# Rendered Reality > Source: > Published: 2026-08-02 00:00:00+00:00 # Rendered Reality The simulation argument, perception, and AI world models In June 2016 Elon Musk told a technology conference that the odds we live in “base reality” were [“one in billions”](https://www.vox.com/2016/6/2/11837608/elon-musk-says-odds-living-in-simulation): games were improving so fast, he reasoned, that fully realistic simulations were inevitable, and the simulated would then outnumber the real. The argument is circular, since it assumes that indistinguishable simulations already exist in order to conclude that we are inside one, but it spread anyway, and over the following decade the simulation hypothesis became both more popular and less frightening. The question has not stayed in one field. It began with philosophers and cosmologists, who asked whether the world was real and whether anyone could tell from the inside. Biology answered it indirectly, since no creature perceives the world as it is. More recently engineers have taken it up, building the renders rather than asking about them. Over that time it stopped being a question about reality and became a description of what is being built. What was Matrix-era dread now circulates as an online meme. The change tracks the technology: between 2023 and 2026 generative AI began producing a large share of the text, images and voices in the shared digital environment, and once the everyday world is visibly rendered, the question of whether the cosmos is rendered loses its charge. The doubt itself is roughly 2,400 years old, and in each era it has borrowed that era’s most advanced machine as its metaphor. In the fourth century BC the Chinese philosopher Zhuangzi gave one of the oldest recorded versions of it: Once Chuang Chou dreamt he was a butterfly, a butterfly flitting and fluttering around, happy with himself and doing as he pleased. He didn’t know he was Chuang Chou. Suddenly he woke up and there he was, solid and unmistakable Chuang Chou. But he didn’t know if he was Chuang Chou who had dreamt he was a butterfly, or a butterfly dreaming he was Chuang Chou. — Zhuangzi, The Complete Works of Chuang Tzu, trans. Burton Watson For Zhuangzi the point was relativism: there must be some distinction between the man and the butterfly, he wrote, but which one is dreaming cannot be settled from the inside. The same doubt appears in Plato’s cave, where prisoners mistake shadows for the world; in Descartes, who in 1641 imagined an evil demon devoted to deceiving him about everything, including mathematics; and in Berkeley, who argued that objects exist only as they are perceived. From Newton onward the universe was often pictured as a machine. It became a clockwork so precise that Leibniz could mock the theology behind it: a God who, [he wrote in 1715](https://www.newtonproject.ox.ac.uk/view/texts/normalized/THEM00226), “wants to wind up his Watch from Time to Time: Otherwise it would cease to move.” Laplace dispensed with the winder and imagined a mind that could compute the whole future from the present. In the nineteenth century the clock became a steam engine running down towards “heat death”; in the twentieth it became a computer. In 1990 the physicist John Wheeler gave the idea a slogan, [“it from bit”](https://philarchive.org/rec/WHEIPQ): reality is, at the root, information, and the universe is not merely like a computer but is one. That is a claim about what the world is made of. The simulation hypothesis is a separate claim about who, if anyone, is running it: that reality is a program built by a more advanced civilisation. Nick Bostrom set out its best-known form in 2003, as [an argument from probability](https://www.simulation-argument.com/simulation.html): if advanced civilisations usually survive and run many detailed simulations of their own history, simulated minds would vastly outnumber real ones, and any given mind would more likely be simulated than not. At least one of three things must therefore be true: civilisations like ours almost always die out first, or those that survive choose not to run such simulations, or we are almost certainly inside one. The question cannot be settled by experiment, though several physicists have tried. They have searched for signatures a simulation might leave: a 2012 paper looked for a cut-off in cosmic-ray energies that a universe computed on a lattice might show; in 2023 the physicist Melvin Vopson proposed a [“second law of infodynamics”](https://pubs.aip.org/aip/adv/article/13/10/105308/2915332/The-second-law-of-infodynamics-and-its) he presented as testable; in 2025 Franco Vazza [calculated](https://www.frontiersin.org/journals/physics/articles/10.3389/fphy.2025.1561873/full) that simulating Earth to the smallest measurable scale would mean converting something like 100,000 stars into energy, making a full simulation “(nearly) impossible.” There is also a structural obstacle: to find evidence that you are simulated, you would look for the corners a simulator cuts, and a competent simulator would smooth any corner the moment you inspected it. A simulator need only render what is being observed, the same way a video game draws detail only where the player is looking, or a language model generates the next word only after something prompts it. The hypothesis also has the structure of a theology. A simulator sits outside space and time, able to read and rewrite the code, with motives no one inside can know; suffering becomes a parameter and good behaviour a wager against deletion. It is intelligent design restated for a secular audience, and like intelligent design, nothing could ever count against it. Some of its believers have acted on it anyway. In 2016 the New Yorker [reported](https://www.newyorker.com/magazine/2016/10/10/sam-altmans-manifest-destiny) that “two tech billionaires have gone so far as to secretly engage scientists to work on breaking us out of the simulation.” No animal perceives reality as it is; perception shows only the slice that keeps the animal alive. In 1909 the biologist Jakob von Uexküll named this the *Umwelt*, the self-world each species assembles from the narrow band of signals its organs can register. His example was the tick, which navigates by three cues alone: the butyric acid of mammal sweat, warmth, and the touch of skin. “Out of the vast world which surrounds the tick,” he wrote, “three stimuli shine forth from the dark like beacons, and serve as guides to lead her unerringly to her goal.” The branch the tick waits on has, for the tick, no colour and no texture; it is only a place to wait for the smell of a passing mammal. Donald Hoffman, a cognitive scientist, gave the idea a mathematical form. He ran evolutionary simulations pitting organisms that perceive the world accurately against organisms tuned only to what bears on survival, and the accurate ones lost. Their [fitness-beats-truth theorem](https://pubmed.ncbi.nlm.nih.gov/33231784/) holds that truthful perception is driven to extinction by perception tuned to fitness, because seeing the world as it is costs effort that seeing only what matters does not. Perception, on this account, is a desktop interface, useful precisely because the icon hides the machinery: “Spacetime is our desktop,” Hoffman writes, “and physical objects, such as spoons and stars, are icons of the interface of Homo sapiens.” His stronger claim, that spacetime is therefore unreal, is contested and mostly rejected; the modest one, that perception is tuned to fitness rather than truth, is mainstream, and it arrives from several directions: evolutionary psychology’s useful errors, the stick taken for a snake, and neuroscience’s account of perception as the brain’s best guess. The neuroscientist Anil Seth calls it a [“controlled hallucination,”](https://www.ted.com/talks/anil_seth_your_brain_hallucinates_your_conscious_reality) noting that “we’re all hallucinating all the time,” and that “when we agree about our hallucinations, we call that reality.” While a hidden computer may or may not be rendering the world, the nervous system certainly is. From a thin stream of signals it builds a workable version of the world, one shaped by evolution to keep its owner alive rather than to be accurate. Nobody has ever stepped outside that interface to see the world plain, so the question is not whether people live behind a render but who builds it. For almost the whole history of life that builder was natural selection, working blind and owned by no one. Engineers have lately joined in, building world models: systems that learn an environment well enough to generate its next frame. In 2018 David Ha and Jürgen Schmidhuber trained an agent [that improved inside its own dream](https://arxiv.org/abs/1803.10122) of a racing game; by 2024 Google DeepMind’s Genie could conjure playable worlds from a single image, and a year later [Genie 3](https://deepmind.google/blog/genie-3-a-new-frontier-for-world-models/) generated interactive three-dimensional environments in real time; a separate model [ran the 1993 shooter Doom](https://gamengen.github.io/) with no game engine beneath it, every frame produced by a neural network. OpenAI was explicit about the aim, describing its Sora generator as [“a promising path towards building general purpose simulators of the physical world,”](https://openai.com/index/video-generation-models-as-world-simulators/) a claim sceptics such as Yann LeCun reject, arguing the models learn the statistics of pixels rather than the laws of physics. Robots, too, learn to walk and to handle objects inside simulated physics before the trained policy is loaded onto real hardware, an entire field of AI depending on simulating the world in order to teach machines to act in it. Researchers have begun putting agents inside these worlds as well: in 2023 a Stanford team filled a simulated town, “Smallville,” with 25 characters driven by language models [that formed memories and threw a Valentine’s party](https://arxiv.org/abs/2304.03442) nobody had scripted, and the next year a startup reported releasing about [1,000 such agents into Minecraft](https://arxiv.org/abs/2411.00114), where, by its own account, they grew an economy and adopted a constitution. What Bostrom had imagined a posthuman civilisation doing, ancestor-simulations full of agents, had become a research demonstration within 20 years of his paper. Building these worlds gives the old question a concrete form: whether a rendered world is any less real than a physical one. The philosopher David Chalmers argues that it is not. In *Reality+* (2022) he holds [that “virtual reality is genuine reality”](https://consc.net/reality/); if a physical object is at bottom a pattern of information, a virtual object is that pattern realised in silicon, and so a real digital object. “What happens in VR really happens,” he writes. The argument arrives at the biologists’ conclusion from the opposite direction: where they find that the physical world is never seen as it is, Chalmers finds the virtual world real even when rendered. Either way, the boundary meant to separate the real from the simulated is hard to locate. Already, millions keep up steady conversations with chatbots built to be friends or confidants, and a 2025 study found [they ease loneliness “on par only with interacting with another person.”](https://academic.oup.com/jcr/advance-article-abstract/doi/10.1093/jcr/ucaf040/8173802) The open web is increasingly machine-written: the analytics firm Ahrefs found [that 74.2% of new web pages](https://ahrefs.com/blog/what-percentage-of-new-content-is-ai-generated) in April 2025 carried AI-generated text. And personal memory is now editable after the fact, by photo tools from Google and Samsung that let people open a shut eye or delete a stranger from a holiday. A Samsung executive met the worry by denying its premise: [“There is no such thing as a real picture.”](https://petapixel.com/2024/02/05/samsung-argues-that-there-is-no-such-thing-as-a-real-picture/) An [MIT study](https://www.media.mit.edu/publications/synthetic-human-memories/) found that AI-edited images roughly doubled the false memories people formed. What is left of “real” is narrower than it was. In 1974 the philosopher Robert Nozick imagined an “experience machine” that could give you any experience you wanted (writing a great novel, making a friend), perfectly and indistinguishably, for the rest of your life. He assumed most people, offered the chance to plug in, would refuse, because they want to do things, not merely to feel that they are doing them, and to stay in contact with a reality outside their own sensations. But in 2010 the philosopher Felipe De Brigard [turned the scenario around](https://www.tandfonline.com/doi/10.1080/09515080903532290): instead of asking whether someone would plug in, he told subjects they were already inside the machine and asked whether they would unplug and return to the real world. Most chose to stay: 87% when the outside life was described as a prison, and 59% even in a neutral version, where subjects were told only that the life they knew was simulated, with nothing said about the world outside. The intuition that reality matters for its own sake is, on this evidence, mostly a preference for the life one already has. Some differences between a physical world and a rendered one remain: a counterpart with an inner life of its own, able to consent or be wronged; a record held in common rather than generated for each user; and consequences that do not reset on reload. None of these is authenticity of feeling, which is what the experience-machine debate had assumed was at issue. The question began as pure metaphysics. Whether the world is real, the biology suggests, was the wrong question from the start. What matters now is what the renders are for, whether a synthetic friend will do, whether a shared record can be kept, and who controls the machines they run on. Bostrom asked how probable it is that we live in a simulation. Now that the simulations are something built rather than supposed, the probability is beside the point. ### The newsletter New pieces as they ship. Unsubscribe whenever.