For about a decade I had a party trick, and the party trick was a misquote.
Somebody would raise it — usually late, usually after the second bottle — and I would lean in with the authority of a man who has read the abstract. An Oxford philosopher, I would say, has shown that we are probably living in a computer simulation. Then I would do the arithmetic bit: any civilization advanced enough to run detailed simulations of its own ancestors would run billions of them, so the simulated vastly outnumber the real, so what are the odds you're one of the originals? I enjoyed the silence that followed. I got very good at the pause.
Then, embarrassingly recently, I read the paper.
It does not say that. It says something narrower, stranger and much more interesting, and the gap between the two is not a rounding error — it is the whole point. What I had done was take a carefully fenced three-way disjunction and ship it as a conclusion, which is precisely the move I spend my working life complaining about when a press office does it to a study. I had been doing it at dinner, for fun, for ten years.
So let me correct it properly. And then let me explain why I now think the metaphysical question — the fun one, the one you clicked for — is a decoy, and the version of it that is actually going to touch your life has a date on it. That date was seventeen days ago.
What the paper actually says, and what I said it said
The philosopher is Nick Bostrom, and the paper is from 2003. It argues that at least one of three propositions is true: that the human species very likely goes extinct before reaching a "posthuman" stage; or that a posthuman civilization is extremely unlikely to run a significant number of simulations of its evolutionary history; or that we are almost certainly living in a computer simulation.
Read that again, because the structure is the argument. It is a trilemma. It says one of these three. It does not say which. And Bostrom himself, in the same paper, declines to pick: in the dark forest of our current ignorance, he writes, it seems sensible to apportion one's credence roughly evenly between the three. Roughly a third. Not "probably."
There is even a tidy corollary he draws that almost nobody repeats, and it is the sharpest sentence in the thing: unless we are now living in a simulation, our descendants will almost certainly never run one. Either we're inside, or nobody ever builds it. That is a genuinely unsettling claim, and it is not the claim I was making over dessert.
What I was making was the popular version, which has been sanded down through twenty-three years of retelling until it fits in a headline — the framing that ends up as roughly even odds tilting toward base reality by a smidgen, which at least has the honesty of showing its Bayesian working. Mine had no working. Mine had a pause.
Physics took two real swings at it
Here is the part I did not expect when I started reading: serious physicists have actually tried to test this, and both times the press coverage hardened the finding on the way out the door.
In 2017, two physicists, Ringel and Kovrizhin, looked at whether certain quantum systems can be simulated efficiently on classical computers at all. Their answer, in Science Advances, was technical and careful: it is believed that not all quantum systems can be simulated efficiently using classical computational resources, and they identified a specific obstruction for a specific class of them. That is a real result about computational complexity. It is not a result about the universe. The abstract makes no claim whatsoever about whether we live in a simulation. The coverage announced that "Quantum Weirdness Once Again Shows We're Not Living in a Computer Simulation".
Then in October 2025 a group including Mir Faizal and Lawrence Krauss went further, arguing from incompleteness theorems that reality rests on a kind of understanding no algorithm can reach. Any simulation is inherently algorithmic, Faizal said — it must follow programmed rules — and so the universe cannot be, and could never be, a simulation. Fascinating argument. It is an argument, in a specialist journal, resting on contested premises about what physics fundamentally is. The headline it produced was "Physicists Just Ruled Out The Universe Being a Simulation" — from the same outlet that had ruled it out once already, eight years earlier, on entirely different evidence.
So in one direction we get "philosopher proves we're in a simulation," and in the other we get "physicists prove we're not," and underneath both is careful, hedged, provisional work that says neither. I am not above this. I was this.
The version of the question that has an answer
Here is where things get interesting, and where I stopped caring about base reality.
Ask "is the universe a simulation?" and you get twenty-three years of unfalsifiable elegance. Ask a smaller question — can you tell whether the thing in front of you was generated? — and you get an experiment, a number, and a result you will not enjoy.
Researchers put 1,276 people in front of images, audio, video and audiovisual clips, some authentic and some synthetic, under ordinary online viewing conditions, and asked them to sort real from fake. People struggled to distinguish between synthetic and authentic media: mean accuracy across everything was 51.2%. On video alone it was 50.7%. Images were worse than guessing — 49.4%. Prior knowledge of synthetic media, the authors found, did not meaningfully improve anyone's performance — the practice you think you have been getting is not practice.
The authors' conclusion is the sentence I would put on a wall: human perceptual capabilities can no longer be relied upon as a useful defense.
You do not need to know whether the cosmos is rendered. You are already living inside generated realities you cannot pick out of a lineup, several times a day, on a phone. That question is settled. The only open part is what anybody is obliged to do about it.
Where a coin flip becomes a verdict
Most days, being wrong about a video costs you nothing. Some days it costs someone their liberty.
On February 17, 2026, New York's highest court decided a family court case in which video clips appeared to show an adult sexually abusing a child. In a 4–3 decision, the court concluded that the video clips were not properly authenticated — on several grounds, including hearsay. But the majority, written by Chief Judge Rowan Wilson, went out of its way to say something new: that the increasing prevalence of deepfakes has now reduced the old technique of matching details in a video against what a witness remembers of the room to a more suspect form of authentication. Fabricated video, the court noted, tends to borrow real rooms.
Nobody in that case proved anything was fake. That is what makes it a landmark. The mere existence of convincing fakes moved the bar.
Judge Singas, dissenting, called the majority's treatment of the technology superficial and simplistic, and warned that any litigant can now cry "deepfake" and shift the burden. Judge Paul Grimm, who has spent years on this, puts the underlying problem plainly: the evidence rules set a low barrier — if something is 51% likely to be what you say it is, it comes in, and it was 49% likely not to be.
And the federal rule that would settle who has to show what? It has been drafted. At its last six meetings, the Judicial Conference's evidence rules committee has been working on the problem, and it has prepared, but not approved, a new Rule 901(c) that would make the challenger produce evidence of fabrication before the burden flips back. Its companion, Rule 707, covering machine-generated evidence, drew testimony at two hearings and more than seventy written comments — and on May 17, 2026, the committee reported that it does not recommend action on the proposed Rule 707 at this time, revising it instead and holding it over, together with deepfakes, to the next meeting. Neither rule is in force. The country's answer to "how do we know it's real" is, at this moment, one judge at a time.
Sixteen days ago, one state started requiring somebody to tell you
Then, on August 2, 2026, something did switch on — and almost nobody noticed, because it arrived on the same day as a much louder European deadline.
The California AI Transparency Act became operative. If you run a generative AI service with more than a million monthly users that is publicly accessible in California — which is to say, if you run any of the ones you have heard of — you must now embed a hidden, machine-readable disclosure in every image, video and audio file your system produces, offer users a visible label as an option, and publish a free, public AI detection tool anyone can use to check whether a file came from your system. Penalties run to $5,000 per violation, per day.
That date is not a coincidence. The law was passed in 2024 to start on January 1, 2026, and then amended so that its operative date was extended to August 2, 2026 to align with the EU AI Act, with duties for large online platforms following in 2027 and for capture devices — cameras and the like — in 2028. California deliberately synchronized its watch with Brussels.
So the machines must now sign their work. Problem solved?
Not quite, and the reason is depressingly mundane. The signature lives in the file's metadata — and metadata is the first thing to die in transit. When files are uploaded to social platforms or passed through ordinary sharing systems, they are routinely stripped of their metadata; implementers of the provenance standard have themselves acknowledged stripping as an obstacle to its robustness. California saw this coming, sort of: the statute requires that the disclosure be permanent or extraordinarily difficult to remove — and then adds "to the extent it is technically feasible," which is where the whole sentence goes to sleep. A law that requires a mark at the moment of creation, in an ecosystem that scrubs marks at the moment of sharing, has built a beautiful lighthouse on a coast nobody sails past.
The strongest case against everything I have just written
Now let me argue the other side properly, because it is better than I would like.
In January 2026, the Federal Judicial Center surveyed federal judges nationwide about exactly this. 931 responded — a 45% response rate, which for a judicial survey is enormous. And 98% of them said they had never encountered a challenge to a video, photograph or audio recording on the grounds that it was a deepfake. Fifteen judges had. Of those fifteen, ten had seen it once.
Fifteen. Out of 931. That is the entire observed federal caseload of the crisis I have just spent a thousand words describing.
So the honest reading is available, and I want it on the record: the panic is running well ahead of the docket. Rules written in a hurry against an imagined problem tend to be gamed by real lawyers within a season, which is more or less Judge Singas's warning. The committee's caution may be wisdom rather than sloth.
Two things keep me from resting there. The first is that 55% of the judges who had never seen a deepfake challenge nevertheless said the authenticity rules should be amended — they are not seeing it yet, and they still want the rule before they do. The second is buried in the same survey: a judge reported a case in which an informant pled guilty to obstruction of justice after supplying false audiovisual evidence, and that fabricated evidence had already produced the indictment of an innocent person. The report does not say a machine made it. It doesn't have to. It tells you what the failure looks like when it lands, and it tells you that it lands quietly — that one surfaced at a sentencing hearing, not on any docket about authenticity.
And there is a third thing, which is the argument that actually moved me. Five of the public comments on the machine-evidence rule went out of their way to ask for a deepfake rule the committee had not asked them about — including the Innocence Project, which spends its working life on convictions that turned out to rest on bad evidence. Its point is not that fabricated video is common. Its point is what it would be for: deepfake technology, it wrote, could be used to alter crime scene surveillance footage to place an innocent person at the scene of a crime — a new-age form of misconduct similar to the planting of contraband to incriminate an innocent suspect. Planted contraband was never common either. We wrote rules about it anyway, because of what it does to the person it happens to, and because the rule is what makes it detectable at all.
Beijing put the duty on the pipe, not just the printer
Here is where a comparison earns its keep, because a different government looked at the same problem three years ahead of us and made a structurally different choice.
China's labeling measures were jointly unveiled in March 2025 by four regulators at once — the internet regulator, the industry ministry, the public security ministry and the broadcasting authority — together with a mandatory national technical standard, and they took effect on September 1, 2025. The measures require both explicit and implicit labeling: explicit labels a user can actually perceive, and implicit labels written into the file's data where a user cannot.
That much resembles California. Two provisions do not.
First, the distribution platforms carry their own duty. Under Article 6, a platform must check the file's metadata for an implicit label and put a conspicuous notice around the content when it finds one — and, remarkably, must also flag content when it finds no label but detects other traces of generation, as suspected synthetic content. The obligation follows the file down the pipe instead of stopping at the printer.
Second, under Article 10, the labels must not be maliciously deleted, altered, fabricated or concealed by anyone, and no one may supply tools for doing it. Stripping the mark is itself the offense.
I am not holding this up as a model, and I want to be exact about why. The same apparatus that guarantees you a label on a synthetic video is a general-purpose content-tracking spine, run by a state that uses it for other things; under Article 5, every file carries the name or code of the service that generated it and a content reference number. That is a serious price, and Americans would be right to refuse it. But the architectural insight underneath is free, and we have not taken it: a provenance mark is worth nothing unless somebody downstream is obliged to look for it and forbidden to rub it off.
Now run it forward
Give it seven years.
A year ago Google DeepMind demonstrated a system that, from a text prompt, generates dynamic worlds that you can navigate in real time — 24 frames a second, 720p, holding itself together for a few minutes. Read that as a baseline rather than a ceiling, because it is only ever the worst version you will see.
By 2033 the phrase "recorded footage" starts to sound like "handwritten letter" — a real category, an increasingly ceremonial one. Your insurer's app doesn't ask for photographs of the dent; it asks your phone to reconstruct the collision from sensor logs and renders it, and the render is what the adjuster watches. Your child's school sends a video of the recital your grandmother could not attend, generated from three angles she wasn't there for, and it is lovely, and nobody thinks of it as fake. A politician's genuine recorded confession lands in October and dies in nine hours under a wave of skepticism nobody had to manufacture — the doubt is now ambient, free, and permanently available to anyone caught doing anything.
And in a courtroom in 2033, a public defender stands up and says the four words that will define the decade: prove it isn't generated. On present course, the judge will have no rule to reach for, and will do what judges do — decide it themselves, differently in Buffalo than in San Diego.
What the people who work on this are actually saying
They do not agree, and the disagreement is the useful part.
The provenance camp — the coalition behind the open standard that establishes the origin and edits of digital content — argues that authenticity has to be proven forward from capture rather than detected backward from the artifact. The privacy analysts who reviewed that same standard warn it is not a lie detector: it verifies who signed a file, not whether the file tells the truth.
The Brennan Center, working the democracy side, documents the second-order damage: as people learn how good fakes have become, false claims that real content is AI-generated become more persuasive too — the "liar's dividend" named by law professors Bobby Chesney and Danielle Citron, and the thing New York's dissenters were frightened of.
From the free-speech side, FIRE makes the objection that will decide how far any of this can go in America: developers may watermark all they like, but the First Amendment restricts the government from compelling companies to do so. California's law will meet that argument, probably soon.
And Europe, having made marking a legal obligation, published a companion code in June 2026 to explain how — adherence to the code is voluntary even though the underlying duty is not. Everyone is trying to build the same fence out of whatever their constitution will let them use.
So what does this actually mean for you?
Not much, cosmically. Quite a lot, practically.
Stop treating your eyes as evidence. You are at 51%. The correct question about a striking video is never "does it look real" but "where did this come from, and who else has it." Provenance beats perception now, and it is not close.
Use the detection tools that just became mandatory. Any large generative service available in California must now offer a free public checker. It only tells you whether a file came from that system — a negative proves nothing — but it is free, it takes fifteen seconds, and until August 2 it did not exist.
Preserve originals, not screenshots. If a photo or recording might ever matter — an accident, a dispute, a threat — keep the original file off the phone it came from, unedited, unposted. Every share, crop and re-upload strips exactly the data that would later prove it was yours.
Agree on a channel, not a password. If a video call or voice note asks you to move money or send documents, verify on a different channel you established in advance. The fake will be excellent. The second channel is what you actually have.
Watch what your state does next. California's rule reaches platforms in 2027 and cameras in 2028, and other legislatures copy California. That second phase — obliging the pipe, not just the printer — is the part that decides whether any of this works.
When someone dismisses a recording as AI, ask what their evidence is. That reflex is the liar's dividend arriving at your kitchen table, and the only defense is the boring one: make people show their work, including the people you agree with.
The lesson, as I read it
I spent ten years asking whether somebody else had simulated my reality, and I got the question wrong twice — first by misquoting the man who posed it, then by assuming it was idle.
It was never idle. It was rehearsal. While we argued about base reality, we built the capacity to manufacture convincing ones, handed it to everybody, and only got around to asking who has to disclose it about twenty minutes ago — in one state, with a detection tool that answers a narrower question than the one you'll be asking, in an ecosystem that deletes the evidence on upload.
Bostrom's trilemma says one of three things is true and we cannot tell which. Fine. Here is a smaller one I am confident about: either we build institutions that can establish what actually happened, or we don't, and if we don't, "prove it isn't generated" becomes the last word in every argument that matters — in a courtroom, in an election, at a kitchen table. That one has only two branches, and unlike Bostrom's, we get to pick which.
My vote? Mark everything, oblige the pipe, and keep the receipts.
Somewhere between "none of this is real" and "all of it is" sits the only question with a paper trail: who has to tell you? The HAIA Foundation works on the answer. If you'd rather not find out the hard way, subscribe — and forward this to the person who still trusts a video.