Convincing someone is expensive. Reminding them you exist is nearly free

It's the wrong question, and it's wrong for a reason the political persuasion literature has documented for years without the public conversation catching up. A large language model writing microtargeted messages does not persuade in a way that's measurably different from one writing generic messages. In a preregistered randomized trial published in PNAS, personalization added nothing statistically distinguishable. The advantage of language models sits in the informational density of the message, not in knowing who you're sending it to, a point the larger Science study on the levers of political persuasion with conversational AI confirms at scale: roughly half the persuasive variance in AI dialogue comes from packing in fact-checkable claims, not from tailoring the pitch to the person.

If you stop there, the story is reassuring. AI has no persuasive superpowers, the fears are overblown, everyone goes back to sleep.

But persuasion isn't the only channel through which an election moves. There's a second one, older than machine learning and far cheaper: mobilization. Not changing the mind of someone who already votes, but getting someone who'd stopped voting to show up. And here the cost calculus flips.

The experimental literature on door to door canvassing measures tiny effects. Real contact raises turnout by a few percentage points, but the real contact rate is low, and the aggregate effect collapses to a fraction of that. Phone calls do less. Mail does even less still, a pattern a recent meta-analysis of voter mobilization tactics confirms across decades of field experiments. For decades, mobilization has been a game of tiny margins, won one human contact at a time, at real cost per contact.

What AI changes isn't the size of the effect per contact. It's the number of contacts a campaign can afford. If writing a thousand variants of a message, running automated conversations, and generating a continuous stream of on brand content costs a fraction of what it cost a year ago, then the economics of mobilization stop being a budget constraint and become an infrastructure constraint, a shift the LSE Public Policy Review's analysis of AI in election campaigns already flags as a structural advantage for whoever adopts fastest and faces fewer reputational costs for doing so. You're not buying more effective persuasion. You're buying volume.

That's where the second piece comes in, the one about who owns the pipe the messages flow through rather than who writes them. There's a structural difference, not a rhetorical one, between buying political advertising, which is regulated, traceable, and (in Europe, since autumn 2025) banned outright on the major platforms, and owning the algorithm that decides what people see. Whoever controls the infrastructure doesn't need to follow ad transparency rules, because what they're distributing isn't an ad. It has been documented, by name, that a platform owner altered the recommendation algorithm to boost his own profile's visibility, a case study laid out in Paddy Leerssen's From Murdoch to Musk and echoed in the DSA Observatory's analysis of shareholder control, which notes the EU's own transparency law regulates market concentration but has nothing to say about a single individual's controlling stake. That part isn't a hypothesis. It's a studied case. What remains open, and here is where I am explicitly stating a hypothesis rather than a fact, is whether the same kind of control has been exercised, with the same logic if not the same platform, to favor a specific party in a specific election. We don't know, because no one outside the company has access to the logs that would show it.

There's a third piece that ties the first two together, and it's the most unsettling because it requires no manipulative intent at all to function. A randomized experiment, users assigned by chance to an algorithmic feed or a chronological one, over seven weeks, thousands of people, found that the algorithm doesn't just shift what people think in the moment. It changes who people follow. And those new follows stick even after the algorithm is switched off. The effect doesn't fade. It settles into a new, stable social network that keeps working on its own long after the initial push is over. That's the finding in Gauthier et al., "The political effects of X's feed algorithm," Nature, and it is, of the three pieces here, the one with the cleanest causal evidence behind it.

Put together, these three pieces sketch a precise mechanism: AI lowers the cost of producing volume; platform ownership routes around the rules built for purchased advertising; repeated exposure in a feed builds networks that outlive the intervention that created them. None of these three steps requires that a voter was "convinced" by machine-written text. It only requires that reaching them cost less effort, and that once reached, they stayed hooked.

That's a hypothesis, and it should be treated as one, not flattened into a neutral timeline that simply lines up a billionaire, a platform, and an election result.

This analysis, though, isn't necessarily tied to Germany's recent elections, not least because the full chain has never been measured for that case: no one has estimated how many of the 170,000 new voters the far right picked up in Saxony-Anhalt passed through this exact mechanism (Infratest dimap voter flow data), and maybe that's exactly why we keep talking about it as if we did.