Latest / The BlackVeil Files / The Real Reasons You Can't Prove What's True Anymore | The Shoggoth Has an Alibi

Transcript

The Real Reasons You Can't Prove What's True Anymore | The Shoggoth Has an Alibi · The BlackVeil Files

0:00
SPEAKER_03: Labor Day Weekend 2025.
0:02
SPEAKER_03: Somebody outside the White House is filming the residents when a second floor window swings open and a black bag comes flying out.
0:10
SPEAKER_04: A window is open to the residents upstairs and somebody is throwing a big bag out the window.
0:15
SPEAKER_04: Have you seen this?
0:16
SPEAKER_03: No, that's probably AI.
0:18
SPEAKER_03: The same day, the Pentagon posts a video of an American airstrike on a speedboat in the Caribbean that the administration says was running drugs out of Venezuela.
0:28
SPEAKER_03: Eleven people on board are killed.
0:30
SPEAKER_03: Within a day, Venezuela's communication department, Freddy ⁇ tells his followers that Americans built that video with artificial intelligence, and he uses an analysis from Google's Gemini as his proof.
0:42
SPEAKER_03: Haney Farid, the Berkeley professor who does image forensics, he looks at the White House clip for CNN.
0:48
SPEAKER_03: Nine days after the strike, Venezuela stops talking about artificial intelligence.
0:53
SPEAKER_03: Its interior minister starts asking whether the men on the boat were smugglers at all.
0:57
SPEAKER_03: Eight years ago, two law professors named this phenomenon in a law review.
1:02
SPEAKER_03: The liar's dividend, they called it.
1:04
SPEAKER_03: And still, nobody uses the term, which is a problem because the thing itself has gone from a hypothesis to an everyday occurrence.
1:12
SPEAKER_03: It needs a name that people will actually use because whatever we end up calling it, it's going to become a lot more common than it already is.
1:21
SPEAKER_03: I've spent the last two weeks researching peer-reviewed experiments, court filings, transcripts, and about 40 clips of powerful people saying the same thing the moment a camera catches them doing something that they wished they hadn't.
1:34
SPEAKER_03: And the worst thing I found isn't a fake.
1:36
SPEAKER_03: It's an experiment from 2020.
1:38
SPEAKER_03: Three political scientists put a one-sentence warning above a video, and then they measured how that sentence affected the people who watched it.
1:46
SPEAKER_03: The warning didn't do the job that they were testing for.
1:49
SPEAKER_03: It didn't help anybody spot a fake.
1:51
SPEAKER_03: What it changed was what those people were willing to believe.
1:55
SPEAKER_03: AI synthesized faces.
1:58
SPEAKER_03: Eight faces.
1:59
SPEAKER_03: Half of them are photographs of people, and half of them came out of an AI model called Style Jan 2.
2:05
SPEAKER_03: Pick the four you'd trust with your house keys and remember which one you picked.
2:10
SPEAKER_03: When they ran it properly, one face at a time, people scored 48.2%, a coin flip.
2:16
SPEAKER_03: So they trained a new group, they told them what to look for, then told them after every single face whether they'd got it right.
2:23
SPEAKER_03: That group reached 59%, and on the last face that they judged, they were as wrong as the first group.
2:29
SPEAKER_03: Then a third group didn't guess at all.
2:31
SPEAKER_03: They only rated how trustworthy each face looked.
2:34
SPEAKER_03: In the fakes one, the most trusted face in the entire set was manufactured.
2:42
SPEAKER_03: Faces are the easy version because you get to stare at them for as long as you want.
2:46
SPEAKER_03: But what about voice clones?
2:48
SPEAKER_03: This is from the Queen Mary Study in London.
2:51
SPEAKER_00: Two recordings.
2:59
SPEAKER_03: One of those is a man reading a passage into a microphone.
3:02
SPEAKER_03: The other is a clone of that same man built from four minutes of his own recordings.
3:07
SPEAKER_03: When the researchers played pairs like that to listeners whose only job was to decide what was human, they thought that the clones were human 70% of the time.
3:16
SPEAKER_03: They guessed the actual human beings were human 72% of the time.
3:20
SPEAKER_03: At Berkeley, a separate team played 604 people, a clone, next to a recording of a real speaker.
3:27
SPEAKER_03: Eight times out of ten, people said it was the same person.
3:30
SPEAKER_03: The next one is called Fooled Twice.
3:32
SPEAKER_03: People can't spot a deep fake, and they're sure they can.
3:36
SPEAKER_03: Video is where people get confident.
3:38
SPEAKER_03: A team in 2021 showed people short clips and told them up front that half were fake.
3:43
SPEAKER_03: They still scored barely better than a coin flip.
3:46
SPEAKER_03: Some were offered cash to get it right, and the money didn't change anything.
3:50
SPEAKER_03: Told outright that half of what they were watching was fabricated.
3:53
SPEAKER_03: They still called most of it real.
3:55
SPEAKER_03: The ones who scored worst were the most confident.
3:57
SPEAKER_03: Last year, Economist ran it on 17,000 readers of a German newspaper.
4:02
SPEAKER_03: Three pairs of pictures, one real and one machine made in each pair.
4:06
SPEAKER_03: Pick the fake.
4:07
SPEAKER_03: Almost nobody got all three.
4:09
SPEAKER_03: A third of them got all three wrong, which is worse than guessing.
4:13
SPEAKER_03: So look at what's actually failing in all of them.
4:16
SPEAKER_03: Nobody is being beaten by a brilliant fake.
4:18
SPEAKER_03: They are being told that fakes exist, and that alone is enough to start turning real things into suspects.
4:24
SPEAKER_03: Still, pictures are the same story.
4:26
SPEAKER_03: Economists showed nearly 3,000 British adults a news article.
4:30
SPEAKER_03: Some of them got the genuine wire service photograph that ran with it, and 47% said a machine made it.
4:36
SPEAKER_03: Then the researchers did the thing that journalism has always done to vouch for a picture.
4:40
SPEAKER_03: They printed the photographer's name underneath it.
4:43
SPEAKER_03: Suspicion dropped by 4%.
4:45
SPEAKER_03: A byline is the oldest provenance system that we have.
4:48
SPEAKER_03: Attached to a real photograph in front of a modern audience, it's worth 4%.
4:53
SPEAKER_03: That isn't a fake fooling anybody.
4:55
SPEAKER_03: That's a real photograph that is losing.
4:57
SPEAKER_03: In 2020, researchers showed 2,000 people the Jordan Peele deepfake of Barack Obama, the one where a synthetic Obama insults Donald Trump.
5:14
SPEAKER_03: Then they asked, did Obama ever actually say that?
5:17
SPEAKER_03: Half got it right, 16% were fooled and said yes, and a third answered that I don't know.
5:23
SPEAKER_03: That third group is the finding.
5:25
SPEAKER_03: They weren't fooled, they were disarmed.
5:27
SPEAKER_03: And they're the ones who came away trusting the news and their feeds less than when they sat down.
5:32
SPEAKER_03: The fake didn't have to work, it only had to exist.
5:36
SPEAKER_03: So you warn people, that's the obvious fix.
5:39
SPEAKER_03: Here's how you test a thing like that.
5:41
SPEAKER_03: You need one video in two versions, one honest and one fake.
5:45
SPEAKER_03: So they hired an actor.
5:46
SPEAKER_03: They filmed him giving a short speech as a politician who doesn't actually exist, and then they made a second copy of that same clip with his face swapped by deep fake software.
5:56
SPEAKER_03: Same man, same speech.
5:58
SPEAKER_03: One version is footage, the other is forgery.
6:00
SPEAKER_03: Then 1,396 Americans each got one of the two, and nobody was told which one they had.
6:06
SPEAKER_03: Then a random half of everyone read this line of text above the player before it played.
6:12
SPEAKER_03: Warning computer scientists are increasingly concerned about deep fake videos.
6:17
SPEAKER_03: With deep fakes, it is now possible to use a computer to convincingly manipulate videos of people to make them appear to say things that they never said.
6:27
SPEAKER_03: The warning didn't help anybody tell the two videos apart.
6:31
SPEAKER_03: What it did was move people toward calling the genuine one fake.
6:35
SPEAKER_03: Then they ran it on a real politician.
6:37
SPEAKER_03: 1,925 people watched Mitt Romney in 2002 promising to preserve and protect a woman's right to choose.
6:45
SPEAKER_01: And I've been very clear on that.
7:14
SPEAKER_00: Hand that judgment to a machine.
7:22
SPEAKER_00: So that's what we did.
7:23
SPEAKER_03: This is a free detector called Zero GPT.
7:26
SPEAKER_03: It's running on my screen.
7:27
SPEAKER_03: And that's the Declaration of Independence going into it.
7:31
SPEAKER_03: In October 2024, a data analyst named Christopher Penn ran the Declaration of Independence, and the results he got back was that it was 97.7% machine written.
7:42
SPEAKER_03: In fairness to the industry, a competitor called GPT-0 looked at the same passage and called it 89% human.
7:49
SPEAKER_03: It got the declaration right.
7:50
SPEAKER_03: Then feeded a section of the Constitution, GPT-0 reports that that text was likely written entirely by AI.
7:57
SPEAKER_03: None of these tools can see artificial intelligence.
8:00
SPEAKER_03: Nothing can.
8:01
SPEAKER_03: What they can measure is how predictable your next word is and how much your sentence lengths bounce around.
8:07
SPEAKER_03: Write clean, even prose, and the machine calls it a machine, which is the problem, because clean is exactly what every English teacher in the country spends 12 years beating into you.
8:17
SPEAKER_03: So the better your English gets, the guiltier you look.
8:20
SPEAKER_03: The company selling the detector says, don't discipline students with the detector.
8:24
SPEAKER_03: Schools try to, anyways, because the alternative was admitting that there's no way to tell.
8:29
SPEAKER_03: OpenAI built one of these two in 2023.
8:32
SPEAKER_03: It caught 26% of machine-written text, and it flagged human writing as machine-written 9% of the time.
8:39
SPEAKER_03: They stopped offering that feature after about six months.
8:42
SPEAKER_03: Then Stanford ran a batch of these detectors on two piles of essays, one pile by American eighth graders and one by foreign students taking the English exam that you need to pass to study here.
8:53
SPEAKER_03: The eighth graders came through almost clean.
8:55
SPEAKER_03: The foreign students got called machine written 61% of the time, and nearly all of their essays were flagged by at least one detector.
9:03
SPEAKER_03: May 2023, Texas AM University Commerce, an instructor named Jared Mum, emails his class to say that he has loaded their final papers into ChatGPT, and he asked Chat if they were AI.
9:15
SPEAKER_03: Grades stopped.
9:16
SPEAKER_03: Several of the students get an X on the transcript instead of a mark.
9:20
SPEAKER_03: Somebody on Reddit took a page out of Mum's own doctoral dissertation and pasted it into ChatGPT and asked the same question.
9:27
SPEAKER_03: The university says nobody failed the class in the end, and one student admitted to using ChatGPT.
9:33
SPEAKER_03: Two years later, at the University of Buffalo, Turnitan flagged three of students' Kelsey Amond's assignments as AI, the last one at 97% chance that she used AI.
9:44
SPEAKER_03: She was weeks from graduating.
9:46
SPEAKER_03: She cleared her name by sending in her entire browser history.
9:49
SPEAKER_03: That's what clearing your name looks like now.
9:51
SPEAKER_03: You hand over your entire browsing history and you hope that everyone believes you.
9:55
SPEAKER_03: Turnatin, the company whose software flagged her, says its false positive rate is about 4% at the sentence level.
10:02
SPEAKER_03: Bloomberg ran its own test in 2024.
10:04
SPEAKER_03: They took 500 college application essays written in the summer of 2022, before ChatGPT was public, so every word of them was human.
10:12
SPEAKER_03: Then they fed them to GPT-0 and copy leaks.
10:16
SPEAKER_03: Some still came back as flagged with near total confidence.
10:19
SPEAKER_03: Then a reporter took an essay that both tools had called 100% human, it ran it through grammarly to make it sound more academic, and watched it come back at 100% AI written.
10:29
SPEAKER_03: In one year, Australian Catholic University opened nearly 6,000 academic misconduct cases, and about 90% of them involved AI.
10:37
SPEAKER_03: One of them was a nursing student named Madeline.
10:40
SPEAKER_03: She believes it cost her a job.
10:42
SPEAKER_03: New nurses there get hired into a first-year hospital program, and you can't apply for one without a finished transcript.
10:48
SPEAKER_03: Hers said results withheld until the hiring was over.
10:52
SPEAKER_03: So how do you prove you wrote it yourself?
10:54
SPEAKER_03: For a student, getting that wrong costs a semester.
10:57
SPEAKER_03: In July, it cost the man$2.4 million.
11:03
SPEAKER_03: A 25-year-old Nigerian PhD student at Southern Methodist University named Jerry Falati, he sold his first book, a crime novel entitled Call Me, I'll Hide the Body.
11:14
SPEAKER_03: 14 publishers bid for it.
11:16
SPEAKER_03: Minotaur won with a deal that deadline put at$2.4 million for that book.
11:21
SPEAKER_03: UTA started setting up the film auction.
11:23
SPEAKER_03: Then an industry newsletter ran about a debut author in AI.
11:27
SPEAKER_03: There was no test and no proof.
11:29
SPEAKER_03: His own agents pulled the book.
11:31
SPEAKER_03: Publishers Lunch is reported to have put the manuscript through a detector and got a 97% machine score.
11:37
SPEAKER_03: Filati told the Wall Street Journal he had used an open source model on his own computer for research and never for the writing.
11:44
SPEAKER_03: And then he ran his own test.
11:46
SPEAKER_03: 97% for his manuscript, 98% and 99% for novels published before ChatGPT existed.
11:53
SPEAKER_03: His agents never said he cheated.
11:55
SPEAKER_03: They said that they can no longer substantiate that he had not.
11:58
SPEAKER_03: The book is gone either way, and there is no test on the market that can close the question in either direction.
12:05
SPEAKER_03: March 2018.
12:18
SPEAKER_03: And in 2023, they asked the court to put that clip in front of Musk under oath.
12:24
SPEAKER_03: Judge Evit Pennypacker wrote back, their position is that because Mr.
12:27
SPEAKER_03: Musk is famous and might be more of a target for deepfakes, his public statements are immune.
12:33
SPEAKER_03: In other words, Mr.
12:34
SPEAKER_03: Musk and others in his position can simply say whatever they like in public domain, then hide behind the potential for their recorded statements being a deep fake to avoid taking ownership of what they did actually say and do.
12:46
SPEAKER_03: The court is unwilling to set such a precedent.
12:49
SPEAKER_03: Tesla settled the case in April 2024, days before the trial was due to start.
12:54
SPEAKER_03: That was a civil case about a deposition.
12:56
SPEAKER_03: And a criminal trial, the move is cruder.
12:59
SPEAKER_03: The evidence against you is a video, so you make the jury wonder whether the video is real.
13:03
SPEAKER_03: Lawyers started trying it almost immediately and failed.
13:07
SPEAKER_03: March 2022, the first January 6th trial, Guy Reffitt went to the Capitol with a gun on his hip and a camera on his helmet, and the government plays his own footage to the jury.
13:18
SPEAKER_03: His lawyer turns to the FBI agent on the stand and asks whether she's heard of deepfakes.
13:23
SPEAKER_03: No expert, no evidence, just the word.
13:26
SPEAKER_03: Reffitt is convicted on all five counts.
13:28
SPEAKER_03: Five months later, another defendant's lawyer won't even agree that a YouTube video from that day is genuine.
13:34
SPEAKER_03: He's convicted too.
13:36
SPEAKER_03: In January of this year, an Iowa appeals court heard from a man who was convicted of sexual abuse.
13:50
SPEAKER_03: The court wrote that he had pointed to no evidence in the record supporting this speculative theory, and that nothing about the video gave even the slightest hint that it was an artificial creation.
14:01
SPEAKER_03: Now the experiment that everybody assumes has been run.
14:04
SPEAKER_03: Put real video in front of a jury, have the defense say the word deepfake with no expert and no evidence, and watch the conviction rate fall.
14:12
SPEAKER_03: I went looking for that study.
14:14
SPEAKER_03: It does not exist.
14:15
SPEAKER_03: Nobody has run it.
14:16
SPEAKER_03: In a 2025 review of everything published on what deepfakes do to people's judgment, the number of studies on jurors in guilt was zero.
14:25
SPEAKER_03: The closest anyone has come is in Wales.
14:27
SPEAKER_03: A research project at Swansea University called True staged an entire war crimes trial on film to find out what jurors actually do with video pulled off the internet.
14:37
SPEAKER_03: They put 132 ordinary people through it as jurors, and 11 separate juries and let them deliberate.
14:43
SPEAKER_03: The evidence against the defendant was a genuine video of an airstrike.
14:47
SPEAKER_03: It was verified by Bellingcat.
14:49
SPEAKER_01: May I ask, please, that the video is played, and I, for the time being, am can say nothing.
14:52
SPEAKER_01: I'm just going to ask you to look at it.
15:04
SPEAKER_03: Nine juries came back not guilty.
15:06
SPEAKER_03: Two hung.
15:07
SPEAKER_03: Not one conviction.
15:09
SPEAKER_03: They believed the expert.
15:10
SPEAKER_03: What they couldn't get past was the video itself.
15:13
SPEAKER_03: Who shot it, whether it had been cut, how it ended up online.
15:17
SPEAKER_03: Nobody in that courtroom had to say the word deepfake.
15:20
SPEAKER_03: February 17th of this year, New York's highest court.
15:23
SPEAKER_03: The evidence was a video the FBI pulled off of a hacker's hard drive, and it showed a 14-year-old girl being abused by her mother's boyfriend.
15:31
SPEAKER_03: A family court judge relied on that video in a case about whether the girl was safe in that home.
15:37
SPEAKER_03: The Court of Appeals threw that finding out, four votes to three, not because the video was fake.
15:43
SPEAKER_03: No court has ever found that it was.
15:45
SPEAKER_03: They threw it out over how it had been authenticated.
15:48
SPEAKER_03: And this is the sentence that did it.
15:50
SPEAKER_03: The increasing prevalence of deep fake videos has only rendered the method of matching circumstantial details in a video, personal observations, a more suspect form of authentication.
16:01
SPEAKER_03: The method is suspect now because deep fakes exist.
16:04
SPEAKER_00: The video didn't change, the standard did.
16:52
SPEAKER_03: Three months after New York, they decided it isn't warranted yet.
16:56
SPEAKER_03: The Federal Judicial Center had just asked 931 federal judges whether this had ever happened in their courtroom.
17:03
SPEAKER_03: Fifteen of them said yes.
17:05
SPEAKER_03: So the rule stays unwritten because in federal court the problem is still rare.
17:10
SPEAKER_03: September 2025, Alameda County, California.
17:13
SPEAKER_03: Judge Victoria Kalikowski is handed video testimony from a witness and she watches it carefully.
17:19
SPEAKER_03: What settled it was the metadata.
17:21
SPEAKER_03: The file said it was captured on an iPhone 6 Plus, a phone that came out years before Apple put any AI software into it.
17:28
SPEAKER_03: She threw the whole case out with prejudice.
17:30
SPEAKER_03: So the courts are now catching fakes presented as real and catching real things being called fake.
17:36
SPEAKER_03: And that rule the committee says isn't warranted.
17:38
SPEAKER_03: Somebody already wrote it.
17:40
SPEAKER_03: It is drafted.
17:40
SPEAKER_03: It's been argued over and it's on hold.
17:42
SPEAKER_03: The committee's own note on that draft says a broad claim of deep fake isn't enough to put a court and the other side to the time and expense of proving a piece of evidence wasn't manipulated.
17:53
SPEAKER_03: And what the note calls the problem?
17:55
SPEAKER_03: Not AI evidence or authentication, the liar's dividend.
17:59
SPEAKER_03: Sitting in a federal rules document.
18:01
SPEAKER_03: Ask the judges who haven't seen one yet what they want if it walked into their courtroom, and more than eight in ten say the same thing.
18:09
SPEAKER_03: If you're going to cry deep fake, then you also have the burden of proof to show them something first before the court spends a day on it.
18:16
SPEAKER_03: So the judges are asking for the rule, and the committee's answer is that there aren't enough cases yet to justify writing it.
18:23
SPEAKER_03: Which brings us to the liar's dividend.
18:26
SPEAKER_03: Can politicians claim misinformation to evade accountability?
18:34
SPEAKER_03: Todd Aikens said that on St.
18:36
SPEAKER_03: Louis television in August 2012, running for the Senate in Missouri.
18:40
SPEAKER_03: Now, here are three things he can say back to you to try to keep your vote.
18:44
SPEAKER_03: One, nothing at all.
18:46
SPEAKER_03: Two, yes, I did say that, and I apologize for making those comments.
18:50
SPEAKER_03: Three, you can't know what's true these days with so much misinformation out there.
18:58
SPEAKER_03: Calen Jackson Schiff, Daniel Schiff, and Natalia Bueno ran this on more than 15,000 Americans across five experiments.
19:05
SPEAKER_03: Using real politicians in real scandals, the only invented part was what the politician said in response.
19:12
SPEAKER_03: It was published in the American Political Science Review, the top journal in the field.
19:16
SPEAKER_03: The third answer, one.
19:18
SPEAKER_03: Claiming misinformation beat saying nothing, and it beat apologizing.
19:22
SPEAKER_03: Opposition to the politician dropped by about a quarter, and it didn't only work on his own voters.
19:28
SPEAKER_03: The politicians were two Republicans and two Democrats.
19:31
SPEAKER_03: It didn't change whether people believed the scandal happened.
19:33
SPEAKER_03: They still thought they did it.
19:35
SPEAKER_03: They backed him anyway.
19:36
SPEAKER_03: Then the good news, and it's the only good news in the whole study.
19:40
SPEAKER_03: When the researchers put a fact check right after the denial, the benefit disappeared.
19:45
SPEAKER_03: Not reduced, it was gone.
19:46
SPEAKER_03: Everything you just heard was people reading about a scandal.
19:49
SPEAKER_03: Then the researchers ran it again with the video.
19:52
SPEAKER_03: This year, researchers in Germany ran the same test on video and nothing else.
19:57
SPEAKER_03: People watched real footage of an American state lawmaker slurring his words in session.
20:10
SPEAKER_03: Half of them then read a statement from him calling the video a deep fake.
20:15
SPEAKER_03: Half read an apology.
20:16
SPEAKER_03: The deep fake claim beat the apology.
20:18
SPEAKER_03: He was rated the better leader for denying it, and the people who read that denial came away more convinced the real footage had been manipulated.
20:26
SPEAKER_03: And there's a second half to this, and it's worse than the vote.
20:30
SPEAKER_03: The week before Ireland's abortion referendum in 2018, researchers put six news stories in front of 3,000 voters.
20:37
SPEAKER_03: Four were real, two were invented, complete with a genuine undoctored press photograph.
20:42
SPEAKER_03: Forty-eight percent of them remembered at least one thing that never happened.
20:46
SPEAKER_03: More than a third could describe the event that they were remembering.
20:50
SPEAKER_03: And the false memories lined up with how they were going to vote.
20:53
SPEAKER_03: Then the researchers warned them that some of what you just read was fabricated.
20:58
SPEAKER_03: Go back and tell us which.
20:59
SPEAKER_03: Among the people who had already formed a false memory of the fake poster story, 69% kept the memory anyway.
21:06
SPEAKER_03: Now the part I did not expect, the same lab went looking for whether deep fake videos make this worse twice.
21:12
SPEAKER_03: It doesn't.
21:12
SPEAKER_03: They ran fake news as plain text, as text with a photograph, and as a deep fake video.
21:18
SPEAKER_03: The video was no better than the sentence.
21:20
SPEAKER_03: In a separate study with invented movie remakes, 49% of people remembered films that were never made.
21:27
SPEAKER_03: And the deep fake clips performed the same as written description.
21:30
SPEAKER_03: The machine isn't what makes you remember a thing that never happened.
21:33
SPEAKER_03: You do that on your own and for free.
21:44
SPEAKER_03: That's the day that CNN reported that Mark Robinson had posted for years on a polyurphy site's message board.
21:50
SPEAKER_03: He's the Republican nominee for the governor of North Carolina, seven weeks out from election day.
21:55
SPEAKER_03: Two days later, CNN put Kalen Jackson's shift on air to talk about Robinson.
22:08
SPEAKER_03: She is one of the three researchers behind that 15,000-person study, and she was there to answer one question.
22:15
SPEAKER_03: When a politician calls a true story fake, do the voters stay with him?
22:19
SPEAKER_03: He lost that November by more than 14 points.
22:22
SPEAKER_03: He sued CNN for$50 million, and then he dropped the suit.
22:26
SPEAKER_03: In March of this year, on a podcast called After the Call, he said he'd lied about the whole thing.
22:32
SPEAKER_03: Then they asked him whether he'd do it differently.
22:35
SPEAKER_03: That's one man, one state, one election.
22:38
SPEAKER_03: Here's the same two words where the evidence is a street full of bodies.
22:42
SPEAKER_03: April 2022, northwest of Kiev, the Russians pull out.
22:46
SPEAKER_03: The bodies are in the street, and the footage goes around the world in a day.
22:50
SPEAKER_03: Russian's Ministry of Defense says the whole thing is a provocation.
22:53
SPEAKER_03: Not one local resident suffered.
22:55
SPEAKER_03: The bodies in the video move.
22:57
SPEAKER_03: Look at the 12th second, they said.
22:59
SPEAKER_03: Look at the corpse on the right.
23:00
SPEAKER_03: They never used the words artificial intelligence.
23:03
SPEAKER_03: In 2022, they didn't have to.
23:05
SPEAKER_03: It took about a day to answer them.
23:07
SPEAKER_03: The New York Times pulled satellite images from the 11th and the 19th of March, and they found 11 bodies lying on Yablonska Street while the town was still under Russian control.
23:17
SPEAKER_03: A group called the Center of Information Resilience matched the shadows in the footage to the position of the sun.
23:23
SPEAKER_03: 161 degrees, 1108 in the morning, the 19th of March.
23:27
SPEAKER_03: That is the price of proving that a street full of dead people is real.
23:31
SPEAKER_03: Satellite passes, solar angles, volunteers working through the night inside of 48 hours.
23:37
SPEAKER_03: That was 2022 against a claim of staging.
23:40
SPEAKER_03: Now put the same street in front of the same people with a generation of AI video behind them and count the days.
23:58
SPEAKER_00: So add it up.
24:01
SPEAKER_03: For about 150 years, a photograph worked as a kind of testimony.
24:05
SPEAKER_03: Wasn't perfect, and people faked them from the beginning.
24:08
SPEAKER_03: But a picture was a witness that you could hold up in a courtroom or print on a front page, and the powerful had to answer it.
24:15
SPEAKER_03: A recording was a witness who couldn't be bought.
24:18
SPEAKER_03: That's the thing that's being repriced.
24:20
SPEAKER_03: Not whether pictures can be faked, which they always could, but who carries the burden when somebody says one was?
24:26
SPEAKER_03: The strongest move available to a guilty man is no longer explaining the evidence.
24:31
SPEAKER_03: It's asking whether the evidence is even real and then letting the rest of us do the work of proving that it is.
24:37
SPEAKER_03: And it runs the other way too at the same time.
24:39
SPEAKER_03: In early 2024, Pikesville High School in Baltimore County, audio goes around of the principal, Eric Iswert, making racist and anti-Semitic remarks.
24:49
SPEAKER_02: Between these ungrateful black kids who can't test their way out of a paper bag.
24:52
SPEAKER_02: And don't let me get started on DJ, I'm gonna drag his black ass out of here one way or another.
24:56
SPEAKER_02: And if I have to get one more complaint from one more Jew in this community, I'm gonna join the other side.
25:02
SPEAKER_03: He's removed from the school.
25:03
SPEAKER_03: He gets threats.
25:04
SPEAKER_03: The recording was made by the school's athletic director, Dazon Darien, with AI, and Darion took a plea and was sentenced to four months.
25:12
SPEAKER_03: Eric Iswert lost his school to a fake that he couldn't disprove fast enough.
25:17
SPEAKER_03: The girl in New York State, she lost the fight showing that her stepfather was abusing her because a court couldn't be certain that her video was real, the same technology in opposite directions, and neither of them had any way to fight it.
25:29
SPEAKER_03: There's one more line in the American study at the very end, and it's the one that should worry any honest person.
25:35
SPEAKER_03: The liar gets the dividend.
25:37
SPEAKER_03: The man telling the truth about a fake gets nothing because he's making the same noise the liar makes.
25:43
SPEAKER_03: Eric Eisworth said the recording was fake, and it was, and nobody could take his word for it.
25:48
SPEAKER_03: He got his name back months later when the detectives traced the audio to his own athletic director, and it's not evenly distributed.
25:56
SPEAKER_03: Proving a video is real costs money, or a forensics lab, or a friend at Berkeley who will look at the shadows for a news network.
26:03
SPEAKER_03: Saying it's fake costs one sentence and no evidence at all.
26:06
SPEAKER_03: So line up who is best placed to use that.
26:09
SPEAKER_03: It's not the student and it's not the 14-year-old girl in New York.
26:12
SPEAKER_03: It's the people who can take the fullest advantage of a world where the camera stopped counting as a witness.
26:19
SPEAKER_03: This channel compares every one of these systems to the shogath.
26:23
SPEAKER_03: Lovecraft's thing that was built to serve and it learned to wear whatever face kept the room calm.
26:28
SPEAKER_03: In the Lovecraft story, the shoggaths were built as tools, and the thing that frightened the people who made them, it wasn't the strength.
26:35
SPEAKER_03: It was that they learned to imitate the voices of their masters.
26:39
SPEAKER_03: So look at where that ended up.
26:40
SPEAKER_03: In the laboratory at Lancaster and Berkeley, the manufactured faces scored higher for trustworthiness than the human ones.
26:48
SPEAKER_03: The mask tested better than the face, and nobody priced the second half of it.
26:53
SPEAKER_03: Once everybody knows a face can be a mask, the man caught on camera gets to say he was wearing one.
26:59
SPEAKER_03: That's how a bag coming out of a window becomes AI, and how a missile strike in the Caribbean becomes AI on the same afternoon.
27:07
SPEAKER_03: Which leaves the question that nobody has answered yet.
27:10
SPEAKER_03: What actually works?
27:12
SPEAKER_04: What happens when we enter a world where we can't believe anything?
27:15
SPEAKER_04: Anything can be fake.
27:16
SPEAKER_04: The news story, the image, the audio, the video.
27:19
SPEAKER_04: Well, in that world, nothing has to be real.
27:21
SPEAKER_04: Everybody has plausible deniability.
27:25
SPEAKER_03: There's no detector coming.
27:26
SPEAKER_00: The detectors are the problem with a subscription fee.
27:35
SPEAKER_03: What the people who study this actually recommend is the opposite direction.
27:39
SPEAKER_03: It works like a notary stamp, except the camera applies it the instant that the shutter fires.
27:44
SPEAKER_03: Here's Robert Chesney on how it would work.
27:53
SPEAKER_03: Leica shipped the first camera that signs its own photographs in October 2023.
27:58
SPEAKER_03: Last August, Google put that same signature on every picture the Pixel 10 takes.
28:02
SPEAKER_03: Sony put the same signature on video for newsrooms last October and tested it with BBC's research department.
28:09
SPEAKER_03: The point of all of it is that by the time a picture reaches you, it arrives carrying its own history the way that a house deed carries a chain of title.
28:17
SPEAKER_03: Google put a warning inside their own announcement about their own product.
28:21
SPEAKER_03: Sign the real thing.
28:22
SPEAKER_03: Don't just flag the fake one.
28:24
SPEAKER_03: Do it the other way around, and you've dug the same hole from the other end.
28:27
SPEAKER_03: And it breaks too.
28:29
SPEAKER_03: Last September, Nikon suspended the feature on the Z63 because the camera could be made to stamp its seal of authenticity onto a picture it hadn't actually verified.
28:38
SPEAKER_03: And what about older technology?
28:40
SPEAKER_03: The Witness.
28:41
SPEAKER_03: August 2025, Will Smith posted a montage of crowds from his summer tour.
28:45
SPEAKER_03: Inside two weeks, the internet has decided that the fans are machine made, that he faked his own audience.
28:51
SPEAKER_03: The truth turned out worse than either side had expected.
28:54
SPEAKER_03: His team had run real crowd photographs through an AI tool that animates still images.
29:00
SPEAKER_03: And YouTube was running an experimental filter over shorts that smeared the faces further.
29:05
SPEAKER_03: YouTube is running machine learning over people's uploaded shorts, sharpening faces and skin without telling the creators that it was doing it.
29:12
SPEAKER_03: Musicians noticed their own faces looking wrong in their own videos.
29:16
SPEAKER_03: The company confirmed it.
29:17
SPEAKER_03: So a real crowd shot by a real photographer went through a model at the creator's end and a second model at the platform's end, and it came out looking like a lie.
29:26
SPEAKER_03: Nobody in that chain set out to fake anything.
29:28
SPEAKER_03: Two of the faces that the internet wrote off as machine made belonged to Patrick and Geraldine from Bern.
29:34
SPEAKER_03: They were in that crowd at a festival in Switzerland, holding up a handmade sign telling Will Smith that one of his songs got one of them through cancer, which points to the part that you can actually do this week.
29:45
SPEAKER_03: When somebody tells you that a picture or a recording is AI, ask them for the file, not the opinion.
29:52
SPEAKER_03: Ask who else was in the room, because in every real story in this episode, there was somebody in the room.
29:58
SPEAKER_03: And stop outsourcing the question to a free website with a confidence score because you've now watched those websites convict the Declaration of Independence, the Constitution, a professor's own dissertation, and 98% of a novel written in 2012.
30:13
SPEAKER_03: None of these things should be hard to prove or disprove.
30:17
SPEAKER_03: Somebody just has to go and look instead of deciding blindly.
30:21
SPEAKER_03: A file can be signed and a signature can be faked.
30:24
SPEAKER_03: A person who was in the room is a different kind of a problem for a liar.
30:28
SPEAKER_03: Eight faces and nobody could pick the machine.
30:30
SPEAKER_03: A detector that read the Declaration of Independence and called it a robot.
30:34
SPEAKER_03: A nursing student handing over her browser history to prove that she'd done her own homework.
30:39
SPEAKER_03: A novelist who lost two and a half million dollars to a rumor that no one had ever had to prove.
30:45
SPEAKER_03: A court in New York that looked at a real video of a real crime and couldn't be certain enough to act.
30:51
SPEAKER_03: That's the Shogoth's alibi.
30:53
SPEAKER_03: The alibi of a thing built to serve that learned to imitate the people who built it and wore their faces back at them.
31:00
SPEAKER_03: But the mask was never the prize.
31:03
SPEAKER_03: The prize is that once the world knows masks exist, every real face gets to claim it's wearing one, and every real thing has to prove it isn't.
31:11
SPEAKER_03: So the next time that somebody with everything to lose looks at a camera and tells you it isn't real, ask who else was in the room.
31:18
SPEAKER_03: Because after all the technology that we've built, the oldest verification we have is the only one that the machine can't counterfeit a witness, somebody who is there and who will stand up and say so.