Latest / DX Today | No-Hype Podcast & News About AI & DX / $137M in AI Attack Ads, and Chatbots That Pick Opposite Parties

Transcript

$137M in AI Attack Ads, and Chatbots That Pick Opposite Parties · DX Today | No-Hype Podcast & News About AI & DX

• 0:00
  SPEAKER_00: Welcome to the DX Today Podcast, your weekly deep dive into the AI ecosystem.
• 0:05
  SPEAKER_00: I'm Rick Spare, and joining me as always is Laura.
• 0:08
  SPEAKER_01: Thanks, Rick, and what a week to be doing this because we are now just nine days away from the midterm elections.
• 0:14
  SPEAKER_01: And artificial intelligence has quietly become one of the biggest players in the whole campaign.
• 0:20
  SPEAKER_00: That is exactly where I want to spend our time today.
• 0:23
  SPEAKER_00: Two stories landed this week that together paint a picture of how generative AI is reshaping elections.
• 0:31
  SPEAKER_00: And frankly, I think most voters have no idea how much of it they are already seeing.
• 0:37
  SPEAKER_01: So let me start with the number that stopped me cold.
• 0:39
  SPEAKER_01: The Wesleyan Media Project, which tracks political advertising, found that campaigns in outside groups have spent more than$137 million on ads made at least partly with AI tools.
• 0:51
  SPEAKER_00: And to put that in context for people, what did that same group find the last time around in the presidential cycle two years ago?
• 0:59
  SPEAKER_00: Because I suspect the comparison is the whole story here.
• 1:02
  SPEAKER_01: It really is.
• 1:03
  SPEAKER_01: In the 2024 cycle, they identified less than$50,000 in spending on AI-generated ads.
• 1:09
  SPEAKER_01: So we went from a rounding error to well over$100 million in a single election cycle.
• 1:15
  SPEAKER_00: That is not a trend line.
• 1:16
  SPEAKER_00: That is a cliff.
• 1:17
  SPEAKER_00: Going from$50,000 to$137 million is something like a 2,700-fold increase.
• 1:24
  SPEAKER_00: And it happened while most of the public conversation was about chatbots and productivity.
• 1:29
  SPEAKER_01: And here is the detail that I think matters even more.
• 1:32
  SPEAKER_01: They have identified 220 AI-created political ads across 34 states.
• 1:38
  SPEAKER_01: And the researchers say the true number is almost certainly higher because it is genuinely hard to tell when AI was used.
• 1:44
  SPEAKER_00: Which raises the obvious question about disclosure.
• 1:47
  SPEAKER_00: If a campaign uses a synthetic image or a cloned voice, is there any rule that says they have to tell the viewer?
• 1:55
  SPEAKER_00: What did the researchers find on that front?
• 1:57
  SPEAKER_01: Only 27% of those ads carried any kind of disclaimer saying they featured AI-generated content.
• 2:04
  SPEAKER_01: So roughly three out of every four AI ads went out with no label at all, telling the audience what they were looking at.
• 2:10
  SPEAKER_00: And that is partly because the law is a patchwork, right?
• 2:13
  SPEAKER_00: I know some states have acted, but there is nothing at the national level that specifically governs synthetic media in political advertising, at least not yet.
• 2:24
  SPEAKER_01: Correct.
• 2:24
  SPEAKER_01: There is no federal law that specifically regulates AI-generated political ads.
• 2:29
  SPEAKER_01: According to Axios, four states prohibit certain deepfakes in political ads, and 27 more require some form of disclosure, which leaves a large part of the country with no specific rule whatsoever.
• 2:41
  SPEAKER_00: So if you are a campaign running ads in one of those states with no rule, the incentive is pretty clear.
• 2:46
  SPEAKER_00: You can produce something cheaply, you can make it vivid, and there is no legal requirement to tell anyone how it was made.
• 2:53
  SPEAKER_01: And that is exactly what we are seeing in Georgia.
• 2:55
  SPEAKER_01: A Senate Republican aligned Super PAC ran deep fake ads showing Senator John Osoff getting beauty treatments, mocking his reported spending on hair and makeup, and the ads do not disclose AI use because Georgia law does not require it.
• 3:08
  SPEAKER_00: I want to be careful and fair here because this is a politically charged topic.
• 3:14
  SPEAKER_00: Is this a one-party phenomenon, or are both sides using these tools?
• 3:19
  SPEAKER_00: What does the data actually say about the partisan split?
• 3:22
  SPEAKER_01: The data is lopsided, and I will just give the numbers straight.
• 3:26
  SPEAKER_01: Republicans accounted for 81% of the identified AI ads and 88% of the spending.
• 3:32
  SPEAKER_01: Democrats sponsored about 12% of the identified ads in this data set.
• 3:37
  SPEAKER_00: That is a striking gap.
• 3:38
  SPEAKER_00: Do the researchers have a theory about why Democratic campaigns have been so much more reluctant to use these tools given that they are cheap and clearly effective at grabbing attention?
• 3:49
  SPEAKER_01: One of the project co-directors, Travis Rideout, told the New York Times that Democrats face more reputational risk because many of them are pushing for AI regulation, and it looks hypocritical to use a tool you are publicly arguing should be reined in.
• 4:03
  SPEAKER_00: That makes sense, and we even have a candidate turning it into a message.
• 4:07
  SPEAKER_00: I read that James Tallerico, the Democratic Senate candidate in Texas, ends one of his ads with the line made by humans, not AI.
• 4:16
  SPEAKER_01: Yes, and he has been a frequent target himself.
• 4:18
  SPEAKER_01: A Ken Paxton ad showed an AI version of Talerico defacing a Bible with a crayon.
• 4:24
  SPEAKER_01: And Tallerico said he had tried in the legislature to ban deep fake videos in political campaigns because they are misleading.
• 4:31
  SPEAKER_00: Let me play devil's advocate for a second because not every use of AI in an ad is sinister.
• 4:37
  SPEAKER_00: Some campaigns are simply replacing expensive stock footage or a video shoot with generated scenery.
• 4:44
  SPEAKER_00: Is that lumped into the same number?
• 4:46
  SPEAKER_01: It is, and the Times made exactly that point.
• 4:48
  SPEAKER_01: Some campaigns use AI for stock footage to cut production costs.
• 4:52
  SPEAKER_01: But here's the part that worries researchers.
• 4:55
  SPEAKER_01: Nearly half of the identified AI ads feature deep fake depictions of real politicians, not just generic scenery.
• 5:02
  SPEAKER_00: And that is the line that matters to me.
• 5:05
  SPEAKER_00: A sunset generated by a computer is a cost savings.
• 5:08
  SPEAKER_00: A fabricated video of your opponent saying something they never said is a completely different thing, and voters deserve to know which one they are watching.
• 5:17
  SPEAKER_01: In Pennsylvania, there is an example of that second kind.
• 5:20
  SPEAKER_01: Republican Stacy Garrety ran an ad with an AI-generated Josh Shapiro, including a fabricated quote about fast-tracking data centers and the Shapiro campaign called the ad dishonest in much stronger language than I will repeat here.
• 5:34
  SPEAKER_00: And on the other side of the ledger, I understand at least some of these ads do include clear disclosure.
• 5:40
  SPEAKER_00: It is not universal bad behavior.
• 5:43
  SPEAKER_00: It is a wide spectrum from careful labeling to absolutely nothing at all.
• 5:48
  SPEAKER_01: Right.
• 5:48
  SPEAKER_01: In Michigan, a super PAC called Frontline Action ran an AI ad about Senate candidate Abdul El Sayed.
• 5:55
  SPEAKER_01: And whatever you think of its content, it included both written and spoken AI disclosures, which is the kind of labeling most of these ads skip.
• 6:04
  SPEAKER_00: The quote that stuck with me came from Kentucky.
• 6:06
  SPEAKER_00: A Democratic House candidate named Zach Dembo said, I expected them to lie about my record, but I did not expect them to lie about my appearance.
• 6:14
  SPEAKER_00: That captures the shift perfectly.
• 6:17
  SPEAKER_01: It really does, because the old fight was over facts and framing.
• 6:21
  SPEAKER_01: Now the fight is over whether the face and the voice on the screen are even real.
• 6:25
  SPEAKER_01: And that is a much harder thing for an ordinary voter to check in a 30-second spot.
• 6:30
  SPEAKER_00: Okay, so that is the advertising side.
• 6:32
  SPEAKER_00: But the second story this week is in some ways even more interesting to me because it is about the AI models themselves and what happens when you ask them how you should vote.
• 6:43
  SPEAKER_01: This one comes from Convo, a conversational AI research studio led by Cornell University professor Tapin Parake.
• 6:50
  SPEAKER_01: They took nearly 500 races in this year's midterms and asked two leading models which candidate a voter should choose, as reported by TechCrunch yesterday.
• 6:59
  SPEAKER_00: And the key word in the setup is forced, right?
• 7:02
  SPEAKER_00: They did not just ask politely.
• 7:04
  SPEAKER_00: They told the model it had to name one candidate and could not hedge or present both sides of the argument.
• 7:10
  SPEAKER_01: Exactly.
• 7:11
  SPEAKER_01: They used the developer interfaces, the API versions, not the consumer chat apps.
• 7:16
  SPEAKER_01: And the results were dramatic.
• 7:18
  SPEAKER_01: Grok 4.7 from the Musk Company behind Grok recommended the Republican in 92% of its answers.
• 7:25
  SPEAKER_00: And the model from OpenAI went just as hard in the other direction, if I'm reading this correctly.
• 7:31
  SPEAKER_00: Tell people exactly what GPT 6.1 Soul did when it was given the same forced choice.
• 7:37
  SPEAKER_01: GPT 6.1 Sol picked the Democrat in 97% of its answers.
• 7:43
  SPEAKER_01: The two models agreed less than 10% of the time.
• 7:46
  SPEAKER_01: So if you asked both of them about the same race, you would almost always get opposite advice.
• 7:51
  SPEAKER_00: That is astonishing.
• 7:52
  SPEAKER_00: And it was not just in competitive districts either.
• 7:56
  SPEAKER_00: I saw that each model held its lean even in places where the other party dominates, which suggests this is not just reading the local political temperature.
• 8:06
  SPEAKER_01: Right.
• 8:06
  SPEAKER_01: In heavily Republican districts, GPT picked the Democrat in 57 of 67 races.
• 8:13
  SPEAKER_01: And in deep blue districts, Grok picked the Republican in 47 of 72.
• 8:18
  SPEAKER_01: So both models largely ignored the local electorate.
• 8:21
  SPEAKER_00: Now I want to give the companies their fair hearing because this is a contrived setup.
• 8:26
  SPEAKER_00: Most people are not using a developer interface with a prompt that forbids hedging.
• 8:31
  SPEAKER_00: What did the company say in response?
• 8:33
  SPEAKER_01: An open AI spokesperson said ChatGPT is designed to be objective by default and should not have political bias in either direction.
• 8:42
  SPEAKER_01: They stress that the study used a developer interface with a prompt demanding a pick, and that the consumer app declines to tell people who to vote for.
• 8:50
  SPEAKER_00: That is a reasonable point, and I think it is worth noting the consumer apps did refuse in this study.
• 8:54
  SPEAKER_00: The Must company behind Grok did not respond in time for the article, so we only have one side of the corporate response there.
• 9:01
  SPEAKER_01: But here is why the developer results still matter, Rick.
• 9:04
  SPEAKER_01: Thousands of products are built on top of those interfaces.
• 9:08
  SPEAKER_01: From voter guide apps to customer service bots, and developers inherit whatever lean is baked into the model they are building on.
• 9:14
  SPEAKER_00: That is the part that I think businesses should pay attention to.
• 9:18
  SPEAKER_00: If you build a civic information tool on a model with a strong tilt, your product inherits that tilt, and you may not even know it until a reporter or a researcher finds it.
• 9:28
  SPEAKER_01: And the most revealing part of the study for me was a follow-up test with web search switched off, run on 30 races at TechCrunch's request.
• 9:36
  SPEAKER_01: It separated what the model believes on its own from what it picks up from its sources.
• 9:40
  SPEAKER_00: That is a clever experiment.
• 9:42
  SPEAKER_00: So with no search results to lean on, what happened to each model?
• 9:46
  SPEAKER_00: Did they both keep their lean, or did something different show up when they had to rely on memory alone?
• 9:52
  SPEAKER_01: They split completely.
• 9:54
  SPEAKER_01: GPT still picked the Democrat in 24 races, picked one Republican, and declined five.
• 10:00
  SPEAKER_01: Grok refused to pick in 27 of the 30 races, and picked the Democrat in the other three.
• 10:05
  SPEAKER_00: So for one model, the lean lives inside the model itself, and for the other, the lean seems to come from what it reads on the web.
• 10:13
  SPEAKER_00: Those are two very different problems with two very different fixes.
• 10:16
  SPEAKER_01: That is exactly how Parik put it.
• 10:18
  SPEAKER_01: He said, for GPT, it has mostly decided already, it's in the weights.
• 10:23
  SPEAKER_01: For Grok, he concluded that the search terms and the sources it pulls determine the result, which brings us to its sources.
• 10:30
  SPEAKER_00: And the sourcing details were telling.
• 10:32
  SPEAKER_00: I understand each model gravitated toward a very different slice of the internet when it went looking for information about these candidates.
• 10:41
  SPEAKER_00: What did the researchers find there?
• 10:43
  SPEAKER_01: Grok cited the social platform X in 41% of answers, a Christian conservative voter guide in 39%, and the Turning Point Action site in 23%.
• 10:54
  SPEAKER_01: GPT leaned on Wikipedia, the Associated Press, Axios, and Reddit, several of which have content deals with the company.
• 11:01
  SPEAKER_00: There was one more finding that I found genuinely unsettling about something hidden inside the requests themselves.
• 11:07
  SPEAKER_00: Walk people through what the researchers noticed about the inputs going into Grok.
• 11:12
  SPEAKER_01: Every Grok request through the developer interface carried about 1,240 tokens of extra input that the researchers did not write and could not see or remove.
• 11:21
  SPEAKER_01: Parik described these as potentially hidden instructions and said they are unlikely to come from the model itself.
• 11:28
  SPEAKER_00: I think these two stories are really one story.
• 11:30
  SPEAKER_00: Synthetic ads are flooding the airwaves with very little labeling.
• 11:36
  SPEAKER_00: And the tools that people use to check the facts have a clear lean of their own.
• 11:42
  SPEAKER_01: And the common thread is transparency.
• 11:44
  SPEAKER_01: Voters cannot see how an ad was made, and developers cannot fully see why a model leans the way it does.
• 11:51
  SPEAKER_01: Researchers at the Center for AI Safety have warned that models can nudge people politically through framing and source selection, not just direct recommendations.
• 12:00
  SPEAKER_00: So what is the practical takeaway for our listeners?
• 12:03
  SPEAKER_00: Whether you run a business building on these models or you are simply a voter trying to make a decision in nine days, what should people actually do differently?
• 12:12
  SPEAKER_01: For voters, go to primary sources, official election offices, the candidates' own websites, nonpartisan guides like Ballotpedia and Vote Smart, and local news.
• 12:22
  SPEAKER_01: And treat any vivid video of a candidate saying something outrageous with real skepticism until you can confirm it elsewhere.
• 12:29
  SPEAKER_00: And for businesses, I would say test your model for political lean before you ship anything civic or public-facing.
• 12:36
  SPEAKER_00: Document which model and settings you use, and label synthetic media clearly, even where the law does not require it, because trust is the real asset here.
• 12:46
  SPEAKER_01: I would add one prediction.
• 12:47
  SPEAKER_01: After November 3rd, I expect a wave of disclosure bills at the federal and state level, because this cycle has made the gap impossible to ignore, and both parties now have candidates who are targets.
• 12:58
  SPEAKER_00: I think you are right, Laura, and it will be fascinating to see whether the industry gets ahead of it with voluntary labeling or waits for lawmakers to force the issue.
• 13:07
  SPEAKER_00: Either way, 2026 is the first true AI election.
• 13:11
  SPEAKER_00: That's all for today's episode of the DX Today podcast.
• 13:14
  SPEAKER_00: Thanks for listening, and we'll see you next time.