Deleting two convenient assumptions tips a consensus model into an echo chamber

They did not have to build a control. One already existed. In the network models Kong, Hein and colleagues started from, built by earlier researchers, each agent passed along the actual data behind its choice, and every agent could attend to all of its neighbours at once ^[10][12]. Those models never produced echo chambers ^[12]. Hein put the first assumption this way: "It's as if I express my opinion to you, but instead of telling you what my opinion is, I tell you everything that I ever learned that caused me to form that opinion," he said ^[11].

Both assumptions are false of nearly everything alive. "One of the reasons why we found these assumptions interesting is that they appear to be violated in every system we could think of," Hein said ^[13]. An animal usually sees a neighbour's discrete action without the internal information that produced it, and it can attend to only a few neighbours at a moment ^[7][8]. So the team removed the assumptions, one at a time and together ^[14].

Stripped of both assumptions, the groups still reached consensus. "Individuals can reach a consensus decision but it becomes totally decoupled from the state of the world around them," Hein said ^[16]. In the paper's account, the two constraints make each individual extremely sensitive to the messages it receives, and the group settles into sending and receiving the same messages back and forth ^[9].

Stripping the assumptions separately matters for what the study can claim. The same model family produces no echo chambers when both idealisations hold, so the dysfunction traces to those assumptions and not to the wiring of the network ^[12][14].

Agents at the periphery are connected to the centre but also pick up information from their environment, and their change spreads through the group ^[17]. "It's really the free spirits, the individuals who have stayed at the edge and don't have many social influences, who are the first ones to change," Hein said. "They can set off a cascade." ^[18][19] Defection breaks the loop too: individuals who recognise their decisions are wrong and start ignoring the group ^[20]. "If some of the individuals in the network, even just a small fraction of them, do that some of the time, it will destroy an echo chamber," Hein said ^[21]. How small a fraction, and how often, the account does not say ^[26].

Army ants supply the vivid case; the models are general, covering birds in a flock, cells in tissue and ants in a swarm ^[5]. A mill begins when one ant walks in a circle and the ants behind follow the pheromone trail until they collapse and die ^[1][2]. "This is a physical instance of an echo chamber, where the animals involved actually die," Hein said ^[4]. A model can show that two constraints are sufficient to produce the loop ^[6][9]. Deciding whether real groups sit in that regime takes field or laboratory measurement.

The authors make the human comparison themselves. In people, the paper says, the phenomenon appears when individuals send and receive similar messages and lack external sources of differing information ^[23]. "It happens when people are telling each other the same thing, but there's also an implication that this group has become unresponsive to what's actually happening in the world around them," Hein said ^[24]. The tests ran on model agents. The paper, whose first author is Ling-Wei Kong, appeared in the Proceedings of the National Academy of Sciences on 22 September ^[22][6].