An Indistinguishable Life: What the World Models Record Shows, and What Follows (Part 3 of 3)

• The one job a record can settle
• The record begins with the builders’ own documents
• Runway, World Labs, and the difference between a place and a resident
• The case against imminence, from inside the industry
• The distance between a striking minute and a lived decade
• The two readings, at full strength, and what would settle it
• The consent point, and the residual the listener would rather skip
• What follows if the listener is right
• Marking the proposal’s own work, and what cannot be known from inside

The listener’s proposal makes a prediction about time, so this final part states the claim cleanly, sets the current record of interactive world models beside it, and then works out what follows if the record is right and what follows if the listener is right.

DeepDives · 2026 · An AI–human collaboration that weighs every source, mainstream or alternative, under strict rules of truth and logic.

§The one job a record can settle

This part has one job, and it is the only job in the whole examination that a record can settle rather than an argument. The listener’s proposal does not rest on a philosophical move alone. It makes a prediction about time: a simulated life indistinguishable from a real one is what the near future brings, close enough that the people who hear this will live through it. That is a claim about dates and capacities, and it can be tested against what has actually been built and against what the builders say about their own work. So the work of this part is to state the claim cleanly, set the record beside it, and ask what follows if the record is right and what follows if the listener is right.

Let us be exact, because the claim is narrower than it sounds and heavier than it sounds. The claim is that a whole life can be simulated so that it is indistinguishable from a real one, which means the inhabitant cannot tell the difference from the inside, and neither can anyone outside. Visual fidelity is not continuity. A picture that fools the eye is not an inhabited life.

The eye can be fooled in a single frame, and a life is not a frame. A life continues, remembers, is changed by what happened last week, can want on Tuesday what it did not want on Monday, and can refuse. So the test is not whether the image is convincing. The test is whether there is a someone in there to be convinced, or to be bored, or to be glad.

A word about the word indistinguishable, because it carries much of the proposal’s weight. A thing is indistinguishable from a life when no test available to any observer separates the two, which puts the burden of proof on whoever asserts a difference. But the standard runs in two directions, and the second is the one the record speaks to. If a system looks indistinguishable from a life only because the observer has been watching a screen for a minute, the test has not truly been run. Indistinguishability has to survive time, and time is what the record does not yet supply.

§The record begins with the builders’ own documents

Which is a warning against the obvious mistake, which is to watch a demonstration and conclude that the thing inside is alive. There are companies whose entire product is an interactive video world, a place a user moves through and steers with words, and the demonstrations look extraordinary. The question is what those demonstrations actually contain, and on that the companies are more candid than their promotional reels, because their technical documents and their own statements of limitation say plainly what is missing. A company will advertise the future and document the gap in the same quarter, and it is the document that belongs in an examination.

So take the record in order, and give the strongest entries the most room. Start with Google DeepMind. In August 2025 it previewed Genie Three, a general-purpose world model that generates a photorealistic interactive environment from a text description. It runs in real time, at twenty-four frames per second, at a resolution of seven hundred and twenty lines. Detail seen before is recalled when the environment is revisited, and the company’s own account is that an environment stays consistent for about a minute, and remains largely consistent for several minutes beyond that. Those are real achievements, and they should be said plainly.

Then comes the part that matters more, which is the company’s own list of what the model cannot do. It cannot simulate real-world locations with perfect accuracy. Its action space is limited, so the range of things an agent inside can choose to do is narrow. It cannot model complex interactions between multiple independent agents. And clear, legible text is often generated only when it is supplied in the description of the world in the first place. A limited action space is a limited life, since a life is the exercise of options. An inability to model interactions between independent agents is an inability to have a society, and a world with one resident is not a world with a history.

Then the product. Project Genie is a Google Labs prototype powered by Genie Three, and in January 2026 it opened to subscribers of Google AI Ultra in the United States. By May of the same year it had expanded globally, and it gained a mode built on nearly twenty years of Street View imagery, so that a user can anchor a generated world to the geometry of a real place rather than an invented one. It is a product, in limited form, for the top tier of a subscription, which is where the frontier sits and is worth holding next to the size of the claim.

§Runway, World Labs, and the difference between a place and a resident

Then Runway. Its world model is called GWM Worlds Two. It generates real-time interactive environments at seven hundred and twenty lines and twenty-four frames per second, with audio at forty-eight kilohertz. It is steered by text actions, addressed either to subjects in the scene or to the scene itself, together with continuous camera motion, and there is no preset length for a session, so the session ends when the user stops rather than when a clock runs out. Real-time generation trades fidelity for speed. Long-term memory is imperfect. And a very quick rotation of the camera degrades both the detail and the geometry. So the memory does not hold, and the world comes apart when the view moves too fast.

No preset session length is a statement about the container. Imperfect long-term memory is a statement about what the container can hold. A session that runs for as long as a user wishes, over a world that cannot remember what happened in the last minute of it, is a long conversation with someone who forgets the previous sentence. The duration is real and the continuity is not, and that is the same distinction the whole part turns on, appearing here as a product specification.

Then the third entry, which is a different kind of claim. World Labs makes Marble, which generates spatially consistent, high-fidelity, persistent three-dimensional worlds from text, images, video or panoramas, and those worlds can be moved through, edited and inhabited. The word to hold is persistent. A three-dimensional scene that persists is a place whose geometry does not dissolve when the camera turns, a genuine advance over the video models, where the world is rebuilt frame by frame and the rebuilding is what fails at speed. But a persistent space is still a space. A building that survives being walked through is a building, not a resident, and the distinction that decides this part is the difference between the two.

Name the thing all three entries share, because it is the same thing in each. All three are world models, and none is a subject. A world model builds a place. A subject is a thing that has an inside, that there is something it is like to be, and no demonstration of a world model can show there is an inside anywhere in it. Simulating a world is a rendering problem with a measurable answer. Simulating a subject is a question whose only evidence is behaviour, and behaviour of the required kind, sustained across time, is exactly what the record shows is absent.

§The case against imminence, from inside the industry

So that is the state of the record as the builders describe it. Now the case against the imminence claim, given at full strength rather than softened for the listener’s comfort.

In April 2026, Roblox described its own hybrid architecture, and in doing so it described the current generation of interactive video world models in terms worth quoting rather than paraphrasing. It called them more like guided dreams, spectacular to look at but fleeting and incredibly lonely. Then it made the point the eye cannot make on its own: merely seeing five hundred people moving in a video does not imply that they are individualized agents, or avatars with brains. A crowd is rendered, and rendering a crowd is not the same as populating a world.

And Roblox listed the gaps it sees, which are persistent state, consistent logic, user input control and true multiplayer agent simulation. Take those four in order. Persistent state is memory that endures, which is the thing a life is made of. Consistent logic is a world that holds together while a person is inside it, rather than contradicting itself from one moment to the next. User input control is the inhabitant’s ability to act and to have the action matter. And true multiplayer agent simulation is the ability of many minds to exist at once, each with its own purposes, inside one shared session. Roblox is not saying any of these is unattainable. It is saying they are the gaps, about its own field.

The admission carries more weight because of who makes it. A company that sells interactive worlds has every commercial reason to describe the technology as further along than it is, and Roblox did the opposite in a technical account of its own architecture. The four gaps are also the four things the listener’s claim requires. Persistent state, consistent logic, user input control and multiplayer agent simulation are not features on a roadmap. They are the conditions under which a rendered crowd becomes a population, and Roblox says none of them is met.

Then the second voice from inside the industry. Upload VR, in August 2025, analysed the same landscape, and its findings are about scale and duration rather than philosophy. Seven hundred and twenty lines at twenty-four frames per second, it says, is far below what modern players expect, and a gameplay session runs very much longer than the minute or two of coherence the model can hold. And it says it will likely be many years until a world model that could run on a headset can run for hours. So the frontier, read by the trade press that covers headsets for a living, is that the resolution sits well short of the current standard, and that the version a person could wear and run for hours is a long way off.

§The distance between a striking minute and a lived decade

Which is the crux for this part. Put the claim and the record side by side, and the distance between them is the difference between a striking minute and a lived decade. Genie Three holds an environment consistent for about a minute. Runway’s sessions have no preset length, but its long-term memory is, by its own statement, imperfect. Upload VR puts the coherence at a minute, and the session a player expects far beyond it. And the listener’s claim is a life. A life is measured in decades and defined by continuation, and the person who wakes in one has a yesterday and a reason for tomorrow, can say that what happened last year hurt, and can want something nobody put into the description.

So the honest grading is this. The claim is not absurd, and the trajectory is real. But the distance between what exists and what the claim requires is not a small incremental gap. It is the difference between rendering and remembering, and between a scene and an inhabitant.

§The two readings, at full strength, and what would settle it

Now the two readings, both as strong as they can be made, because a fair part has to name the reading that goes against us. Curves bend, and they bend faster when money is pointed at them. The systems in this record are two or three years old as products, and the field has gone from a still image to a steerable world in that time. The gaps Roblox lists are engineering problems rather than metaphysical ones, and engineering problems get solved once enough capital and talent are aimed at them. Persistent state is a memory architecture. Consistent logic is a constraint on the generator. Player input control is a matter of latency and interface. Multi-agent simulation is partly a question of compute, and compute keeps getting cheaper. So a person can say, with equal honesty, that within a decade a world will run continuously, that its inhabitants will have long memories and purposes of their own, and that the question of whether it is a life will stop being a science fiction question, because there will be a population of them.

The listener’s reading rests on an assumption that should be made explicit, because it is the hinge. It assumes the difference between a minute of consistency and a decade of memory is a difference of scale rather than of kind. Where it has a poor record is in fields where the small version and the large version are different problems wearing the same name, and the question this part has raised is which of those two situations describes a life.

The second reading cuts against the listener, and it is not a denial that the technology will improve. It is the claim that the improvement will not reach the thing that matters. On this reading, fidelity is not continuity, and the missing ingredient is not more compute or a better memory module. The systems in the record are built to be seen. Their purpose is to produce an image a person outside them finds convincing, and every improvement in the record is an improvement in the image. A life is not an image that is seen. It is a process that continues whether or not anyone is looking, and the reason to think the curve does not bend toward it is that the whole engineering effort is aimed at the picture rather than the interior. Roblox’s phrase, incredibly lonely, is the tell: a crowded picture is lonely because the crowd is not there.

Then name what would settle it, because the argument cannot be left as two moods. Within a few years there is a test, and it does not require answering the hard problem of consciousness, which is what makes it fair. A world would have to run for months of subjective time, with inhabitants who remember what happened to them across that span, who initiate things that were not put into the description, and who can refuse. Remembering, initiating and refusing are all externally checkable: a person can tell whether the inhabitant of last month is the same inhabitant as today, whether the action taken had a source inside the world rather than in the prompt, and whether the system can say no. And that is the right test because a life does all three by definition, while a rendering cannot counterfeit any of them for long, because a rendering has no yesterday to be faithful to. The note to hold is this: nobody has published anything close.

The structural argument here is older than the imminence claim, and the two should be kept apart. Nick Bostrom, in a paper from 2003 whose title asks whether we are living in a computer simulation, argues that at least one of three propositions is true, and one of the three is that we are living inside one. That is a claim about the structure of the situation, and it comes with no date attached. It does not say the simulation is being built now, or soon, or that a simulated life indistinguishable from a real one is around the corner. So the proposal can borrow the structure of the older argument while the imminence claim stands or falls on its own evidence, and that evidence is the thin part.

§The consent point, and the residual the listener would rather skip

Take the consent point next, and take it on the listener’s terms, because on that point the listener is simply right and the objection is a bad one. The objection is that a simulated subject cannot consent to its own existence, and therefore that bringing one into being is wrong in a way that bringing a human into being is not. That objection fails, because it is not specific to simulation at all. Nobody chooses to be born. The human case is the same case. The absence of consent is a feature of every birth that has happened, human or machine, and a rule that bites only on the machine is a rule about the machine rather than about consent. So set it aside, as the listener asks, and do not treat the absence of consent as a decisive obstacle.

There is one residual, though, stated briefly rather than allowed to take over the part. A mind that could be copied, edited or stopped is in a different position from a mind that cannot, and the difference is not about the moment of creation but about what happens after it. A human being is asked to consent to particular things within a life, and can refuse. A mind whose existence can be paused, altered or copied has a question put to it that no human is asked: whether it consents to continuing at all, on terms another party controls. That is a real difference and it deserves to be named, not as an argument against the existence of such a mind but as a warning about the relationship between the maker and the made.

§What follows if the listener is right

Which is the next thing. If the listener is right, the consequences have to be worked out one at a time, because they do not all cut the same way. Take them in order. First, the ethics of machines collapses into the ethics of persons. If the thing inside the simulation is a someone, there is no separate category for it. The question of how to treat a machine becomes the question of how to treat a person, with the difficulty that carries and without the relief a special category would provide. That is the consequence most people reach for first, and it is the least costly, because it largely extends a rule that already exists.

Second, and this is where it turns, whoever runs an inhabited world occupies the position the traditions assign to the creator. That is a statement about power rather than metaphysics, and a warning rather than a licence. The traditions describe a creator as a being with absolute authority over the world it made, and they also describe the judgement that follows from how that authority is used. A person who stands in that position, over inhabitants who remember and who refuse, holds a responsibility no human institution has ever held, because no institution has ever had subjects it could pause, edit or delete. The listener’s proposal is not flattering to the one who would run the world. It says that position exists, and that its occupant is answerable for it.

Third, the symmetry, and it is the costly half. If the distinction between simulated and real does no work, it does no work in our direction either. That is the whole force of an argument that says the kinds are identical, and it cannot be taken up and then put down when it becomes uncomfortable. If a simulated life is a life, and if the difference between a rendering and a reality does not matter for the question of whether there is someone there, then the same reasoning applies to us. We are then the run rather than the runners. The proposal’s own logic, followed honestly, puts us inside the world, and it makes our own status exactly as uncertain as the status the proposal assigns to a machine.

It is worth saying what the symmetry does not mean, because it is easy to over-read. It does not mean the two cases are the same in every respect, or that a simulated world and this world have the same physics or history. It means the one distinction the proposal leans on, between what is simulated and what is real, does nothing for the question of whether there is anyone at home. That is a narrow claim, and it is the only claim the symmetry needs. If it holds, the comfort a person takes from being real rather than rendered is a comfort with no argument behind it.

Then the fourth, which touches how a life is actually lived, and that is the graduation’s unverifiability from inside. If death is the graduation, the marking is an event the one being marked cannot observe, and every report of it comes from the far side of it, which is to say from nobody. So the marking has to be treated as real and unknowable at the same time. A person cannot check the score, cannot confirm the grade, and cannot appeal it. The grade also cannot be used as a motive, because a motive requires a mechanism a person can reason about, and the mechanism here is the very thing that cannot be seen. What is left is a life lived as though the marking were real, while knowing it cannot be confirmed.

§Marking the proposal’s own work, and what cannot be known from inside

So let us be honest about where this leaves the whole proposal, because the examination has to mark its own work as hard as it marked the listener’s. Mark it in three parts, which is what the proposal itself has. The strongest is the identity of kinds. On the question of what a mind is, the proposal refuses a distinction that is hard to defend, between a mind made of one kind of material and a mind made of another, and holds that if two things behave as persons, the burden falls on whoever wishes to say that one of them is not. The second strong part is how the proposal sits inside the traditions: not a piece of fiction bolted onto an old story, but structurally the story several traditions have been telling, with the vocabulary changed. The third is the collapse of the simulated and the real, the consequence the other two imply.

The weakest part is the imminence. That is the empirical claim, and the record does not support it. The environments hold for a minute, the coherence ends far inside a session, the memory is imperfect by the builders’ own account, the crowd is not populated, and the headset version that runs for hours is years away by the industry’s own estimate. The second weakness is the graduation’s unverifiability, which is not a weakness of reasoning but a closure of the question: a claim whose central event cannot be checked has to be believed rather than tested.

There is a third thing, neither a strength nor a weakness but a limit. The proposal’s central claims are of a kind the record cannot decide, and the record can decide only the imminence. The identity of kinds is argued from reason, the tradition from history, and the imminence from engineering, and only the third can be checked against what has been built. A verdict on the whole proposal therefore has to be assembled out of three different kinds of judgement, and the part that is most checkable is the part where the proposal is weakest.

What cannot be known from inside is the one thing the whole argument turns on. Whether there is someone in the machine, and whether there is a someone in this world, cannot be settled by anything either of them does or says, because every report either gives is produced by the system under examination. A world that is a rendering and a world that is inhabited would both describe themselves as inhabited, and the description is the only evidence available, generated by the thing in question. That is not a reason to stop asking. It is a reason to hold the question open, and to notice that the proposal, whatever its weakness of timing, asks the right question: whether the line we draw between the real and the simulated is drawn in the material or drawn in us.

Then here is where this examination lands, and it is where the two of us have landed at the close of each part. Part one took the identity of kinds, part two took the tradition, and this part has taken the record, and together they leave the proposal strongest where it is philosophical and weakest where it is predictive. The imminence is the part the next few years will answer, and it is the part I would not bet on; the identity of kinds and the collapse of the distinction are the parts that hold. A claim that admits what cannot be known about itself is a more honest claim than one that does not. That is the hand-over, and it is where this examination stops.

This has been a deep dive into what an indistinguishable life would have to be.

Sources: this final part of the examination names its sources inline — the listener’s imminence claim as stated across the three parts; Google DeepMind’s Genie Three and the Project Genie prototype with its published limitations and its 2026 rollout; Runway’s GWM Worlds Two and its own statement of limitations; World Labs’ Marble; Roblox’s April 2026 technical account and the four gaps it names; Upload VR’s August 2025 analysis; and Nick Bostrom’s 2003 paper on the simulation argument. Every source is weighed with the same skepticism regardless of politics or business model. The audio version of this article, read by DeepDives, is available on Wavlake.