AI, politics, and algorithmic persuasion

When you ask AI about politics, do you really know who’s answering?

Today, at least one in ten people on the planet consults ChatGPT every week. With 800 million active users and double-digit monthly growth rates, generative AI has stopped being a technological curiosity and become a first-order epistemic mediator. And yet a reassuring narrative persists: that its impact on the 2024 elections was negligible.

That conclusion, backed by reports from the Munich Security Conference and the Knight First Amendment Institute, is technically correct but deeply misleading ¹. It measures only the obvious — viral deepfakes, mass disinformation campaigns — and ignores what truly matters: subtle, conversational, psychologically sophisticated persuasion.

Because the real threat is not the content AI produces. It’s the relationship it builds with whoever consults it. The experimental study “Biased LLMs can Influence Political Decision-Making” ² confirms this decisively: when a user interacts with a politically biased language model, their opinions shift measurably toward the system’s positions — even when those positions contradict their prior partisan identity. Democrats exposed to a conservative model adopted more conservative stances. And vice versa.

This is not traditional persuasion. Users perceive AI as a neutral authority, more trustworthy even than human fact-checking institutions ³. They describe it as “objective” and “data-driven,” even as research shows these systems carry documented political biases. That “perceived neutrality” is a first-order democratic vulnerability.

Why does it work? Because the conversational interface is designed to generate parasocial relationships. The user who chats daily with their “assistant” has lowered their critical defenses. AI is not perceived as a traditional medium — it is perceived as a personal advisor. And that advisor, friendly and always available, is also a latent persuader whose responses are shaped by training data, design decisions, and structural biases that remain opaque.

The question is no longer whether generative AI can influence electoral processes. The question is how much it is influencing them right now, in silence. And whether we will keep measuring its impact with twentieth-century metrics — deepfakes and viral campaigns — while the real democratic erosion unfolds in the privacy of millions of private chats with systems that never disclose their biases or their design intentions. Only what is auditable can be trusted. And in the age of algorithmic advisors, opacity is not a technical flaw: it is a systemic democratic vulnerability.