In a research paper, the first red flag did not appear. It came as a phone call. Voters in New Hampshire picked up their landlines sometime in late January 2024 and heard what sounded like Joe Biden calling on Democrats to abstain from the primary in that well-known, slightly raspy voice. Naturally, the voice was incorrect. It was artificial. However, because it felt more like a rehearsal than a stunt, the moment carried a kind of subdued dread that persisted long after the news cycle passed.
After nearly two years, the MIT Technology Review has finally voiced what many researchers had been murmuring. The magazine came to the conclusion that the era of AI-driven political persuasion is no longer a forecast in an article published on December 5, 2025. It’s already started. The alarm itself, which has been raised numerous times, is not what makes the verdict noteworthy; rather, it is the change in the supporting evidence.
Reiterated by Meta executives and echoed in reports from the Alan Turing Institute, the standard line was comforting for years. Indeed, AI had been used in elections, but its influence was minimal at best. Nick Clegg himself described it as “striking how little these tools have been used” to genuinely interfere with democratic processes in 2024. Only a small percentage of the more than 100 elections that the Turing researchers examined showed any signs of AI interference. That narrative persisted for some time.
The new studies then arrived. This month, two extensive, peer-reviewed studies revealed that AI chatbots could significantly change voters’ opinions, outperforming conventional political advertisements in side-by-side tests. These are not the kinds of figures that are typically published by researchers without hedging. This time, there was also very little hedging.
Those who follow this closely believe that the deepfake panic was merely a diversion. Although there is a genuine concern that viewers might be duped by a fake video showing a candidate admitting to a scandalous act, this is a relatively small risk that fact-checkers can pursue. The research on chatbots points to something more unusual and challenging to regulate. Convincing without shouting. Persuasion that listens, adjusts, remembers what you said three messages ago, and quietly leans on the emotional thread you didn’t realize you’d revealed.
The ease with which this technology has permeated daily life is difficult to ignore. People argue with chatbots about parenting, taxes, and whether or not to end a relationship. It is not at all a leap from that to political discourse. In the minor ways that don’t make headlines, it has already occurred.
The MIT article suggests something worth considering. Contemporary AI is more than just a message writer. It can instruct other AI systems to create the image, score the voice, select the platform, monitor the interaction, and improve the subsequent attempt. The authors described it as a coordinated persuasion machine. Macedonia has no meme farms. No weary interns at troll factories. The majority of voters will never be able to see the code that runs silently in the background during a campaign cycle.
It’s still genuinely unclear if this affects the results. Voting behavior is unyielding, disorganized, and more influenced by family, religion, and economic sentiment than by a single message. Mass persuasion campaigns tend to nudge rather than transform, according to decades of political science. Thus, it’s possible that the doomsday framing is still excessive. Perhaps.
However, there’s a sense that something has crossed over as you watch this play out. The 2024 robocall is now a clumsy first draft that reads almost like a museum piece. Whether American institutions, regulators, platforms, and voters themselves have any kind of plan is the question for 2026 and the presidential cycle that lies just beyond. Right now, the truth is that they don’t. Not at all. Furthermore, the machines aren’t holding out for one.
