The Meme Machine
AI video got realistic and remixable at the same time.
AI video passed the visual Turing test sometime in 2025. In its State of Generative Media Volume 1, fal.ai documented 985 new model releases last year, with major generative AI breakthroughs arriving every four to six weeks.
Even though new models can create media indistinguishable from reality, realism isn’t the main story here.
Remixing is. The entire visual language of internet culture is now in a production pipeline that costs very little to run. Any media is now remixable and memeable through face swaps, style transfer, and character consistency across scenes.
The impact is what we’re seeing in the Texas primaries and in Loudoun County right now: AI that is simultaneously realistic enough to look real and cheap enough to be absurd. Campaigns are using both—and capturing attention like never before.
The Signals This Week
1. Texas Primaries Become the First AI Ad Proving Ground
Texas Tribune reviews five months of Facebook and YouTube ads across major races
The Texas Tribune reviewed political advertising across the state’s March primary and found AI-generated content at every level, from the U.S. Senate race down to state House contests. Ken Paxton’s campaign released a synthetic video of John Cornyn dancing with Jasmine Crockett. Cornyn posted an AI deepfake of Wesley Hunt as a “show dog” with no disclosure. Crockett ran an AI ad featuring herself in a massive crowd of supporters, and her campaign declined to confirm whether the crowd was synthetic.
Some ads are obvious satire while others are sophisticated enough that, as the Tribune noted, the quality makes it difficult to confirm whether specific elements are AI-generated. A Texas House bill requiring AI disclosure in political ads passed the House in 2025 but stalled in the Senate.
Last month, Crockett got flack for using AI to generate a crowd in her video. Starting with the obvious… If you’re a politician, using AI to generate your crowd is not a good PR move. It’s even worse to deny it when you’re called out. The campaigns that are open about AI use will excel in an environment of shifting norms and policy.
Read more: Texas Tribune
2. X’s Algorithm Shifts Political Attitudes, and the Effect Doesn’t Reverse
Nature publishes landmark field experiment on 4,965 users showing persistent conservative drift
A randomized field experiment published in Nature this week assigned nearly 5,000 active X users to either the algorithmic “For You” feed or a chronological feed for seven weeks in 2023. Users switched onto the algorithm became 4.7 percentage points more likely to prioritize Republican-favored policy issues like crime, immigration and inflation afterwards. They grew more favorable toward Trump and shifted in a pro-Russia direction on Ukraine. Conservative posts were roughly 20% more likely to appear in algorithmic feeds, while traditional news was demoted by 58%.
Turning the algorithm on shifted attitudes, but turning it off did not reverse them. Users exposed to the algorithmic feed followed more conservative activist accounts and kept following them after switching back to chronological. The algorithm permanently reshaped their information environment.
Previous experiments on Meta’s platforms found no comparable political effects, suggesting X’s recommender system operates in a structurally distinct way.
Read more: Nature
3. Meta’s Manus Acquisition Signals the End of the Media Buyer As We Know It
$2 billion deal for an AI agent company points to full-stack automation of ad operations
Meta closed a $2 billion-plus acquisition of Manus, a Singapore-based AI agent startup generating over $100 million in annual recurring revenue just eight months after launch. The most interesting application is in their ads platforms.
This may mark a shift in the role of media buyers. Manus has a connector to Meta Ads Manager where it can analyze campaign performance and generate reporting. Tools like it are likely to collapse the remaining manual work: toggling budgets, optimizing placements, routing creative across channels. Media buyers will level-up into full-stack growth strategists or be replaced by the systems they currently operate.
Read more: CNBC
4. The First-to-Know Tax: Journalism’s 3-Minute Monetization Window
A $211,900 investigation can be extracted by AI in seconds
It costs $211,900 to produce a news investigation and $0.02 for an AI to extract its value. That ratio, 10 million to one, is what one observer calls the “First-to-Know Tax.” Within minutes of publication, AI systems crawl, index, summarize and serve the essential value of a story, often before a reader considers clicking through.
For 200 years, the business model of journalism was the advantage of knowing first and monetizing the information gap. AI collapsed that advantage to minutes. The monetization window after publication is shrinking toward zero.
Of course, if paywalls aren’t the goal, AI amplifies earned media faster and further than ever before, which means a well-placed story in a credible outlet or owned investigation has compounding value as AI systems cite it in perpetuity. The trust tax that used to protect publishers now protects your message.
Read more: Francisco Marconi
5. Generative Media Hits 985 New Models in One Year
fal.ai’s State of Generative Media report documents the pace of change across image, video, audio and 3D
2025 was a major year for generative AI. fal.ai published the first comprehensive State of Generative Media report, documenting 985 new model endpoints integrated into their platform in 2025 alone: 450 for video, 406 for image, 59 for audio, and 35 each for 3D and speech. Major video generation releases arrived every four to six weeks. Performance leadership changed hands multiple times.
AI video quietly passed the visual Turing test for untrained observers in 2025. Image editing achieved real-time speeds with character consistency and style transfer. Audio synthesis reached production quality. The tools used to create content in Texas and Loudoun County (below) were barely functional 12 months ago.
Read more: fal.ai State of Generative Media
6. Loudoun County Republican Committee Released AI-Generated Attack Video on Spanberger
Off-the-shelf tools used to create inflammatory video which quickly went viral
The Loudoun County Republican Committee has been running a series of AI-generated satire targeting Virginia Governor Abigail Spanberger. Their first video features vignettes of her “taxing dog walkers,” “grabbing guns,” and “hating Virginians,” ending with a scene of her burning a picture of Jesus. The video is cheap, shocking, and exactly the kind of content that makes editorial review seem like a quaint concept. It also went viral. AI makes it possible to produce vivid, visceral content for almost nothing, and in this environment, absurdity is its own distribution strategy.
That combination, low-cost, plus viral reach regardless of quality, is what makes cheapfakes worth watching heading into the midterms.
Watch: Loudoun County Republicans
The primaries are the first real environment where all of this is running simultaneously: realistic enough to pass as true, absurd enough to go viral, cheap enough that any county committee can do it. The campaigns that treat that as a creative opportunity rather than a liability will have the upper hand in the battle for attention.
Best,
Ben