Episode Transcript
[00:00:00] Sa.
[00:00:31] Hello, I'm Dr. Alan Badot and this is AI Today, three weeks from tonight, Americans finish voting in the midterm elections in many states. Ballots are already being cast. And somewhere between now and then, you're gonna see a face on your screen saying something that a face maybe never said.
[00:00:53] You may already have seen it. You may have laughed at it, you may have shared it.
[00:00:58] That's the point.
[00:01:00] Now let me show you the season you're living in right now.
[00:01:07] Not the version on the cable panels, okay? The one in dad Buys. And I'm going to show you both sides of the aisle because both sides are doing it in Texas, considering I'm in Austin right now. Republican Lt. Gov. Dan Patrick's campaign ran an AI ad depicting his Democratic opponent, opponent State Representative Vicki Goodwin, in a shirt that reads taxes instead of Texas.
[00:01:38] Goodwin filed a criminal complaint under the state's deepfake law. The campaign took the ads down, called them obvious parody and says none ran inside the law's 30 day window.
[00:01:52] In New Jersey, the race for governor last year, Democratic Congressman Josh Gotten Heimer opened his TV campaign with AI generated scenes of himself in a boxing ring trading punches with Donald Trump.
[00:02:09] To his credit now he did label the ad as AI generated.
[00:02:15] In Wisconsin, Republican Congressman Derek Van Orden posted AI generated videos of his Democratic opponent, Rebecca Cook, saying things that she never said. Her lawyer sent a cease and desist letter. His videos did carry an automatic AI label. He answered the letter with an AI generated duck saying cease and desist.
[00:02:41] Now they're getting very creative, right? In San Francisco, Democratic State Senator Scott Weiner's campaign launched an AI chatbot that answered thousands of questions about his opponent, Connie Chan.
[00:02:56] It drew sharp criticism, though he pulled it, saying it missed the mark.
[00:03:03] We've heard that before, haven't we?
[00:03:05] Usually when things don't go well and in upstate New York, you can watch the whole cycle play out. In one district last summer, Democratic Democrat Blake Jen Den Bine posted an AI video of Republican Congresswoman Elise Stefanik dancing on the House floor.
[00:03:30] Her team called it sexist and deceitful. This August, Republican Anthony Constantino, running for the same seat, answered with a far harsher deepfake of Gendenbein and said he was responding to Gendenheim's video.
[00:03:48] So we've reached a point. The, you know, in American politics where the answer to a deep fake seems to be another deep fake.
[00:03:59] Yes, no, the numbers.
[00:04:03] Because anecdotes are how you get fooled, okay? And data is how you stop.
[00:04:11] The Wesleyan Media Project It's a nonpartisan research group, and they've counted at least 164 political ads this cycle that were generated over or enhanced with AI. Just under $80 million in spending.
[00:04:26] In May, the same researchers had counted 33 ads and about $20 million.
[00:04:35] Now, that curve is not subtle, okay?
[00:04:39] They call their own figure an undercount. And here's why. Because 69% of those ads or, you know, 78% of the money carried no disclaimer telling you AI was involved at all. All that's a problem.
[00:04:55] Now you deserve a partisan split, right? Because we're going to hit both sides. And here it is. Okay?
[00:05:03] In the paid ads, Wesleyan could identify Republican candidates and allied groups sponsored 80% of the AI ads and 83% of the spending.
[00:05:15] Now, that means that Democrat sponsors 11% and 14%.
[00:05:22] But understand what that count actually measures. Okay? It covers paid advertising. The researchers could find broadcasts, streaming digital ads. It does not count the AI content. Leaders of both parties post for free on their own accounts every single day.
[00:05:42] And that is is where some of the most aggressive material in American politics lives.
[00:05:49] So last year, the President of the United States posted AI videos of his opponents. Okay?
[00:05:56] House Democratic leader Hakeem Jeffries and a sombrero. Former President Obama in handcuffs. The White House says the memes will continue.
[00:06:05] Meanwhile, the Democrat governor of California turned AI mockery into a press office strategy. The President has Marie Antoinette. And last December, the President and Secretary of War Pete Hegseth and Stephen Miller in handcuffs.
[00:06:25] Yeah, now I'm not here to adjudicate taste, okay? That's not what we do on this show. I'm here to tell you what it means. Structurally normalization isn't coming from the fringe. It's coming from the podium. And people need to see that.
[00:06:40] And if you rebalance that, it's almost 5050 on the attack ads because some are free, some are paid.
[00:06:49] Both podiums.
[00:06:51] Now here is what actually surprised me when I was looking at this. So for years, people in my field warned all the time about the perfect fake, the flawless forgery, the video so no real human could tell.
[00:07:09] We spent fortunes quite honestly preparing for it. I've talked about it in other segments. I've talked about it in other shows. Almost two years ago, we talked about came the Wesleyan researchers found hyper realistic deep fakes in the cycles ads.
[00:07:27] But that isn't all that arrived.
[00:07:31] What also arrived is a fake Dr. Fauci running through the state fair with a giant syringe. Liz Cheney, Mitt Romney, Mike Pence carrying Pitchforks on the White House lawn.
[00:07:44] Texas Democrat James Telo in a blue dress styled like Maria from the Sound of Music.
[00:07:52] A Democratic candidate for governor in a boxing ring with the president.
[00:07:56] A congresswoman dancing on House floor.
[00:08:01] And those.
[00:08:03] Those are just the ones you can see and are easy to find. The quieter uses should worry you even more.
[00:08:12] So Democratic Congresswoman Jasmine Crockett was accused of using AI to enhance the crowd in the closing shot of one of her ads.
[00:08:22] Researchers found other ads with generated crowds and constituents who never existed.
[00:08:30] Nobody thinks to question a crowd, do they?
[00:08:36] What changed isn't only quality. It's also cost.
[00:08:40] A convincing fake once took a skilled team of AI engineers weeks worth of work.
[00:08:48] Today takes a laptop and maybe an afternoon, maybe lunch.
[00:08:54] When the cost of manufacturing a face falls to almost nothing, the number of faces manufactured goes toward infinity.
[00:09:07] And that is a big problem.
[00:09:12] So a researcher at nyu, they've got a name for it, right? They call it Slopaganda. And I love that name. Slopaganda.
[00:09:20] Cheap synthetic content built not to prove anything, but to make you feel something.
[00:09:27] The Deepfake era didn't arrive only as a forgery. It also arrived as a cartoon.
[00:09:34] And this is the part I want you to understand because it is the most counterintuitive idea in tonight's program. The cartoon may work better than the forgery.
[00:09:46] That's the reality.
[00:09:48] A forgery can be debunked. A fact checker can prove the real candidate was somewhere else that day. But nobody fact checks a feeling when you see your opponent in a dress, in handcuffs, dancing on the house floor. The claim isn't this happened. The claim is this is who they are.
[00:10:08] And there's no forensic test for that.
[00:10:11] The fake doesn't need to fool you. It only needs to stick with you.
[00:10:17] And even when you're trying, you can't tell. A study revealed last year by iProof, you know, which is a company that sells identity verification technology, tested 2,000 people in the United States, in the United Kingdom, on a mix of real and synthetic types of video.
[00:10:35] 1/10 of 1% identified every real item in every fake item correctly. That's right. 1/10 of 1%.01% spotted every fake.
[00:10:48] That is awful.
[00:10:50] Now consider the source. A company that sells detection has an interest in your doubt. But the finding matches what I have seen in my own work. Human eyes lost this race a long time ago. And the software built to catch fakes is trained on yesterday's fakes, while the forgers are already working on tomorrow's Fakes.
[00:11:13] We're already behind.
[00:11:15] So here's what I want to plant in the question that runs through the rest of tonight. So, since 2002, federal law has required candidates to appear in their broadcast ads and say a familiar phrase. You know it by heart. Okay, I approve this message. It's called the standby. Your ad.
[00:11:33] And the idea was simple. If you're going to say something about your opponent, you got to put your face on it, own it. Simple as that.
[00:11:42] That law asked the candidates to sign the message. Accountability. Right. Great.
[00:11:47] It never asked anyone to sign the face, though.
[00:11:51] Every face and every ad I just described was made by a machine. The machine did the synthesis beautifully, quickly and cheaply.
[00:12:01] That is what these systems can do. And they own the part of the work. Now, what none of these ads come, you know, or show is a person who stands behind the face. A person who says, I choose to put my opponent's face or likeness in a scene that never happened. And I'll answer for it.
[00:12:24] Never happens in almost every other serious profession. We solved this a long time ago. A doctor signs a chart. An engineer stamps the drawing. A lawyer signs the filing and answers every word in it. Political advertising is one of the last places in American life where a machine can author a human being. And nobody signs, nobody cares.
[00:12:47] Well, you need to care.
[00:12:50] That's the missing signature.
[00:12:52] And it's missing in a country that for the first time in history, can manufacture a candidate's face, their voice, and a crowd faster than you can count votes.
[00:13:06] Now, when we come back, I'll show you who is supposed to be enforcing these rules and why.
[00:13:13] Three weeks before our national election, the honest answer is almost no. Nobody.
[00:13:18] So stay with me. We'll be back after a few messages from our sponsors.
[00:13:48] Foreign.
[00:13:57] Welcome back. I'm Dr. Alan Badot.
[00:13:59] And in sports, the scariest thing that can happen isn't a dirty player.
[00:14:09] It's a game where the referees, they didn't show up right, because then everyone learns fast what the rules are worth.
[00:14:20] So let's walk the field.
[00:14:23] Let's count the referees.
[00:14:26] Let's start with Congress. Okay.
[00:14:29] Easy place to start.
[00:14:31] There is no federal law on deepfakes and political ads.
[00:14:35] If. If you're a lawyer, maybe you can find one. I was not able to.
[00:14:40] There's not a ban, there's not a label. And bills have been introduced, but none of them have reached the floor for a vote.
[00:14:51] Now, Congress can pass a deep fake law when it wants to. In 2025, it passed the Take It down act and that's against. AI generated intimate images with overwhelming support from both parties, as it should have.
[00:15:05] And it really helped form some of the foundations for Cerberus, our DigitalNet AI tool that we're using for helping folks overcome some of those challenges. So thank you for that.
[00:15:16] So the vote exists for deepfakes, just not these types of defects, right?
[00:15:23] One observation, very simple, about every member of every party, the people who would have to write a law against campaign deepfakes are the people who run the campaigns. Simple as that.
[00:15:40] Now let's move on to the next one. The Federal election Commission.
[00:15:47] In 2024, the FEC declined to write a new rule on AI.
[00:15:53] Instead, it said its existing ban on fraudulent misrepresentation applies no matter what technology is used.
[00:16:03] Except that the FEC needs four commissioners to act.
[00:16:07] And since the spring of 2025, after a series of departures, it hasn't had them.
[00:16:14] Two of the six seats are filled.
[00:16:17] They can't do anything even if they wanted to.
[00:16:20] Now, without a quorum, the nation's campaign finance regulator can't enforce its own rules, can't write new ones, or even hold hearings.
[00:16:33] By every indication, it will sit out this entire election.
[00:16:40] Then the FCC, the Federal Communications Commission.
[00:16:44] In 2024, the FCC did two things. It ruled that AI clone voices count as artificial under the federal robocall law. Thank Google, which made unsolicited AI voices call your cell phone illegal without your consent.
[00:17:02] That ruling, it's still in force. And it proposed a rule requiring broadcasters to disclose AI generated content and political ads.
[00:17:12] Now, that one stalled.
[00:17:15] Now watch this one very closely. This month, the FCC is weighing a petition from the Club for Growth, and that is a conservative advocacy for an expedited waiver.
[00:17:30] If it's granted, political groups could call your cell phone with artificial or pre recorded voices, including AI voices, without your consent. There would be a cap of three calls per caller every 30 days and an opt out reply. Comments are due October 19, which is about six days from now, and a decision could come before election Day.
[00:17:56] That won't make me happy because I was pushing previously about the robocalls in the first place. Turning them back on is useless. We don't want that. The public does not want that.
[00:18:10] Now let me give you both sides as far as I can, okay? The Club for Growth president puts it this way. We think the Constitution protects a conversation with a voter at least as highly as a delivery update.
[00:18:25] The group says that the change would help campaigns reach voters in more languages.
[00:18:30] That's a real argument. I'll give them that.
[00:18:34] The FCC's lone Democratic commissioner, Anna Gomez, warns of, you know, a tsunami of robocalls and misinformation. A lawyer known for defending companies and robocall lawsuits, hardly an anti business voice, calls it a terrible idea that is certain to be a abused.
[00:18:56] And consumer advocates point to the arithmetic, the math. The cap is three calls per caller.
[00:19:04] There are thousands of callers.
[00:19:08] The waiver also wouldn't belong to one party though every campaign, every committee left and right could use it.
[00:19:21] How many opt outs are we going to have to do?
[00:19:25] Terrible, terrible idea.
[00:19:28] Now recently on this program I asked who answers for agents that get out of control or out of their rooms. Remember, here's the same question at a national scale. An AI voice that can hold a conversation in your language about your issues, calling you three times a month from every committee that wants your vote.
[00:19:51] Somebody approves the script for one minute.
[00:19:54] Who approves what it says to you in the fourth minute is a conversation that probably nobody has written.
[00:20:04] Nobody, nobody proves.
[00:20:09] Minute four.
[00:20:13] And let me tell you, this is not, this isn't hypothetical or partisan. Okay, so in 2023, Democrat Shemaine Daniels, running for Congress in Pennsylvania put an AI caller named Ashley on their phones. It held conversations with thousands of voters in a single weekend.
[00:20:32] Now to its credit, Ashley introduced herself itself as an AI agent.
[00:20:39] Whether the next one does is up to whoever pays for it.
[00:20:44] And if you think enforcement will catch really the worst of it, let's consider the most famous case. Okay, so in January 2024 before the new Hampshire primary, thousands of voters received a robocall in a cloned Joe Biden voice telling them to skip the primary.
[00:21:04] The political consultant behind it, Steve Kramer, had worked for Democrat Dean Phillips primary campaign, which of course denied any involvement. Now Kramer admitted making the call.
[00:21:19] The FCC fined him $6 million.
[00:21:22] In June of last year, a New Hampshire jury acquitted him on all 22 criminal counts. He testified he made the calls to warn the country about AI. A federal court later ordered him to pay, you know, three voters. I think it was, you know, $7500 each.
[00:21:42] And he said publicly he won't pay it or he won't pay the FCC fine either.
[00:21:50] So the most famous election, deep fake in American history has so far cost its creator.
[00:21:57] Let me do the math. Oh yeah, 00.
[00:22:06] The lesson a consultant could draw.
[00:22:09] The downside's the headline and the headline is free advertising and the cost is nothing.
[00:22:15] Hmm, that's great.
[00:22:18] So what's left?
[00:22:19] Well, the states, last year the US Senate voted 99 to 1 to preserve the state's power to regulate AI. Then Congress left them to do the job alone.
[00:22:32] About 31 states have now passed AI laws. And we've talked about this, especially the ramifications of allowing states to have multiple AI laws that are different and different, touching and different reaching and impact to creators of AI. We've talked about those.
[00:22:51] And since December, a White House order has the Justice Department really challenging state AI laws that considers too burdensome.
[00:23:01] So even this patchwork, it lives under a cloud.
[00:23:06] And the people that you know are reaching for these laws, they are not only Democrats. Okay? In Tennessee, Republicans, you know, Senator Marsha Blackburn, now running for governor, demanded that a PAC pull an AI ad casting her as a paid pharmaceutical pitchman, citing the state's new AI disclosure law.
[00:23:26] And this is where the evidence, quite honestly, turns really uncomfortable.
[00:23:32] The Wesleyan researchers compared AI disclosure rates in states with these laws. In states without them. In states without a law, 32% of AI ads are disclosed.
[00:23:45] In states with a law, 29%.
[00:23:54] Wow.
[00:23:57] The laws sure don't appear to be changing any behavior, do they?
[00:24:03] The researchers are careful to call that comparison imperfect, of course, but it's hard to read it as a success.
[00:24:12] AI disclosure, 32% without a law, 29% with one.
[00:24:19] That's messed up.
[00:24:22] Now let's look at why. Okay, part of it is the patchwork. Colorado's law covers only deep fakes of candidates. And, you know, an ad that runs on a national streaming Service lands in 31 rule books and 19 states without one part of its timing.
[00:24:41] Texas makes it a crime to create and publish a deepfake video within 30 days of an election with intent to injure candidate or influence that result.
[00:24:52] But people don't all vote on election day anymore, right? Because there's write in ballots, there's early voting, there's all that stuff. Mail ballots go out weeks earlier. Which is exactly why Goodwin argues the clock should start 30 days before early voting begins.
[00:25:07] But there's no common sense in this, right? Why would we do that?
[00:25:12] Yeah.
[00:25:13] Now, part of it's also definition. Minnesota bans political deepfakes.
[00:25:21] X, of course, sued to overturn it. So far, the law remains in force, but the law reaches only content so realistic that a reasonable person would believe it.
[00:25:31] In May, a synthetic video of Democratic Senate candidate Peggy Flanagan aired, and nobody can say for certain whether it broke the law.
[00:25:43] That's right, 0.1%.
[00:25:46] Remember that statistic from the first segment?
[00:25:50] Think about what that standard rewards.
[00:25:56] The law punishes the convincing fake and Protects the ridiculous one.
[00:26:01] So campaigns have learned.
[00:26:05] Just be ridiculous enough.
[00:26:07] A duck, a dancing congresswoman, a pair of handcuffs.
[00:26:13] And then there's the wall that every one of these laws eventually hits. In 2024, California passed one of the most aggressive election deepfake laws in the country.
[00:26:27] Last year, a federal judge struck it down under the First Amendment, finding it's just too broad.
[00:26:34] And noting that counterspeech, the fact check in the community note, was the less restrictive tool.
[00:26:42] Here's the uncomfortable constitutional truth.
[00:26:46] So in 2012, in the United States versus Alvarez, the Supreme Court held that lies are not automatically outside the First Amendment.
[00:26:58] The government can't simply ban falsehood.
[00:27:02] A political speech gets the highest protection here.
[00:27:08] The Constitution protects your right to lie about yourself. The question in front of this country is whether it protects your right to manufacture your opponent.
[00:27:23] Wow.
[00:27:28] Free speech advocates and serious ones. Okay, say mostly yes, and they have a point.
[00:27:36] Satire is as old as the republic.
[00:27:39] Thomas Nast's cartoon of boss tweeds weren't forgeries.
[00:27:44] They were commentary. And they were helping bring down a political machine.
[00:27:50] If you don't know what I'm talking about, go read the history books, especially about New York politics.
[00:27:56] Give the government power to decide which images of a candidate are deceptive, and you've handed the people in power a weapon against the people trying to replace them. I understand that.
[00:28:09] It's the same thing with the flag debate, okay? It really is.
[00:28:14] People have the right to burn it.
[00:28:17] My right, my job is to defend their right to be able to do that until some sort of constitutional amendment is in there. Right? I don't have to agree with it, but that's just the realities of it.
[00:28:27] Election integrity advocates answer that. A cartoon announces itself. A deep fake impersonates.
[00:28:34] The first is an argument. The second is counterfeit.
[00:28:38] Both of them are right.
[00:28:39] And while they argue, the fields stay empty.
[00:28:45] Now, next, I'm going to take you outside of this country because the rest of the world is already living in a future we're arguing about.
[00:28:54] So stay with us. We'll be back after a few messages from our sponsors.
[00:29:31] Welcome back. I'm Dr. Alan Badeau, and this is AI. Today I promise to take you outside the country. So let's. Let's go on a little trip.
[00:29:41] Five stops.
[00:29:42] Each one is a preview of a future we haven't decided to prevent.
[00:29:48] Stop. 1. Slovakia, September 2020 3. Two days before a national election, an audio recording spreads online. It appears to capture the leader of a liberal opposition party discussing how to rig the vote.
[00:30:01] It was fake.
[00:30:03] Synthetic audio.
[00:30:04] But it landed inside Slovakia's pre election silence period. The 48 hours when campaigning and election coverage are restricted.
[00:30:15] The fake had a stage, though the rebuttal didn't.
[00:30:19] This party finished second.
[00:30:21] Nobody can prove the recording changed the outcome.
[00:30:25] Remember that sentence because it's gonna matter.
[00:30:29] Stop. 2.
[00:30:31] Romania, December 2024. The country's Constitutional Court annulled the first round of presidential election.
[00:30:39] Not over ballot counting, over what happened online.
[00:30:43] Declassified intelligence alleged a coordinated influence campaign had lifted a little known candidate through social media.
[00:30:52] The election was rerun following the spring sit. With that, a European democracy canceled its own election because of what had been done to its voters through their screens.
[00:31:08] Some called it the system protecting itself. Others called it the system overruling the voters.
[00:31:18] Both are partly true.
[00:31:21] Stop. 3.
[00:31:23] Ireland, October 2020. Five days before the presidential election, a video spreads on Facebook styled as a broadcast from rte now that's Ireland's national broadcaster. And it was complete with AI versions of real presenters. It reports that the front runner, who was Kathleen Connolly, had withdrawn and that the election was off. It was viewed nearly 90,000 times before social media platforms took it down.
[00:31:58] Now she won anyway, decisively. Okay, but notice the format.
[00:32:03] It didn't fake the cam, you know, candidate's opinion. It faked the news.
[00:32:08] It impersonated the institution whose job was to tell you what's real.
[00:32:13] That's the escalation to watch. First they faked the candidate, then they faked the referee. Ireland's version was caught.
[00:32:22] The next one is going to be better though.
[00:32:25] Now stop. War is kind of three places at once. Okay, just a little easier.
[00:32:32] In INDONESIA in early 2024, a political party, they released an AI generated video of Suharto, right, The country's former dictator, dead since 2008, urging people to vote for its candidates.
[00:32:47] In New York in 2023, the city's then mayor, Eric Adams, elected as a Democrat, sent robocalls in languages he doesn't speak in his own clone voice.
[00:32:57] And in Illinois this year, Democrat Jesse Jackson Jr. Used AI to restore the voice of an endorser, former Congressman Bobby Rush, whose own voice had been damaged by illness.
[00:33:10] A dead man endorsing the living. A living man speaking in languages he doesn't know.
[00:33:16] A man given back the voice he lost.
[00:33:20] One feels grotesque, one feels almost generous.
[00:33:25] One feels like a gift.
[00:33:27] Maybe.
[00:33:29] But here's the hard question.
[00:33:32] What's the difference?
[00:33:34] In each case, a voter heard a voice say words its owners never said.
[00:33:40] The line between accessibility and deception isn't in the technology it's in the intent.
[00:33:49] And intent is the one thing machine can't certify.
[00:33:56] Just like everything else we've talked about on this show.
[00:34:00] Now stop. Five happening right now.
[00:34:06] Yeah, Hope you're paying attention.
[00:34:08] Brazil is in the middle of its national elections this month with a presidential runoff on October 25.
[00:34:16] Under the same, or I guess I would say some of the most strictest AI campaign rules of any major democracy, deepfakes of real people are banned from campaign material. A candidate who uses one risks losing their registration or their seat from 72 hours before the vote until a day after posting or boosting new synthetic content that uses a candidate's image or voice is barred.
[00:34:52] And in a dispute, the burden can, of course, shift. The publisher may have to prove the content's general.
[00:35:02] Now, read that rule twice. In Brazil, when a video is challenged, the publisher may have to prove it's real.
[00:35:10] That's the future I'll describe in our final segment. So make sure you don't go away. But it's already written into law in Brazil. A deep fake can cost you your candidacy. In the United States, it might cost you two. Second disclaimer if you're bothered to include one at all.
[00:35:31] And seven and ten don't. Right.
[00:35:34] That's what we find out, law or no law.
[00:35:38] Terrible.
[00:35:39] Now, I'm not telling you that Brazil has solved this, okay?
[00:35:45] Experts say there are, you know, fines that are just too low, the forensics are too slow, and private messaging apps, that's a gray zone. So it may not count. It may count.
[00:35:58] I'm telling you, though, that one of the largest democracies decided that a candidate's face is not raw material.
[00:36:05] We haven't decided anything in the U.S.
[00:36:08] now, the idea that all these steps can put, you know, a stop to anything is kind of funny.
[00:36:20] But the idea that ties all five of these stops together, it's kind of the most dangerous idea in tonight's program.
[00:36:32] So in 2018, two law professors, Bobby Chesney and Danielle Citron, they gave it a name, and it's called a liar's dividend.
[00:36:45] That's right. A liar's dividend.
[00:36:49] So here's how it works. Once the public knows convincing fakes exist, anyone caught on a real recording can say it's fake.
[00:36:58] And the more people that learn about deep fakes, the more valuable that excuses become.
[00:37:06] Every public awareness campaign about deepfakes, including, and I'll admit this, you know, raises the value of the lie.
[00:37:16] And notice, who collects the liar's dividend has no party it pays out to whoever gets caught, which means every politician and every party has a quiet interest in keeping the public just confused enough to doubt it.
[00:37:32] The dividend also has a twin, though, right? And it's kind of just as ugly.
[00:37:38] The fake that convicts, right?
[00:37:40] In January, after an ICE agent shot and killed a woman in Minneapolis, people online, many of them furious at the agency, used AI tools to unmask the agent from a blurry video.
[00:37:54] The faces were invented.
[00:37:58] One was matched to a gun shop owner in Missouri who had never been to Minnesota and never worked for ICE.
[00:38:07] He was harassed.
[00:38:09] His Facebook account was taken down.
[00:38:12] PolitiFact raided the claim. Pants on fire, Right?
[00:38:18] Nobody approved that face either.
[00:38:21] Now, there's an old case that shows you exactly where the road ends.
[00:38:27] In Galvan, at the end of 2018, the President had been out of public view for months. After a stroke, a New Year's Eve video finally appeared.
[00:38:40] It looked off, right? Critics said it was a deep fake. A week later, soldiers attempted a coup, and they pointed to the video as part of their justification.
[00:38:52] Experts who examined it could never establish that it was a fake.
[00:38:58] And as a video no one could prove was fake.
[00:39:04] Help start a coup.
[00:39:06] The most deep, you know, the most dangerous deep fake isn't the one you believe.
[00:39:12] It's the one real video you refuse to believe for a century. A recording could end an argument. The tape, the photograph, the video. We have it on camera. That used to be the sentence that settled things, right? Oh, it's on camera.
[00:39:31] That sentence is dying.
[00:39:33] And when evidence dies, what's left isn't truth, it's loyalty.
[00:39:39] You'll believe the video if it's about the other side and dismiss it if it's about yours.
[00:39:45] We know that now, the future. And I'm going to be more specific than is quite honestly going to be comfortable for a lot of folks, because a prediction is only useful if you can be wrong about it. Okay, so prediction one.
[00:40:03] By the 2028 presidential campaign, the most consequential AI in politics won't be an ad.
[00:40:09] It'll be a conversation.
[00:40:12] An agent that calls you, speaks your language, knows your issues, argues with you, remembers you, and has that conversation with 10 million other people at once.
[00:40:21] Not one candidate. 100 million versions of one candidate. The early drafts already exist.
[00:40:28] Agents can be deployed just like that today.
[00:40:33] They come from every single direction, too. In 2024, a super PAC backing Democrat Dean Phillips launched an AI intimidate, you know, imitation of a candidate. Until OpenAI cut off its developer. Okay. The push to let AI voices call your cell phones. Comes from a conservative group. Nobody owns this future. Everybody is building it.
[00:40:57] Prediction 2.
[00:40:59] Somewhere in the democratic world, a major candidate will be caught on a genuine recall, genuine recording of calling it AI and win anyway, because nobody could prove otherwise fast enough for it to matter.
[00:41:17] And some of you are going to tell me that this has already happened. You might be right, but that's the point.
[00:41:23] And then point three, a losing side. Somewhere. We'll contest an election on a claim that AI did it. Fake videos, fake calls, fake crowds, fake votes, maybe even. And the truth is going to be unknowable, not false. Unknowable.
[00:41:40] Slovakia's lesson at a national scale. Nobody can prove it changed the outcome, and nobody can prove it didn't.
[00:41:49] Now, in the final segment, I'll tell you how we're going to fight back.
[00:41:56] Some of the answers will make free speech advocates and election officials equally uncomfortable.
[00:42:02] That's usually how you know you're close, when both sides don't feel really good about what the solution is.
[00:42:09] So stay with us. We'll be back after a few messages from our sponsors.
[00:42:28] Foreign.
[00:42:45] Welcome back. I'm Dr. Alan Badot, and this is the last segment of tonight's program. Now, I've spent three segments on the problem.
[00:42:53] Now, let me spend this one on the fight, because it is going to be a fight. And I'll start with the idea, I believe, more firmly than you know any other in this last hour.
[00:43:04] We are going to lose the war on fakes. We just are.
[00:43:09] I build AI for a living, and I say this as someone who knows exactly how good the forgers and their tools have become.
[00:43:19] And I mean it technically, okay? Detection is an arms race. And in an arms race, the forger always moves second, which means the forward.
[00:43:32] They always get the last move.
[00:43:35] Every detector that's published becomes a training target for the next generation of fakes. It's just reality. It's the same in cyber security.
[00:43:43] Human eyes have always lost eyewitness testimony. Not as good.
[00:43:50] Used to be. Oh, it's on camera anymore.
[00:43:54] Software detectors are losing. If our plan for democracy is to catch every fake in time, we don't have a plan. We really don't.
[00:44:02] So we got to stop trying to. To prove what's fake.
[00:44:07] Let's try to start proving what's real now. The technology for this exists, okay? And it's called content, provenance. A camera, a phone, a broadcaster attaches a cryptographic signature to a piece of media at the moment it's created.
[00:44:26] And every edit after Gets recorded.
[00:44:32] Oh, sound like Janice. AI's trustworthiness and its agents and how it tracks behavior. You betcha. It's the exact same thing.
[00:44:43] Think about it as a chain of custody for a video.
[00:44:47] Just like a custody chain of thought or an AI orchestrator or an agent's performance.
[00:44:55] Same thing.
[00:44:58] So the law is starting to move in California. The AI Transparency act became operative this August.
[00:45:10] The largest generative AI providers now have to embed hidden machine readable disclosures in what their systems produce, often visible labels and give the public a free tool to check whether the content came from their system.
[00:45:28] Large online platforms come next. It's coming in January and by 2028, makers of cameras and phones selling in California will have to offer the option to sign what their devices capture.
[00:45:43] Read the direction of that.
[00:45:47] Within a few years, the burden flips.
[00:45:51] Real media will carry a signature.
[00:45:54] Unsigned media will be presumed unverified.
[00:45:59] Brazil has already written an early version of that into their election rules. Right. Challenge a video and the publisher has to prove it's real.
[00:46:07] But I promise you a little discomfort. So I don't like to not deliver right. So here's the catch. Nobody in the Providence business likes to say out loud when reality requires a signature, reality becomes a product.
[00:46:23] The newest phones will sign their videos. The major networks will sign their broadcasts. The well funded campaign will sign its ads. And the bystander filming misconduct on a seven year old phone. What happens?
[00:46:40] The the whistleblower with a leaked recording.
[00:46:45] Their truth arrives unsigned.
[00:46:48] In a world that presumes unsigned means fake. The people with the least power have the hardest time being believed.
[00:46:58] That's right.
[00:47:01] Verified reality could become a luxury good.
[00:47:08] The powerful get provenance.
[00:47:12] Everyone else, they get rumors.
[00:47:15] If we build this, we have to build it for everyone.
[00:47:20] Or we gotta have an automated way that the oldest inequalities you know can can come. And that is? You know whose words count?
[00:47:31] No, the faces.
[00:47:33] Here's my proposal.
[00:47:35] Built on an idea that I've argued on this program for months. And we built into our own AI platforms with Janus AI and Atlas and all the other ones. The machine owns the synthesis, the human owns the accountability.
[00:47:47] I call it stand by your face. Very simple. And it's got three parts. Approve it.
[00:47:53] Any political ad that puts a real person's synthetic face or voice on a screen must be personally approved by the candidate. Simple as that. Or by the committee's chair, on the record and by name. Very simple. Not by a consultant, not by the vendor. Every party, every pack. The same rule. I approve this message becomes I approve this face.
[00:48:16] Next, you gotta label it. Not a two second disclaimer at the top that vanishes before you read it.
[00:48:23] A persistent label on the screen the entire time the synthetic face is visible, as clear as a food label.
[00:48:34] If you're proud of your satire, you'll have no trouble leaving the label on.
[00:48:42] Next. You gotta own it because liability sits with the person who approved it if the ad approves. Deceptive under what standards the court allows. The name on the approval is the name that answers. It's not a shell committee that dissolves on November 4th anymore.
[00:49:00] Because that's ridiculous.
[00:49:03] Notice what this doesn't do, though. It doesn't ban anything.
[00:49:08] It doesn't let the government decide which jokes are funny. Requiring disclosure of who is speaking has a long history in American election law. It's why you hear paid for at the end of every ad. Even the Supreme Court that decided, you know, the Citizens United upheld disclosure 8 to 1. This proposal stays on that road and it puts the burden where it belongs on the human who chose to use the machine. Just like everything else that we are trying to do on this show.
[00:49:38] You cannot push accountability down the road.
[00:49:42] Doesn't work.
[00:49:44] It doesn't work. Especially for elections.
[00:49:47] Especially doesn't work.
[00:49:53] What would change tomorrow if this was law? Okay, the duck would still be legal.
[00:49:58] The dancing congresswoman would still be legal. The handcuffs on either side would still be legal. But every one of them would carry a name on screen the whole time.
[00:50:11] Now, a researcher at the Starling Lab who argues against bans says a candidate's creative choices should simply reflect on the candidate.
[00:50:21] I agree.
[00:50:22] So make them stand where the voters can see them and let the voters judge the joke.
[00:50:27] Is it funny or is it stupid?
[00:50:32] That's right, I said stupid. Because oftentimes it is because they think that we're sometimes. Well, most of the time they think we're idiots. Right? We believe anything that we see. Well, that's not the case anymore, especially if you put their name on it Now.
[00:50:49] No new, you know, no new synthetic depictions of real candidates in the final three days before an election.
[00:50:56] I like that. I think we should do what Brazil does.
[00:50:59] Now, the argument against it, the First Amendment, of course, and a real risk that a ban on synthetic content turns into a ban on satire.
[00:51:08] The argument for, you know, the final 72 hours are when a lie can outrun any correction. And Slovakia has taught us that. So, you know, we got to draw the line somewhere. And maybe it's 24 hours, maybe it's 72. Maybe it's 48, but there's got to be a compromise in there. But there's got to be something like that.
[00:51:27] Disqualification is the next one I think would be very interesting to explore. Brazil allows its electoral court to strip a candidate's registration over deepfakes should a candidate who knowingly deploys a deceptive defect of an opponent lose their place on a ballot?
[00:51:49] Maybe. The argument against, though is pretty powerful. You would be handing elections to judges, and I don't feel good about that either. And so every losing campaign is, of course they're going to sue and the argument is going to be simple. The penalty for counterfeiting your opponent should be larger than the benefit. Right now it's really close to zero.
[00:52:09] So we'll see about that.
[00:52:13] Now, the last one around. Thinking about the AI announcing itself. If the FCC opens the door to AI voices calling your cell phone, all of us are going to be really ticked off again.
[00:52:28] But every one of those calls should have to say in its first sentence, I am an artificial voice, slick. The argument against campaign say it would hurt engagement.
[00:52:39] Now that's also an argument for it, though.
[00:52:42] No?
[00:52:45] Here is the uncomfortable truth. Okay? A 2024 YouGov poll for tech Policy Press cited by the Wesleyan Researchers also found 78% of registered voters favored banning AI content that makes deceptive claims about candidates.
[00:53:04] If that still holds, I bet it's higher though. But if that still holds, the public's already made up their mind.
[00:53:10] The courts have signaled that broadbands run straight into First Amendment.
[00:53:16] So the fix most Americans want is the fix the Constitution won't allow.
[00:53:20] That gap is where the next decade of this fight is going to be.
[00:53:26] And it's not going anywhere.
[00:53:28] So we have to think about that in those contexts now.
[00:53:35] Until then, here's what you can do in the next three weeks. Okay?
[00:53:41] When a clip makes you furious, wait about an hour before you share it.
[00:53:45] Fury is the delivery mechanism.
[00:53:50] Ask who signed it, not who posted it.
[00:53:53] Who stands behind it. If nobody does, treat it as unverified, no matter how real it looks.
[00:54:01] And last one, be suspicious in the final 72 hours. Especially anything that surfaces in the last three days before an election has timed so you wouldn't have time to check it.
[00:54:15] The timing's the confession, okay? That's what a lot of fraud based algorithms are built on. Timing.
[00:54:23] Okay?
[00:54:24] Think about that now before I go. Tonight's wardrobe, sponsored by Canali tailoring, is honest work. A suit or a sweater either fits or it doesn't. And the House that made it puts its name inside the jacket or on the tag.
[00:54:44] My thanks to them for always dressing the show. And a reminder that every episode of AI Today is available on demand at Nalmedia tv and many immediately after it airs.
[00:54:57] Now, here's what I'm going to leave you with. In 2022, or in 2002, I mean, Congress asked candidates for a single sentence. I approve this message. It wasn't technology law, it was accountability law. It said, if you're going to speak about your opponent, show your face.
[00:55:13] 24 years later, the face is the thing that now can be faked.
[00:55:20] The machine can make any face, any voice, any crowd, any moment that never happened.
[00:55:27] That's synthesis. And the machine is already better at it than will ever be.
[00:55:35] What the machine cannot do is stand behind it.
[00:55:39] Someone has to approve the face. Someone has to sign the real document. Someone has to sign the real commercial.
[00:55:46] And on November 3rd, someone has to decide what to believe. Okay, make sure it's you.
[00:55:54] I'm Dr. Alan Badot. This has been AI today. Thank you for watching. Good night.
