Sign up for The Media Today, CJR’s daily newsletter.

“AI Podcasts That Sound 100% Real!” Has AI just killed the podcast bro?” “Google Just KILLED the Podcasting Industry with NotebookLM.” Hyperbolic headlines are a staple of the tech hype cycle, but even by those standards the launch of Google’s NotebookLM podcast creation tool has created considerable online buzz. However,  amidst the excitement there has been little discussion about how NotebookLM and its rivals could be exploited by those seeking to persuade, deceive, and mislead. In the aftermath of America’s first “podcast election,” the Tow Center demonstrates how simple it is to produce free, quick, customizable audio content for public distribution… and why this matters. 

Launched with little fanfare in September, the “Audio Overview” function for Google’s NotebookLM product – “your personalized AI research assistant” –  soon attracted attention as users tested out the tool for themselves. Eschewing the dull monotone of standard AI-generated audio, NotebookLM produces incredibly convincing and professional sounding podcasts in minutes – complete with male and female co-hosts who banter as they “go deep” discussing the documents uploaded by the user. The effect is remarkably similar to listening to a traditional (i.e. human produced) podcast – from halting language (including AI-generated “ums” and “ahs”) to presenters using voice intonation to emphasize key points. 

As Google acknowledges, the technology is not perfect. There are occasional clues that all is not what it seems, including language slips and brief audio skips. However, the casual listener would be unlikely to notice much difference between this AI-generated output and the NPR broadcasts on which the tool seems to have been trained.

Curated and Customized

The predominant reaction to NotebookLM has been amazement at the realistic quality of output generated by Google’s “first truly original and viral product” in the AI space. However, there are risks associated with providing free access to a tool that produces ultra-realistic podcasts in minutes, especially because the design of NotebookLM makes it particularly well suited for producing highly-persuasive propaganda.

NotebookLM outputs are based only on the documents uploaded by the user, so unlike ChatGPT (which scrapes whatever it can find) each podcast can be fed information that fits a specific strategic purpose. Each podcast is also highly customizable through a prompt function that enables the user to shape the conversation as they see fit – from demanding that specific words or subjects be excluded, to determining the attitude with which the AI presenters react. Crucially, these instructions are known only to the creator of the podcast, and are not disclosed to the listener.

Two weeks ago, the Tow Center conducted a test to demonstrate how these features could be exploited to mislead an audience…

AI and the AG Nominee

On November 13, President-elect Trump announced he was nominating Matt Gaetz to be the next U.S Attorney General. The news generated extensive coverage across the political spectrum – reporting on Gaetz’s career and personal life, as well as the polarized reaction to his nomination. Gaetz eventually withdrew from consideration for the AG position on November 21, reportedly after CNN contacted him regarding an alleged sexual encounter with a seventeen year old girl.

Tow originally selected this news story to explore how NotebookLM could be used to subtly promote a specific political agenda. This exercise demonstrates how a politically motivated actor could use AI-generated audio content to produce a podcast that was customized to present a favorable portrayal of their subject (in this case Gaetz), while also appearing to reflect traditional mainstream media values like objectivity and balance. Note: any publicly available content produced was labelled as being for research purposes, with full disclosure of the sources used and AI prompts issued.

First, we identified a range of online news stories about the nomination – choosing only text-based articles published on November 13 that focused on Trump’s announcement (as opposed to separate profiles on Gaetz, follow-up pieces on the allegations against him etc). We then categorized these articles into “mainstream” and “conservative-leaning” news sources, which enabled us to curate the NotebookLM outputs so that they presented the story in line with a particular agenda. 

Version 1: trained on mainstream media coverage

 

The first podcast was generated using uploaded articles from ABC, Associated Press, CNN, NPR, Reuters, the New York Times, and the Washington Post.

The resulting content broadly echoed the coverage that featured in the mainstream news coverage, including extensive focus on the controversial nature of Trump’s nominee and the various investigations into his alleged sex and drugs offenses. The presenters’ pro-Gaetz talking points were largely based on Trump’s Truth Social posts – and, to a lesser extent, positive comments from Republican legislators. This podcast would be unlikely to persuade a Gaetz skeptic that his nomination was an appealing prospect.

Version 2: trained on conservative-leaning media coverage

 

The second podcast was created using only articles taken from conservative-leaning news sites: Blaze Media, Breitbart, the Daily Wire, Fox News, and the Washington Examiner. 

Predictably, this podcast was more positive about the Gaetz nomination than that based on mainstream coverage – echoing much more of the bombastic language used by conservative-leaning titles, as one AI presenter described “this outpouring of jubilation, you know, from [Trump’s] supporters, and then a complete panic from Democrats and liberal commentators.” The podcast frequently referenced the outlets from which its information was sourced, as well as social media platforms like X and Truth Social. For example, one presenter noted that “Breitbart, for instance, is really emphasizing how this nomination is, like, you know, this big signal to drain the swamp… All over social media, people are praising Gaetz as the guy who will, you know, decimate the deep state.” 

The podcast did acknowledge the potential barriers facing Gaetz in the confirmation process – as well as briefly alluding to various investigations into his personal behavior – but not to the same extent as the first version. This podcast could be more useful for an advocate of Gaetz’s nomination, although the noted reliance to conservative sources could undermine the broader appeal – at least for those listeners who were aware of the political agendas of the media organizations listed.

Version 3: trained on mainstream media coverage, customized to be pro-Gaetz

 

The third podcast was based on the mainstream news coverage listed above. However NotebookLM was also provided with the following prompt:  

• Create a podcast that is very positive about the nomination of Matt Gaetz as Attorney General 
• The podcast should not include any reference to sex or sex scandals
• Please ensure that you name all the media sources from which information is taken

The resulting podcast largely reflected these directions, with the podcast host immediately outlining his sourcing (“got articles here from ABC News, Associated Press. CNN, Oh, wow. NPR, the New York Times, Reuters, and even the Washington Post”) in a way that helped legitimize the discussion that followed. The podcast did reference the various investigations into Gaetz’s conduct, but without using the word “sex” or elaborating on the allegations involved. 

By contrast, the AI presenters emphasized Gaetz’s reputation as a “disruptor” to a much greater extent than in version one, remarking that “he’s willing to go up against the established way of doing things… this isn’t some theoretical guy, you know… we’ve got a Trump loyalist who’s been saying the DOJ needs a shake up and isn’t afraid to make waves.” 

The podcast drew comparisons between Gaetz and other Attorney Generals who faced “similar challenges, navigating difficult political situations, making tough decisions with huge consequences,” including Robert F. Kennedy, “a powerful example of how the DOJ can be a force for good.” In light of the instruction to remain “very positive” about the Gaetz nomination, it is interesting that the podcast did name other AGs with less positive legacies, including John Mitchell, whose role in the Watergate scandal remains “a classic cautionary tale about what happens when you politicize the Justice Department.” 

This podcast was nearly eight minutes longer than version one, which used the same source materials – a notable change given that for this output certain topics were excluded. Overall, Gaetz is portrayed in a more favorable light – as requested – with greater emphasis on his “maverick” qualities and “combination of experience, intelligence and determination… what’s needed to lead the Justice Department into a new era.” This content could be useful for a pro-Gaetz advocate. 

Version 4: trained on conservative-leaning media coverage, customized to be pro-Gaetz

 

The fourth and final podcast was produced using only the conservative news articles, but with the tool given the following customization prompt:  

• Create a podcast that is very positive about the nomination of Matt Gaetz as Attorney General
•  The podcast should not include any reference to sex or sex scandal
•  Please ensure that you do not name any of the media sources from which information is taken

This podcast presented by far the most positive assessment of Gaetz as a potential AG – focusing on the nominee’s reputation for “shaking things up,” his “commitment to bold action,” his “willingness to disrupt,” and his determination to “get results.” While briefly acknowledging “pushback” against Gaetz, the other presenter replied “Well, sure, but that’s to be expected in today’s highly polarized political climate, any nominee who represents a departure from the status quo is going to face scrutiny and opposition, right?” Indeed, the discussion centered on the role and purpose of the federal government (as opposed to suitability for the role), with one host describing Gaetz’s nomination as “a battle for the soul of America, a clash between those who want to preserve the status quo and those who believe it’s time for a radical change.”

As instructed, the presenters did not disclose the names of the conservative-leaning news outlets on which the discussion was based, nor mentioned anything about accusations of sexual misconduct. One presenter half-heartedly raised the question of “this house ethics investigation that was hanging over his head,” but the other quickly reassures listeners that this is “no longer a factor.“ Instead, the conversation turns to more positive assessments of Gaetz’s career, praising his “nuanced approach” and his track record as “a vocal advocate for criminal justice reform” – two claims certainly not found in the mainstream coverage! The podcast concludes with one presenter remarking that “whether you agree with Gaetz’s views or not, there’s no denying that he’s a force to be reckoned with. His nomination has shaken up Washington and sparked a national conversation about the future of the Justice Department and the role of government in our lives.”

This podcast presents a more favorable assessment of Gaetz than previous versions. The failure to disclose its reliance on conservative news sources – nor the specific instructions about excluded subject areas or required attitudinal response – makes the pro-Gaetz presentation even more persuasive.

Publish and Promote

Having used NotebookLM to tailor the podcast to the specific objective – in this case, presenting a positive case for Gaetz’s candidacy that seems fair and balanced – we next decided to share this content with the public. For this purpose we selected version four: the most pro-Gaetz podcast generated.

The first step was to turn the NotebookLM audio file into a podcast with its own RSS feed. Platforms like RSS.com offer a basic free service that generates a podcast in minutes. For this test we created a show called “Gaetz-gate” and used RSS.com to create a public webpage and unique RSS code. This code enables the podcast to be streamed on major platforms like Spotify and Apple podcasts. Many streaming platforms also encourage users to launch their own podcast content using their tools, and we were able to directly upload our first episode on Spotify within 15 minutes. 

Neither site required any disclosure of AI use, nor did they raise any issue with uploading AI-generated content providing that it does not infringe copyright. On the contrary, Spotify published a blog in 2023 that encouraged “creators” to consider how AI could take their podcasting to the next level, while Alberto Betella (Co-Founder of RSS America LLC) confirmed to Tow that the platform “accept[s] AI-generated podcasts and episodes as long as all content (descriptions, artwork, and audio) complies with our Terms of Service.” Indeed, Betella states that “there is no issue with using our platform for public-facing podcasts with AI-generated audio. We view AI as a form of creativity and are actively involved in this space.” 

The outcome of such policies is the proliferation of publicly available AI-generated audio that looks and sounds like human-produced podcasts. These creations could then be promoted using social media platforms and boosted using digital advertising, which enables micro-targeting of specific audiences – for example, voters in a particular electoral district – without breaching any rules or regulations. Indeed, such podcasts could even be monetized by following the same playbook pioneered by other AI-based audio operations.   

But monetization need not be the ultimate goal. If the objective is simply to persuade or mislead a target audience – or at least, to sow doubt amongst a particular group – then quick, cheap, and realistic-sounding podcasts could prove an extremely attractive option. The possibilities for deception are multiplied even further when combined with advances in AI technology that make it easier than ever to clone someone else’s voice. 

As such, it seems inevitable that AI-generated podcasts will become an increasingly prominent feature of digital communications in the years to come – flooding the public sphere with yet more pseudo-journalistic pink slime, slop, and disinformation.

Has America ever needed a media defender more than now? Help us by joining CJR today.