is the point of connection between computation in merely physical systems and in those we deem alive and how do they differ?

By far the best answer to these questions that I’ve found was the case made by John E. Mayfield in his book The Engine of Complexity: Evolution as Computation.  Mayfield views both physics and biology in terms of the computation of functions. What distinguishes historical processes such as life, evolution, culture and technology from mere physical systems is their ability to pull specific structure from the much more general space of the physically possible. A common screwdriver, as an example, at the end of the day just a particular arrangement of atoms. It’s a configuration that while completely consistent with the laws of physics is also highly improbable, or, as the physicist Paul Davies put it in a more recent book on the same topic:

“The key distinction is that life brings about processes that are not only unlikely but impossible in any other way.”

Yet the existence of that same screwdriver is trivially understandable when explained with reference to agency. And this is just what Mayfield argues evolution and intelligence does. They create improbable yet possible structures in light of their own needs.

Fans of Richard Dawkins or students of scientific history might recognize an echo here of William Paley’s famous argument for design. Stumble upon a rock and its structure calls out for little explanation, a pocket- watch with its intricate internal parts, gives clear evidence of design. Paley was nothing like today’s creationists, he had both a deep appreciation for and understanding of biology, and was an amazing writer to boot. The problem is he was wrong. What Darwin showed was that a designer wasn’t required for even the most complex of structures, random mutation plus selection gave us what looks like miracles over sufficiently long periods of time.

Or at least evolution kind of does. The issue that most challenges the Theory of Natural Selection is the overwhelming size of the search space. When most random mutations are either useless or even detrimental, how does evolution find not only structure, but the right structure, and even highly complex structure at that? It’s hard to see how monkeys banging away on typewriters for the entire age of the universe get you the first pages of Hamlet let alone get you Shakespeare himself.

The key, again it seems, is to apply the metaphor of computation. Evolution is a kind of search over the space of possible structures, but these are structures of a peculiar sort. What makes these useful structures peculiar is that they’re often computationally compressible. The code to express them is much shorter than the expression itself. As Jordana Cepelewicz put it for a piece at Quanta:

“Take the well-worn analogy of a monkey pressing keys on a computer at random. The chances of it typing out the first 15,000 digits of pi are absurdly slim — and those chances decrease exponentially as the desired number of digits grows.

But if the monkey’s keystrokes are instead interpreted as randomly written computer programs for generating pi, the odds of success, or “algorithmic probability,” improve dramatically. A code for generating the first 15,000 digits of pi in the programming language C, for instance, can be as short as 133 characters.”

In somewhat of an analogy to what Turing’s “primitives” are for machine based computation, biological evolution requires only a limited number of actions to access a potentially infinite number of end states. Organisms need a way of exploring the search space, they need a way to record/remember/copy that information, along with the ability to compress this information so that it is easier to record, store and share. Ultimately, the same sort of computational complexity that stymes human built systems, may place limits on the types of solutions evolution itself is able to find.  

The reconceptualization of evolution as a sort of computational search over the space of easily compressible possible structures is a case excellently made by Andreas Wagner in his book Arrival of the Fittest. Interestingly, Wagner draws a connection between this view of evolution and Plato’s idea of the forms. The irony, as Wagner points out, being that it was Plato’s idea of ideal types applied to biology that may have delayed the discovery of evolution in the first place. It’s not “perfect” types that evolution is searching over, for perfection isn’t something that exists in an ever changing environment, but useful structures that are computationally discoverable. Molecular and morphological solutions to the problems encountered by an organism- analogs to codes for the first 15,000 digits of pi.

Does such evolutionary search always result in the same outcome? The late paleontologist and humanist, Stephen Jay Gould, would have said no. Replay the tape of evolution like what happens to George Bailey in “It’s a Wonderful Life” and you’d get radically different outcomes. Evolution is built upon historical contingency like Homer Simpson who wishes he didn’t kill that fish.

Yet our view of evolution has become more nuanced since Gould’s untimely death in 2002. In his book Improbable Destinies, Jonathan Losos lays out the current research on the issue. Natural Selection is much more a tinkerer than an engineer. It’s solutions, by necessity, clugee and a result of whatever is closest at hand. Evolution is constrained by the need for survival to move from next best solution to next best solution across the search space of adaptation like a cautious traveler walking from stone to stone across a stream.

Starting points are thus extremely important for what Natural Selection is able to discover over a finite amount of time. That said, and contra Gould, similar physics and environmental demands does often lead to morphological convergence- think bats and birds with flight. But even when species gravitate towards a particular phenotype there is often found a stunning amount of underlying genetic variety or indeterminacy. Such genetic variety within the same species can be levered into more than one solution to a challenging environment with one branch getting bigger and the other smaller in response to say predation. Sometimes evolution at the microscopic level can, by pure chance, skip over intermediate solutions entirely and land at something close to the optimal solution thanks to what in probabilistic terms should have been a fatal mutation.

The problem encountered here might be quite different from the one mentioned above, not finding the needle of useful structures in the infinite haystack of possible one, but how to avoid finding the same small sample of needles over and over again. In other words, if evolution really is so good at convergence, why does nature appear so abundant with variety?

Maybe evolution is just plain creative- a cabinet of freaks and wonders. So long as a mutation isn’t detrimental to reproduction, even if it provides no clear gain, Natural Selection is free to give it a whirl. Why do some salamanders have only four toes? Nobody knows.

This idea that evolution needn’t be merely utilitarian, but can be creative, is an argument made by the renowned mathematician Gregory Chaitin who sees deep connections between biology, computation, and the ultimately infinite space of mathematical possibility itself. Evolutionary creativity is also something that is found in the work of the ornithologist Richard O. Prum. His controversial The Evolution of Beauty arguing that we need to see the perception of beauty by animals in the service of reproductive or consumptive (as in bees and flowers) choice as something beyond a mere proxy measure for genetic fitness. Evolutionary creativity, propelled by sexual selection, can sometimes run in the opposite direction from fitness.

In Prum’s view evolution can be driven not by the search for fitness but by the way in which certain displays resonate with the perceivers doing the selection. In this way organisms become agents of their own evolution and the explored space of possible creatures and behaviors is vastly expanded from what would pertain were alignment to environmental conditions the sole engine of change.

If resonance with perceiving mates or pollinators acts to expand the search space of evolution in multicellular organisms with the capacity for complex perception- brains- then unicellular organisms do them one better through a kind of shared, parallel search over that space.

Whereas multicellular organisms can rely on sex as a way to expand the evolutionary search space, bacteria have no such luxury. What they have instead is an even more powerful form of search- horizontal gene transfer. Constantly swapping genes among themselves gives bacteria a sort of plug-n-play feature.

As Ed Yong in his I Contain Multitudes brilliantly lays out, in possession of nature’s longest lived and most extensive tool kit, bacteria are often used by multicellular organisms to do things such as process foods or signal mates, which means evolution doesn’t have to reinvent the wheel. These bacteria aren’t “stupid”. In their clonal form as bio-films, bacteria not only exhibit specialized/cooperative structures, they also communicate with one another via chemical and electrical signalling– a slime internet, so to speak.

It’s not just in this area of evolution as search that the application of the computational metaphor to biology proves fruitful. Biological viruses, which use the cellular machinery of their hosts to self-replicate bear a striking resemblance to computer viruses that do likewise in silicon. Plants too, have characteristics that resemble human  communications networks, as in forests whose trees communicate and coordinate responses to dangers and share resources using vast fungal webs.

In terms of agents with limited individual intelligence whose collective behavior can quite clearly be considered quite sophisticated, nothing trumps the eusocial insects, such as ants and termites. In such forms of swarm intelligence, computation is analogous to what we find in cellular automata where a complex task is tackled by breaking up a problem into numerous much smaller problems solved in parallel.

Dennis Bray in his superb book Wetware: A Computer in Every Living Cell has provided an extensive reading of biological function through the lens of computation/information processing. Bray meticulously details how enzyme regulation is akin to switches, their up and down regulation, like the on/off functions of a transistor, with chains of enzymes acting like electrical circuits. He describes how proteins form signal networks within an organism that allow cells to perform logical operations. Components linked not by wires, but molecular diffusion within and between cells. Structures of cells such as methyl groups allow cells to measure the concentration of attractants in the environment, in effect acting as counters- performing calculus.

Bray thinks that it is these protein networks rather than anything that goes on in the brain that are the true analog for computer programs such as artificial neural nets. Just as neural nets are trained, the connections between nodes in a network sculpted by inputs to derive the desired output, protein networks are shaped by the environment to produce a needed biological product.

The history of artificial neural nets, the foundation of today’s deep learning, is itself a fascinating study on the interitative relationship between computer science and our understanding of the brain. When Alan Turing first imagined his Universal Computer it was the human brain as then understood by the ascendent behavioral psychology that he used as its template.

It wasn’t until 1943 that the first computational model of a neuron was proposed by McCulloch and Pitts who would expand upon their work with a landmark paper in 1959 “What the frog’s eye tells the frog’s brain”. The title might sound like a snoozer, but it’s certainly worth a read as much for the philosophical leaps the authors make as for the science.

For what McCulloch and Pitts discovered in researching frog vision was that perception wasn’t passive but an active form of information processing pulling out distinct features from the environment such as “bug detectors”, what they, leaning on Immanuel Kant of all people, called a physiological synthetic a priori.

It was a year earlier in 1958 that Frank Rosenblatt superseded McCulloch and Pitts earlier and somewhat simplistic computational model of neurons with his perceptrons whose development would shape what was to be the rollercoaster like future of AI. Perceptrons were in essence single-layer artificial neural networks. What made them appear promising was the fact that they could “learn”. A single layer perceptron could, for example, be trained to identify simple shapes in a kind of analog for how the brain was discovered to be wired together by synaptic connections between neurons.

The early success of perceptrons, and it turned out the future history of AI, was driven into a ditch when in 1969 Marvin Minsky and Seymour Papert in their landmark book, Perceptrons, showed just how limited the potential of perceptions as a way of mimicking cognition actually were. Perceptrons struggled over all but the most simple of functions, and broke down when dealing with anything without really sharp boundaries such as AND/OR functions- XOR. In the 1970s AI researchers turned away from modeling the brain (connectionist) and towards the symbolic nature of thought (symbolist). The first AI Winter had begun.

The key to getting around these limitations proved to be using multiple layers of perceptrons, what we now call deep learning. Though we’ve known this since the 1980’s it took until now with our exponential improvement in computer hardware, and accumulation of massive data sets for the potential of perceptrons to come home.

Yet it isn’t clear that most of the problems with perceptrons identified by Minsky and Papert have truly been solved. Much of what deep learning does can still be characterized as “line fitting”. What’s clear is that, whatever deep learning is, it bears little resemblance to how the brains of animals actually work, which was well described by the authors at the end of their book.

“We think the difference in abilities comes from the fact that a brain is not a single, uniformly structured network. Instead, each brain contains hundreds of different types of machines, interconnected in specific ways which predestin that brain to become a large, diverse society of specialized agencies.” (273)

The mind isn’t a unified computer but an emergent property of a multitude of different computers, connected, yes, but also kept opaque from one another as a product of evolution and as the price paid for computational efficiency. It is because of this modular nature that minds remain invisible even to their possessor although these computational layers may be stacked with higher layers building off of end product of the lower, possibly like the layered, unfolding nature of the universe itself. The origin of the mysteriously hierarchical structure of nature and human knowledge?

If Minsky and Papert were right about the nature of mind, if in an updated version of their argument recently made by Kevin Kelly, or a similar one made AI researcher François Chollet that there are innumerable versions of intelligence many of which may be incapable of communicating with one another, then this has deep implications for the computational metaphor as applied to thought and life and sets limits on how far we will be able to go in engineering, modeling, or simulating life with machines following principles laid down by Turing.

The quest of Paul Davies for a unified theory of information to explain life might be doomed from the start. Nature would be shown to utilize a multitude of computational models fit for their purposes, and not just multiple models between different organisms, but multiple models within the same organism. In the struggle between Turing universality and nature’s specialized machines Feynman’s point that traditional computers just aren’t up to the task of simulating the quantum world may prove just as relevant for biology as it does for physics. To bridge the gap between our digital simulations and the complexity of the real we would need analog computers which more fully model what it is we seek to understand, though isn’t this what experiments are- a kind of custom built analog computer? No amount of data or processing speed will eliminate the need for physical experiments. Bray himself says as much:

“A single leaf of grass is immeasurably more complicated than Wolfram’s entire opus. Consider the tens of millions of cells it is built from. Every cell contains billions of protein molecules. Each protein molecule in turn is a highly complex three-dimensional array of tens of thousands of atoms. And that is not the end of the story. For the number, location, and particular chemical state of each protein molecule is sensitive to its environment and recent history. By contrast, an image on a computer screen is simply a two-dimensional array of pixels generated by an iterative algorithm. Even if you allow that pixels can show multiple colors and that underlying software can embody hidden layers of processing, it is still an empty display. How could it be otherwise? To achieve complete realism, a computer representation of a leaf would require nothing less than a life-size model with details down to the atomic level and with a fully functioning set of environmental influences.” (103)

What this diversity of computational models means is not that the Church-Turing Thesis is false (Turing machines would still be capable of duplicating any other model of computation), but that its extended version was likely false. The Extended Church-Turing Thesis claims that any possibly efficient computation can be efficiently performed by a Turing machine, yet our version of Turing machines might prove incapable of efficiently duplicating either the extended capabilities of quantum computers gained through leverage of the deep underlying structure of reality or the twisted structures utilized by biological computers “designed” by the multi-billion years force of evolutionary contingency. Yet even if the Extended Church-Turing Thesis turns out to be false, the lessons derived from reflection on idealized Turing machines and their real world derivatives will likely continue to be essential to understanding computation in both physics and biology.   

Even if science gives us an indication that we are nothing but computers all the way down, from our atoms to our cells, to our very emergence as sentient entities, it’s not at all clear what this actually means. To conclude that we are, at bottom, the product of a nested system of computation is to claim that we are machines. A very special type of machine, to be sure, but machines nonetheless.

The word machine itself conjures up all kinds of ideas very far from our notion of what it means to be alive. Machines are hard, deterministic, inflexible and precise, whereas life is notably soft, stochastic, flexible and analog. If living beings are machines they are also supremely intricate machines, they are, to borrow from an older analogy for life, like clocks. But maybe we’ve been thinking about clocks all wrong as well.

As mentioned, the idea that life is like an intricate machine and therefore is best understood by looking at the most intricate machines humans have made, namely clocks, has been with us since the very earliest days of the scientific revolution. Yet as the historian Jessica Riskin points out in her brilliant book The Restless Clock, since that day there has been a debate over what exactly was being captured in the analogy. As Riskin lays out, starting with Leibniz there was always a minority among the materialist arguing that life was clock like, that what it meant to be a clock was to be “restless”, stochastic and undetermined. In the view of this school, sophisticated machines such as clock or automatons gave us a window into what it meant to be alive, which, above all, meant to possess a sort of internally generated agency.

Life in an updated version of this view can be understood as computation, but it’s computation as performed by trillions upon trillions of interconnected, competing, cooperating, soft-machines- constructing the world through their senses and actions, each machine itself capable of some degree of freedom within an ever changing environment. A living world unlike a dead one is a world composed of such agents, and to the extent our machines have been made sophisticated enough to possess something like agency, perhaps we should consider them alive as well.

Barbara Ehenreich makes something like this argument in her caustic critique of American culture’s denial of death, Natural Causes:

“Maybe then, our animist ancestors were on to something that we have lost sight of in the last few hundred years of rigid monotheism, science, and Enlightenment. And that is the insight that the natural world is not dead, but swarming with activity, sometimes even agency and intentionality. Even the place where you might expect to find solidity, the very heart of matter- the interior of a proton or a neutron- turns out to be flickering with ghostly quantum fluxuations. I would not suggest that the universe is “alive”, since that might invite misleading biological analogies. But it is restless, quivering, and juddering from its vast patches to its tiniest crevices. “ (204)

If this is what it means to be a machine, then I am fine with being one. This does not, however, address the question of danger. For the temptation of such fully materialist accounts of living beings, especially humans, is that it provides a justification for treating individuals as nothing more than a collection of parts.

I think we can avoid this risk even while retaining the notion that what life consists of is deeply explained by reference to computation as long as we agree that it is only ethical when we treat a living being in light of the highest level at which it understands itself, computes itself. I may be a Matryoshka doll of molecular machines, but I understand myself as a father, son, brother, citizen, writer, and a human being. In other words we might all be properly called machines, but we are a very special kind of machine, one that should never be reduced to mere tools, living, breathing computers in possession of an emergent property that might once have been called a soul.

 

The Ethics of a Simulated Universe

Zhuangzi-Butterfly-Dream

This year one of the more thought provoking thought experiments to appear in recent memory has its tenth anniversary.  Nick Bostrom’s paper in the Philosophical Quarterly “Are You Living in a Simulation?”” might have sounded like the types of conversations we all had after leaving the theater having seen The Matrix, but Bostrom’s attempt was serious. (There is a great recent video of Bostrom discussing his argument at the IEET). What he did in his paper was create a formal argument around the seemingly fanciful question of whether or not we were living in a simulated world. Here is how he stated it:


This paper argues that at least one of the following propositions is true: (1) the human species is very likely to go extinct before reaching a “posthuman” stage; (2) any posthuman civilization is extremely unlikely to run a significant number of simulations of their evolutionary history (or variations thereof); (3) we are almost certainly living in a computer simulation. It follows that the belief that there is a significant chance that we will one day become posthumans who run ancestor-simulations is false, unless we are currently living in a simulation.

To state his case in everyday language: any technological civilization whose progress did not stop at some point would gain the capacity to realistically simulate whole worlds, including the individual minds that inhabit them, that is, they could run realistic ancestor simulations. Technologically advanced civilizations either die out before gaining the capacity to produce realistic ancestor simulations,  there is  something stopping such technologically mature civilizations from running such simulations in large numbers or, we ourselves are most likely living in such a simulation because there are a great many more such simulated worlds than real ones.

There is a lot in that argument to digest and a number of underlying assumptions that might be explored or challenged, but I want to look at just one of them, #2. That is, I will make the case that there may be very good reasons why technological civilizations both prohibit and are largely uninterested in creating realistic ancestor simulations. Reasons that are both ethical and scientific. Bostrom himself discusses the possibility that ethical constraints might prevent  technologically mature civilizations from creating realistic ancestor simulations. He writes:

One can speculate that advanced civilizations all develop along a trajectory that leads to the recognition of an ethical prohibition against running ancestor-simulations because of the suffering that is inflicted on the inhabitants of the simulation. However, from our present point of view, it is not clear that creating a human race is immoral. On the contrary, we tend to view the existence of our race as constituting a great ethical value. Moreover, convergence on an ethical view of the immorality of running ancestor-simulations is not enough: it must be combined with convergence on a civilization-wide social structure that enables activities considered immoral to be effectively banned.

I think the issue of “suffering that is inflicted on the inhabitants of a simulation” may be more serious than Bostrom appears to believe. Any civilization that has reached the stage where it can create realistic worlds that contain fully conscious human beings will have almost definitely escaped the two conditions that haunt the human condition in its current form- namely pain and death. The creation of realistic ancestor simulations will have brought back into existence these two horrors and thus might likely be considered not merely unethical but perhaps even evil. Were our world actually such a simulation it would confront us with questions that once went by the name of theodicy, namely, the attempt to reconcile the assumed goodness of the creator (for our case the simulator)  with the existence of evil: natural, moral, and metaphysical that exists in the world.

Questions as to why there is, or the morality of there being, such a wide disjunction between the conditions of any imaginable creator/simulator and those of the created/simulated is a road we’ve been down before as the philosopher Susan Neiman so brilliantly showed us in her Evil in Modern Thought: An Alternative History of Philosophy.  There, Neiman points out that the question of theodicy is a nearly invisible current that swept through the modern age. As a serious topic of thought in this period it began as an underlying assumption behind the scientific revolution with thinkers such as Leibniz arguing in his Essays on the Goodness of God, the Freedom of Man and the Origin of Evil that “ours is the best of all possible worlds”. Leibniz having invented calculus and been the first to envision what we would understand as a modern computer was no dope, so one might wonder how such a genius could ever believe in anything so seemingly contradictory to actual human experience?

What one needs to remember in order to understand this is that the early giants of the scientific revolution, Newton, Descartes, Leibniz, Bacon weren’t out to replace God but to understand what was thought of as his natural order. Given how stunning and quick what seemed at the time a complete understanding of the natural world had been formulated using the new methods it perhaps made sense to think that a similar human understanding of the moral order was close at hand. Given the intricate and harmonious picture of how nature was “designed” it was easy to think that this natural order did not somehow run in violation of the moral order. That underneath every seemingly senseless and death-bringing natural event- an earthquake or plague- there was some deeper and more necessary process for the good of humanity going on.    


The quest for theodicy continued when Rousseau gave us the idea that for harmony to be restored to the human world we needed to return to the balance found in nature, something we had lost when we developed civilization.  Nature, and therefore the intelligence that was thought to have designed it were good. Human beings got themselves into trouble when they failed to heed this nature- built their cities on fault lines, or even, for Rousseau, lived in cities at all.


Anyone who has ever taken a serious look at history (or even paid real attention to the news) would agree with Hegel that: “History… is, indeed, little more than the register of the ‘crimes, follies, and misfortunes’ of mankind”. There isn’t any room for an ethical creator there, but Hegel himself tried to find some larger meaning in the long term trajectory of history. With him, and Marx who followed on his heels we have not so much a creator, but a natural and historical process that leads to a fulfillment an intelligence or perfect society at its end. There may be a lot of blood and gore, a huge amount of seemingly unnecessary human suffering between the beginning of history and its end, but the price is worth paying.

Lest anyone think that these views are irrelevant in our current context, one can see a great deal of Rousseau in the positions of contemporary environmentalists and bio-conservatives who take the position that it is us, that is human beings and the artificial technological society we have created that is the problem. Likewise, the views of both singularitarians and some transhumanists who think we are on the verge of reaching some breakout stage where we complete or transcend the human condition, have deep echoes of both Hegel and Marx.

But perhaps the best analog for the kinds of rules a technologically advanced civilization might create around the issue of realistic ancestor simulations might lie not in these religiously based philosophical ideas but  in our regulations regarding more practical areas such as animal experimentation. Scientific researchers have long recognized that the pursuit of truth needs humane constraints and that such constraints apply not just to human research subjects, but to animal subjects as well. In the United States, standards regarding research that use live animals is subject to self-regulation and oversight based on those established by the Institutional Animal Care and Use Committee (IACUC). Those criteria are as follows:

The basic criteria for IACUC approval are that the research (a) has the potential to allow us to learn new information, (b) will teach skills or concepts that cannot be obtained using an alternative, (c) will generate knowledge that is scientifically and socially important, and (d) is designed such that animals are treated humanely.

The underlying idea behind these regulations is that researchers should never unnecessarily burden animals in research. Therefore, it is the job of researchers to design and carry out research in a way that does not subject animals to unnecessary burdens. Doing research that does not promise to generate important knowledge, subjecting animals to unnecessary pain, doing experiments on animals when the objectives can be reached without doing so are all ways of unnecessarily burdening animals.

Would realistic ancestor simulations meet these criteria? I think realistic ancestor simulations would probably fulfill criteria (a), but I have serious doubts that it meets any of the others. In terms of simulations offering us “skills or concepts that cannot be obtained using an alternative” (b), in what sense would a realistic simulations teach skills or concepts that could not be achieved using something short of such simulations where those within them actually suffer and die? Are there not many types of simulations that would fall short of the full range of subjective human experiences of suffering that would nevertheless grant us a good deal of knowledge regarding such types of worlds and societies?  There is probably an even greater ethical hurdle for realistic ancestor simulations to cross in (c), for it is difficult to see exactly what lessons or essential knowledge running such simulations would bring. Societies that are advanced enough to run such simulations are unlikely to gain vital information about how to run their own societies.

The knowledge they are seeking is bound to be historical in nature, that is, what was it like to live in such and such a period or what might have happened if some historical contingency were reversed? I find it extremely difficult to believe that we do not have the majority of information today to create realistic models of what it was like to live in a particular historical period, a recreation that does not have to entail real suffering on the part of innocent participants to be of worth.

Let’s take a specific rather than a general historical example dear to my heart because I am a Pennsylvanian- the Battle of Gettysburg. Imagine that you live hundreds or even thousands of years into the future when we are capable of creating realistic ancestor simulations. Imagine that you are absolutely fascinated by the American Civil War and would like to “live” in that period to see it for yourself. Certainly you might want to bring your own capacity for suffering and pain or to replicate this capacity in a “human” you, but why do the simulated beings in this world with you have to actually feel the lash of a whip, or the pain of a saw hacking off an injured limb? Again, if completely accurate simulations of limited historical events are ethically suspect, why would this suspicion not hold for the simulation of entire worlds?

One might raise the objection that any realistic ancestor simulation would need to possess sentient beings with free will such as ourselves that possess not merely the the ability to suffer, but to inflict evil upon others. This, of course, is exactly the argument Christian theodicy makes. It is also an argument that was undermined by sceptics of early modern theodicy- such as that put forth by Leibniz.    

Arguments that the sentient beings we are now (whether simulated or real) require free will that puts us at risk of self-harm were dealt with brilliantly by the 17th century philosopher, Pierre Bayle, who compared whatever creator might exist behind a world where sentient beings are in constant danger due to their exercise of free will to a negligent parent. Every parent knows that giving their children freedom is necessary for moral growth, but what parent would give this freedom such reign when it was not only likely but known that it would lead to the severe harm or even death of their child?

Applying this directly to the simulation argument any creator of such simulations knows that we will kill millions of our fellow human beings in war and torture, deliberately starve and enslave countless others. At least these are problems we have brought on ourselves,but the world also contains numerous so-called natural evils such as earthquakes and pandemics which have devastated us numerous times. Above all, it contains death itself which will kill all of us in the end.    

It was quite obvious to another sceptic, David Hume, that the reality we lived in had no concern for us. If it was “engineered” what did it say that the engineer did not provide obvious “tweaks” to the system that would make human life infinitely better? Why, are we driven by pain such as hunger and not correspondingly greater pleasure alone?

Arguments that human and animal suffering is somehow not “real” because it is simulated seem to me to be particularly tone deaf. In a lighter mood I might kick a stone like Samuel Johnson and squeak “I refute it thus!” In a darker mood I might ask those who hold the idea whether they would willingly exchange places with a burn victim. I think not! 

If it is the case that we live in a realistic ancestor simulation then the simulator cares nothing for our suffering on a granular level. This leads us to the question of what type of knowledge could truly be gained from running such realistic ancestor simulations. It might be the case that the more granular a simulation is the less knowledge can actually be gained from it. If one needs to create entire worlds with individuals and everything in them in order to truly understand what is going on, then the results of the process you are studying is either contingent or you are really not in possession of full knowledge regarding how such worlds work. It might be interesting to run natural selection over again from the beginning of life on earth to see what alternatives evolution might have come up with, but would we really be gaining any knowledge about evolution itself rather than a view of how it might have looked had such and such initial conditions been changed? And why do we need to replicate an entire world including the pain suffered by the simulated creatures within before we can grasp what alternatives to the path evolution or even history followed might have looked like. Full understanding of the process by which evolution works, which we do not have but a civilization able to create realistic ancestor simulations doubtless would, should allow us to envision alternatives without having to run the process from scratch a countless number of times.

Yet, it is in the last criteria (d) that experiments are  “designed such that animals are treated humanely” that the ethical nature of any realistic ancestor simulation really hits a wall. If we are indeed living in an ancestor simulation it seems pretty clear that the simulator(s) should be declared inhumane. How otherwise would the simulator create a world of pandemics and genocides, torture, war, and murder, and above all, universal death?

One might claim that any simulator is so far above us that it takes little concern of our suffering. Yet, we have granted this kind of concern to animals and thus might consider ourselves morally superior to an ethically blind simulator. Without any concern or interest in our collective well-being, why simulate us in the first place?

Indeed, any simulator who created a world such as our own would be the ultimate anti-transhumanist. Having escaped pain and death “he” would bring them back into the world on account of what would likely be socially useless curiosity. Here then we might have an answer to the second half of Bostrom’s quote above:


Moreover, convergence on an ethical view of the immorality of running ancestor-simulations is not enough: it must be combined with convergence on a civilization-wide social structure that enables activities considered immoral to be effectively banned.

A society that was technologically capable of creating realistic ancestor simulations and actually runs them would appear to have on of two features (1) it finds such simulations ethically permissible, (2) it is unable to prevent such simulations from being created. Perhaps any society that remains open to creating such a degree of suffering found in realistic ancestor simulations  for anything but the reason of existential survival would be likely to destroy itself for other reasons relating to such lack of ethical boundaries.

However, it is (2) that I find the most illuminating. For perhaps the condition that decides if a civilization will continue to exist is its ability to adequately regulate the use of technology within it. Any society that is unable to prevent rogue members from creating realistic ancestor simulations despite deep ethical prohibitions is incapable of preventing the use of destructive technologies or in managing its own technological development in a way that promotes survival. A situation we can perhaps see glimpses of in our own situation related to nuclear and biological weapons, or the dangers of the Anthropocene.

This link between ethics, successful control of technology and long term survival is perhaps is the real lesson we should glean from Bostrom’s provocative simulation argument.