This is one of those warnings that researchers don’t give lightly. In a recent paper, a group of scientists from Berkeley, Oxford, Cambridge, Harvard, and Yale, along with Nobel laureate Maria Ressa, argued that something truly novel is emerging online: artificial intelligence (AI) agents that can act like coordinated groups of people, learn a community’s habits, and influence its opinions without ever disclosing what they are.
Calm reading is an odd experience. No far-off, theoretical risk is being described by the researchers. They mention elections in Taiwan, India, and Indonesia where early versions of the software have appeared, and they describe software that already exists in rough form. It’s not that bots will yell louder than people. It’s that they will sound just like people, gradually introducing skepticism into comment threads and group chats until a community’s consensus subtly changes.
Coordination is what sets this moment apart from the deepfake panic of a few years ago. It is possible to disprove a single phony video. It is much more difficult to identify and even more difficult to reverse a thousand tiny, patient accounts that mimic regional slang, post during human hours, and gradually steer a conversation. AI-driven accounts pushing young voters toward a sort of forced neutrality on delicate political issues have already been observed by researchers examining Taiwan’s information environment. This may seem harmless at first, but telling people to “stay out of it” is a position in and of itself.
Not everyone who is researching this believes it will happen soon. Some propaganda experts note that the majority of campaigns continue to employ antiquated, antiquated strategies rather than state-of-the-art AI swarms, primarily due to political operatives’ reluctance to entrust campaign strategy to a machine they do not fully control. A practical question that permeates the entire debate is how much real-world votes are influenced by online discussions as opposed to television, door-knocking, or a neighbor’s viewpoint over dinner. It’s still unclear if technology has outgrown its real impact or if influence is simply more difficult to quantify than alarm.
However, those closest to the underlying technology seem less dubious. Building these bot networks is now practically trivial, according to AI researchers who focus on multi-agent systems; it’s a weekend project for someone with basic coding knowledge and access to current language models. Real harm typically occurs in the space between the difficulty of building something and the difficulty of defending against it. Ransomware caused it to occur. Ten years ago, it occurred with coordinated disinformation farms. Those who have previously observed this pattern feel that it is happening again, albeit more quickly.
Platforms and governments are not completely immobile. A few tech companies have once again promised to identify and eliminate coordinated fake accounts prior to elections, and some lawmakers have pushed for the mandatory labeling of AI-generated political content. Platforms have previously made similar pledges and discreetly reduced enforcement once attention shifted, so it remains to be seen if those promises hold up under actual pressure.
The gap between capability and oversight is closing, and the 2028 U.S. presidential election is often cited as the moment when this technology could be used on a large scale. It appears that none of the interviewees believe the issue has been resolved or even nearly resolved. Most seem to concur on something more straightforward and unsettling: the tools are prepared well in advance of the regulations, and this discrepancy seldom resolves itself amicably.
