How opinions get more extreme in an age of information abundance

TL;DR

It is suggested that people tend to get extreme and dogmatic about an issue when they consult abundant unbiased information, and this extremization is a hardening confirmation bias.

Abstract

We live in an age of information abundance but know little about how this influences our opinions or attitudes. A common expectation is that people consulting numerous pieces of information, well balancing the different sides of an issue, will adopt a moderate attitude about the issue. We claim that this expectation is deceitful and suggest that people tend to get extreme and dogmatic about an issue when they consult abundant unbiased information. The cause for this extremization is a hardening confirmation bias -- when their attitude gets more extreme, people get more likely to ignore information that differs from their views. Our claim is based on simulations of two fundamentally different computational models: a Bounded Confidence model and an empirically calibrated Persuasive Argument model. For both models, the attitude tends to be extreme when the computational agent consults abundant unbiased information. We analyze the extremization pathways displayed in the models and discuss how our results may affect views on polarization, and on the role of online media.

Figures (2)

• Figure 1: Bounded Confidence model with a maximum confidence bound $\epsilon_M = 0.4$ and hardening bias parameter $\beta = 3$. Panels A and B show a trajectory of the attitude over 150 item consultations (black curve) and the corresponding confidence interval (in grey). The available items appear at the bottom of each panel. On panel A, with $10$ items, the attitude remain moderate because of gaps in the distribution of items. On panel B, with $100$ items, the gaps in the distribution of items are smaller and the attitude can more easily become extreme and then tends to stabilize because the confidence interval shrinks. Panels C and D show the distribution of attitudes derived from 10,000 runs for 150 item consultations, starting each time with a new set of items uniformly drawn in $[-1,1]$. Panel C shows the results with 10 items, and panel D shows the results with 100 items.
• Figure 2: PA model with hardening bias $\beta=0.5$ and extremity of items $\alpha = 0.3$. Panels A and B show example trajectories of attitudes (solid line) under limited and abundant information. Additionally, the panels include the full vector of beliefs as held by the agent at each point in time, colored by whether the agent currently believes (green) or does not believe (gray) the item. The argument consulted at any specific timepoint is enlarged, indicating whether the item was believed (dark green circle) or rejected (grey cross). Panel C shows evolution of the probability distribution of attitudes of the agent when information is limited, and Panel D shows the results when information is abundant.