A few months ago someone showed me a video of a public figure saying something monstrous. I will leave the figure and the words out of it, because what matters here is not the scandal but what happened in my own mind in the next half-second. Before I had finished hearing the sentence, before I had decided anything, a thought had already formed: that's probably fake. Not a conclusion I reached. A reflex. And I should be honest about what that reflex was doing, because in the moment it felt like caution — like the healthy refusal to be taken in that five nights of this series ought to have trained into me. But it was not caution. It was closer to relief. If the video was fake, I did not have to deal with it. I did not have to believe something hard about a real person. The doubt gave me a way out.
And then a second thought arrived. What if it was real? What if a person really did stand there and say that monstrous thing, on the record — the one situation that, for the whole of my life and my parents' lives, a liar could not escape, because there it was, on tape, in their own face and their own voice? And I had waved it away, reflexively, comfortably. Because I now live in a world where that's probably fake is always available and always plausible — and available not only to me, avoiding a hard truth, but to the person in the video, avoiding accountability. I never said that. It's a deepfake. A sentence that ten years ago would have convinced no one, and now convinces almost anyone, because everyone knows the fakes are good and getting better.
Here is what I think I had run into, and I want to put it as a suspicion, not a verdict, because testing it is the whole of tonight. The deepfake's worst damage may not be the lie it tells me. It may be what its mere possibility does to the true recording standing next to it. A fake does not have to fool anyone to do this damage; it only has to be possible. Once it is possible, a true recording can be waved off as a fake, and no one can quite prove otherwise. A recording used to be the thing that could end an argument. It may have become, without anyone announcing it, one more thing to argue about. Whether that is actually true — whether the recording has really lost its old power, or whether the feeling is a panic I am talking myself into — is exactly what two philosophers will dispute tonight, and I will tell you now that I cannot yet say which of them is right.
That is the question I mean to pursue to the end tonight. On the first night of this arc I said the panic had two halves, and we have only dealt with the first. The first half was: the fakes will fool me into believing false things. That one we have spent five nights taking apart — the verification crisis, which assumes you were once standing on ground you had checked for yourself, an assumption we dismantled. But there was a second half, which I raised on the first night and have not returned to since: that once you know the fakes exist, you begin to disbelieve true things. And not in your own private store of knowledge — we dealt with that — but in the shared, public store: the evidence we all rely on when we say there's footage, it's on the record, I have it in writing.
This is Philosophy for Us — philosophy for everyone, no degree required. This is the seventh of eight episodes on how you know anything you did not see for yourself, in a world where recordings have begun to lie. Last night we asked what trust is, and found that it is not a score you keep on a person but a relationship you enter, in which someone stakes themselves and you let them — a relationship a forgery imitates without being able to supply. And we ended at a limit: that relational account goes silent right where most of what you know comes from — the textbooks, the institutions, the records, the vast impersonal machinery in which no one is turning toward you and telling you anything. The intimate case is clear; the public domain is not. Tonight we go into that public domain. The philosopher who set off this whole panic was never really talking about your sister on the phone. He was talking about something larger and colder, standing outside every particular relationship of trust — and he says it is failing beneath us. Tonight we set out, as precisely as it can be put, what he means. And then we ask whether he is right, because a philosopher standing next to him says he is not.
Let me begin by being fair to the version of the fear we have already dealt with, because I do not want to smuggle it back in.
The naive version of the deepfake panic runs like this: the fakes are getting so good that I will not be able to tell them from the real thing, so I will be fooled, so I cannot know what is true. We have spent five nights showing why that version, on its own terms, rests on a false assumption — the assumption that before the fakes arrived you were reliably telling true from false by checking for yourself. You were not. Almost everything you know reached you on someone's word, unchecked, and you managed. So "I might be fooled" cannot be the deep harm, because being unable to verify for yourself was always your condition. The deepfakes did not create it; they only made you notice it. If that were the whole threat, the honest reply would be: this is the situation you were always in.
But it is not the whole threat, and tonight I have to show you the part that remains after the false assumption is removed. To see it, stop thinking about the lie the fake tells and consider instead what the fake does to the truth beside it. The worst thing a good forgery does is not done to the false thing. It is done to the true thing.
Consider what a recording was, in terms of knowledge, before any of this. For about a century and a half, a photograph, a recording, a piece of film was a particular kind of thing. It was, in a precise sense, a high-information source. A philosopher named Don Fallis has built the clearest account of the threat out of just this idea, in an essay published in 2021 called "The Epistemic Threat of Deepfakes." The idea, stated plainly, is this. A source of information is valuable to the exact degree that it discriminates — that it comes out one way when the world is one way, and the other way when the world is otherwise. A thermometer is useful because it reads high when it is hot and low when it is cold; a thermometer that read high whatever the temperature would tell you nothing. For a hundred and fifty years a recording was an extremely good discriminator. If footage of an event existed, the event had almost certainly happened, because more or less the only way to obtain footage of an event was for the event to occur in front of a camera. The existence of the recording was tightly bound to the reality of what it showed. That binding was the whole of its value. It is what made a recording able to end an argument: not that it was vivid, but that it was hard to fake, so its presence was strong evidence that the thing was real.
Now consider what a good-enough forgery does to that. It does not simply add one false recording to a world of true ones, the way a single counterfeit note joins a supply of real ones. It severs the binding. It makes it possible — cheaply, undetectably, routinely — for footage of an event to exist whether or not the event happened. And the moment that is possible, the recording stops discriminating. It now comes out the same way — here is footage — whether the world is one way or the other. A source that comes out the same way regardless of the facts is, in the precise sense, a low-information source. It has been drained: not turned into a liar, but reduced to telling you nothing. The thermometer now reads the same number in the heat and the cold. Fallis's point is not the obvious one, that there will be convincing fakes. It is the deeper one: that the arrival of convincing fakes lowers the value of every recording, including all the true ones, because what gave a recording its worth was never the image itself but the reliability of the link between the recording and the reality — and it is that link the technology cuts. The real footage is still real. It has simply lost the power to prove that it is real, because it now exists in a world where footage of that kind no longer means what it meant.
This is the second half of the fear I raised on the first night, now named and now precise. The damage is not in the fakes you believe. It is in the truths you can no longer use. The deepest casualty of the deepfake is not the person fooled by a false video; that is the shallow harm, the one the naive panic fixes on. The deepest casualty is the true video that can no longer do its job — the genuine recording of the real crime, the real promise, the real words, which now arrives in a world where everyone knows such things can be faked, and so persuades no one, ends no argument, and saves no one. The liar was once caught by the tape. Now the liar has a new and permanent move, and it is the most damaging sentence there is: that's a deepfake. He does not even have to produce a fake. He only has to point at the possibility of fakes, which costs nothing and which we have all agreed is real, and the true recording of what he actually did becomes one more contested thing. Two legal scholars, Robert Chesney and Danielle Citron, named this in a 2019 article: the liar's dividend. The liar is paid, automatically, by the mere existence of the technology, every time the truth appears on tape. He collects it without lifting a finger.
So hold the shape of this clearly, because the rest of the episode builds on it. The threat that survives five nights of disarming is not I will believe a lie. You were always exposed to lies; trust was always a bet. The threat that survives is I will lose the ability to be corrected by the truth — that a particular instrument, the recording, which for a century and a half could enter a dispute between a liar and the truth and settle it on the side of the truth, is being quietly disabled. And it is an impersonal instrument. This is not about your sister, or anyone turning toward you and staking themselves. It is about the impersonal public machinery that stood outside all of that and held it up. That is the impersonal domain we reached at the end of last night, where the relational account of trust went silent. Tonight's philosopher goes straight into it, and says: that machinery had a name and a job, and it is what we are losing. That is where we turn next.
The philosopher is Regina Rini, and the essay, published in 2020, is called "Deepfakes and the Epistemic Backstop." The whole weight of the pessimist's case rests on that one idea, so let me build it carefully and at full strength before anyone tests it.
Start with a fact about how we live: we run almost entirely on each other's word. That has been the central claim of this whole arc. You believe what people tell you, mostly, because you have to — checking everything yourself is not available to one person facing a whole world of facts. But now ask a question we have not asked. If we run on each other's word, and people sometimes lie, what stops the whole practice from collapsing into unaccountable claims that no one can check? What keeps testifiers — especially public ones, whose word reaches millions — even roughly honest? It is not that lying is difficult; lying is easy. Something must be holding the practice up from outside, some standing pressure that makes honesty the safer choice for most people most of the time. Rini's answer is that for the last century and a half, one of the main things holding it up has been the recording — not because we constantly consult recordings, but because the recording stands behind the whole practice as a backstop.
The word is exact, and worth taking literally. A backstop is something that stands behind an activity and catches what gets past everyone else; you do not consult it or use it directly on most occasions, but its presence changes how the activity goes. Rini's claim is that recorded media has served as the backstop behind testimony in two distinct ways, and both matter. The first is correction after the fact: when a powerful person lies, the recording can catch them. It is the thing you hold up against their word — here is your own face, here is your own voice — and it ends the denial. That is the obvious one. The second is deeper, and does most of the work, and it operates even when no recording is ever produced. It is deterrence. Because a public figure knows a recording might exist — anyone might have a camera; the record is a standing possibility — they are constrained in advance, all the time, by the mere chance of being caught on tape. The backstop disciplines the liar who is never actually caught, in the way a security camera changes the behaviour of people it never films. Most of the honesty the system gets out of powerful people, Rini argues, it gets not by catching them but by the standing threat of catching them. The tape that is never made is doing work every day.
Now you can see what the backstop is, and that it is exactly the impersonal thing we reached at the end of last night — the apparatus standing outside any particular relationship of trust, holding the whole structure up. No one is turning toward you; no one stakes themselves. It is public, mechanical, and indifferent, and it has been regulating the honesty of everyone whose word you depend on and will never meet. And here is the pessimist's claim, stated plainly: the deepfake does not merely add some false videos. It removes the backstop. It takes away both jobs at once. It ends correction, because the decisive recording can now be waved off as a fake — that is the liar's dividend from the last segment. And it ends deterrence, which is worse, because the powerful person no longer has to fear the tape: whatever surfaces, they can deny, and the standing threat that kept them roughly honest simply evaporates. Remove the backstop from behind the practice, and the practice does not collapse tomorrow — a backstop is invisible until it is gone — but the thing that disciplined it from outside is no longer there, and in time the liars learn as much.
And here is why this is the hardest form the fear takes — the one everything we have built has to answer. Grant everything we have built. Grant Reid: that trusting is the default, the thing you cannot get beneath. Grant Moran: that when someone tells you something they stake themselves, and that is a relationship a forgery cannot reach. All of that concerns the intimate case, the person turned toward you. But the backstop was never in the intimate case. The backstop was the public, impersonal apparatus that let you trust well in the vast domain where no one is turning toward you — where you depend on strangers, institutions, officials, broadcasts, records. The relational account said: trust is unavoidable, and the real question is whether you trust well. Granted. But trusting well, out in the public domain, was never simply a private virtue. It had material conditions. It relied on there being some way, somewhere in the system, that a public liar could be caught and knew it. And the pessimist's claim is that this specific condition, which made trusting well in the impersonal domain possible at all, is being removed. Trust as the default still holds. But the public apparatus that made trusting well possible on top of that default — out where you deal with strangers and institutions you cannot check — may be failing.
And do not reach here for the obvious remedy — the technical fix, the detector, the forensic tool that catches the fake. The pessimist has an answer, and it is grim and it is correct. There is no detector a good-enough forgery cannot pass; and even if there were, you would then be trusting the detector — one more instrument, one more thing built and vouched for by people you cannot check, one more piece of testimony at the bottom of the pile. It is a contest between forgers and detectors that the forger is structured to win, because the forger has only to defeat the detector once, whereas the detector must catch every forgery forever — and the moment a detector becomes reliable, it becomes the exact training target that teaches the next forgery to beat it. You cannot engineer your way back to the backstop. What made the recording trustworthy was never a feature you can rebuild; it was the difficulty of faking it, and difficulty is precisely what the technology removes. So the pessimist rules out each escape in turn and leaves you in the open: the backstop is going, there is no replacing it, and the deepest harm is not the lies you will believe but the truths you will lose the power to defend.
That is the case, complete. It is not a moral panic. It is precise, it is mechanical, and every step follows. Let it stand, because now — and only now that it is standing — I will turn you to the philosopher beside Rini, who has heard every word of it and says the recording was never the backstop she takes it to be.
His name is Joshua Habgood-Coote, and his essay, published in 2023, carries a title that announces his attitude to the panic: "Deepfakes and the Epistemic Apocalypse" — apocalypse said the way you would say it of someone who has predicted the end of the world a few too many times. I want to be clear about what he is and is not doing, because it is easy to mishear him as the reassuring voice, the one who says it will all be fine — and if you hear him that way you will badly underrate him. He is not offering comfort. He is making an argument, a hard one, and it deserves exactly the weight I have just given Rini. He is not saying deepfakes are harmless. He is saying that the backstop story — the specific claim that recordings were load-bearing in the way Rini needs, and that removing them brings the structure down — is overstated, probably wrong, and that believing it may be the most dangerous part of the whole situation. He has three objections, and I will give you all three, each at its strongest.
The first objection is historical, and it is the simplest and the hardest to answer. We have always lived with fakes. The premise the apocalypse rests on — that we are leaving a golden age of reliable recordings and entering a new dark age of forgery — is, he says, a story we tell ourselves that does not survive contact with the actual past. Photographs have been staged, doctored, and faked since almost the moment photography existed. The Soviet authorities removed people from official pictures. Propaganda ministries have produced convincing falsehoods for as long as there has been mass media. Confidence men and forgers are not an invention of the 2020s. And we coped — not perfectly, but we coped, and we did it the way we always have: not by trusting recordings blindly because they could not be faked, but by asking where a thing came from, who vouches for it, whether it fits what else we know, whether the source has been reliable before. Competent knowers were never naive about media; they routed the recording through the same social and institutional checks they applied to everything else. So the deepfake does not take us from safety into catastrophe. It intensifies a problem we have always had and always had ways of handling. The novelty is real, but it is a difference of degree, and we have absorbed differences of degree before.
The second objection goes straight at the backstop, and it is the one that goes deepest philosophically. Was the recording ever really the backstop Rini describes? Consider how you actually come to trust that a piece of footage is genuine. You almost never authenticate it yourself; you could not, lacking the tools and the access. You trust it because of where it came from — a news organisation with a reputation to lose, a court that admitted it through a documented chain of custody, a process with people and institutions standing behind it. In other words, what made a recording credible was never the bare physical fact that it was hard to fake. It was the web of institutional and social vouching around it — which is precisely the testimonial trust we have been discussing all along. The recording was always already embedded in trust; it was never a trust-free bedrock underneath trust. And if that is right, the backstop was never a separate, load-bearing foundation the deepfake can knock out. It was one more thing held up by the same web that holds everything else up — and that web is still there. The decisive recording never worked by being unfakeable; it worked by being vouched for. Deepfakes make the vouching harder, yes, but they do not remove a foundation, because the foundation was social trust all along, and social trust has no single technological support you can kick out from under it.
The third objection is the one that should genuinely unsettle you, because it turns the pessimist's own warning into the danger. The deepest threat, Habgood-Coote argues, may not be the fakes at all. It may be the panic about the fakes. The mechanism is this. The actual number of consequential, world-changing deepfakes has so far been far lower than the apocalypse predicted; the technology is real, but the great volume of perfect, undetectable, history-altering forgeries has largely not appeared in the way the loudest warnings said it would. What has spread, fast and everywhere, is the belief that you can no longer trust anything you see. And that belief does the damage directly, without a single fake being made. It is what let me wave away the video in the cold open — not a deepfake, only the idea of deepfakes, doing the liar's work for nothing. A population told that nothing can be trusted does not become careful; it becomes cynical. And a cynic is not harder to manipulate but easier, because he has given up on the difference between true and false and will believe whatever flatters him. The collapse of trust, if it comes, will not be something the deepfakes did to us. It will be something we talked ourselves into — and those who prophesy the apocalypse, however well-meaning, are part of how we talk ourselves into it. The warning does not merely describe the problem. It may help cause it.
So stand now with both of them, because this is where I have to leave you, and I will not pretend it resolves. Rini says: a real and specific apparatus that disciplined public honesty is being removed, the harm is the true things we lose, and there is no rebuilding it. Habgood-Coote says: that apparatus was never the freestanding thing you take it to be, we have always coped with fakes by other means, those means are still here, and the real wound is the despair the panic itself produces. And here is the honest and uncomfortable state of the field: both positions are live, in print, defended by serious people, and neither has been refuted. This is not a settled question I am dramatising. It is genuinely open, and open partly because it turns on facts that are not yet in — how good the fakes actually get, how fast, how far the institutions adapt, how people actually behave when they cannot trust the footage. Philosophy has done its job, which is to show you what the two of them disagree about and why it matters. It cannot do the other job, the one you keep wanting done, which is to tell you which of them is right. No one knows yet. And a show that told you it knew would be doing to you precisely what one of these two thinkers says is the real danger.
So where does tonight leave you? I have no settled answer to give — the whole point is that we cannot yet tell whether the apparatus out here is failing or intact.
What tonight did was take the crisis we began with and finally state it in the form that survives. The naive fear — the fakes will fool me — we disarmed five nights ago; you were never the auditor who could not be fooled, so losing a security you never had is no catastrophe. But beneath it lay a second fear, which does not dissolve, and tonight we named it. It is not that you will believe false things. It is that true things may lose their power to correct false ones — that a specific, impersonal, public instrument, the recording, which for a hundred and fifty years let a liar be caught and, more importantly, let a liar fear being caught, is being drained of its force, with no way to rebuild it once it goes. That is the pessimist's case, Fallis and Rini, and I gave it to you undiluted because it deserves to stand. And then I set beside it Habgood-Coote, who says the instrument was never freestanding, that it was always held up by the same web of vouching that holds up everything we know, that we have coped with fakes since the camera was invented, and that the real damage is the despair the warning itself spreads. And I told you the truth about where that leaves us: in the middle of a live dispute, with no arbiter, on a question that turns partly on facts that have not yet arrived.
Now hear what I am not doing with that, because the temptation runs two ways and both are dishonest. I am not handing you the apocalypse — the backstop is gone, the public record can no longer be trusted. I cannot hand you that, because the strongest objection to it stands right there unrefuted, and someone who waved Habgood-Coote away to preserve the drama would not be rigorous, only a seller of alarm. But I am also not handing you the reassurance — relax, we have always coped, it is only panic. I cannot hand you that either, because Rini's mechanism is real and precise, and the danger she describes is genuinely frightening, and someone who reassured you to make the discomfort stop would be selling you the cheaper thing, comfort. Both are the same error: ending the argument early because standing in it is unpleasant. The argument does not end tonight. It does not end because it is not finished — not in the journals and not in the world — and I will not manufacture a result.
And notice — because this matters, and the careful listener has earned it — nothing tonight cancels the checker. Last night I said that the person who asks for the second source, who will not take an extraordinary claim on ordinary word, is doing real work, and nothing tonight retracts that. If anything, the pessimist gives that person more to do, not less: in a world where the recording cannot vouch for itself, the patient work of the careful knower — where did this come from, does it fit — is more valuable, not obsolete. What has died, if Rini is right, is not checking; it is one particular thing that used to do some of our checking for us, automatically and in the background. The work returns to us. The checker is not the fool here. The fool is the one who thought the backstop meant he never had to be the checker.
Two questions I will point at and not walk through, because they belong to other arcs and entering them tonight would cheat you of the real ones. There is a question inside all of this about whose word is believed in the first place — the one I marked two nights ago, about whose testimony is treated as credible and whose is doubted by default, and what that has to do with power. That is real, and large, and not for tonight; I am keeping it whole for its own week. And there is another question about whether you are to blame when a fake fools you — how much you owed it to look closer before you believed. Also real, also not tonight. Tonight was deliberately narrower than either: not whose word counts, not whether you are guilty, but only this — whether a specific apparatus that made trusting well possible out in the impersonal public domain is failing, and whether it was ever bearing the weight we placed on it.
What you carry out is this. The next time you reach for it — and you will, this week, almost certainly — the next time you settle an argument or steady yourself with there's footage, it's on the record, I saw the video — stop for a moment on what you are leaning on. You are leaning on the backstop. You are trusting that the recording still means what recordings used to mean. And you cannot, in that moment, tell whether that support is still under you or has gone out from under you — whether you are Rini's knower, reaching for an instrument being disabled, or Habgood-Coote's, reaching for something that was always just held up by trust like everything else and is no worse off than before. You will not know which. And that not-knowing — not everything is fake, and not nothing has changed, but you can no longer be sure the old reflex is safe, and you reach for it anyway — is the actual condition the deepfake era puts you in. Neither thinker has earned the right to carry you out of it early. Not yet. The dispute is not finished, and neither are you.
Because next time is the last night, and next time I do not bring you a new rival — I bring all of them into one room. The one who said verify everything, and cannot. The one who said trust is the ground, which it is, but which is blind on its own. The one who said trust is a relationship someone stakes themselves in, which the forgery imitates and cannot supply. And tonight's, who said the public apparatus around the whole thing is failing — perhaps. Four answers, no winner, and the question you walked in with on the first night — the phone ringing at eleven — handed back to you sharper than you can put it down. I will not resolve them for you next week either. I will do something harder.
Tonight ends out in the impersonal public domain, where the relational account went silent, where no one is turning toward you — reaching, as you have reached your whole life, for the one impersonal thing that used to catch a liar, and unable to tell, as you reach, whether it is still there. That is not a comfort. It was never going to be a comfort. But it is exactly where the honest form of the fear has been waiting for you since the first night, and now you are standing in it with your eyes open.
Thanks for listening. I'll see you next time.