Introduction
In traditional social psychology, "group polarization" is a well-documented phenomenon: when individuals discuss an issue within a group, the collective tends to adopt a position more extreme than the initial average of its members. As public discourse shifted to the internet, many researchers expected this trend to intensify, fueled by "information cascades" where early opinions trigger a herd mentality. However, this paper challenges that assumption. By analyzing millions of online reviews on Amazon and IMDB, the authors demonstrate that online opinions actually moderate over time—a phenomenon driven by the unique behavioral mechanics of digital platforms.
The Mechanism: Costly Expression Bias
To understand why online discourse resists polarization, the authors propose the concept of "costly expression bias." In an offline setting, the cost of participating in a conversation is relatively low. Online, however, writing a detailed review requires significant effort.
This creates a "Voter’s Paradox" for reviewers: why spend the time to write a review if thousands of people have already expressed the same sentiment? The authors argue that users are primarily motivated to contribute when their opinion significantly differs from the existing consensus. If a user agrees with a 4.8-star average, their contribution feels redundant. But if they believe the product deserves a 1-star rating, their review becomes a tool to "correct" the perceived inaccuracy. Consequently, the online environment naturally selects for contrarian voices, leading to a moderation of the average score over time.
Methodology
The researchers conducted a longitudinal, observational study focusing on the temporal dynamics of user ratings:
- Amazon Dataset: The study tracked the first 20 ratings for over 16,000 best-selling books. By analyzing the sequential order of these ratings, the authors could measure how the average changed as the volume of reviews increased.
- IMDB Dataset: To isolate the impact of effort, the authors compared two types of feedback on IMDB: simple star ratings (low cost) and written reviews (high cost). They analyzed the 50 highest-rated and 50 lowest-rated movies released after 2000 to see if the moderation effect held true for both "good" and "bad" products.
Key Findings
1. The Decay of Extremity on Amazon
The analysis of Amazon reviews revealed a consistent downward trend for highly-rated books. The average rating typically dropped by approximately 0.4 stars over the first 20 reviews. Statistically, in an independent sequence of random opinions, the deviation from the average should decrease over time. Instead, the authors found that the expected deviation of a new rating increased as the number of reviews grew, proving that later reviewers were intentionally deviating from the established pack.
2. Moderation in Both Directions on IMDB
While the Amazon data primarily highlighted a downward trend in high-rated books, the IMDB data provided a more complete picture. The researchers found that while "good" movies saw their ratings drop over time, "bad" movies saw their ratings rise. This confirms that the trend is not a general shift toward negativity, but a specific shift toward the center (moderation).
3. Effort as a Driver of Contrarianism
The IMDB data provided the strongest evidence for costly expression bias. When users provided low-effort feedback (just clicking a star), they tended to follow the existing consensus. However, when users took the time to write a full comment, their ratings were systematically more extreme in the opposite direction of the majority. Commenters on high-rated movies were more negative than the average, while commenters on low-rated movies were more positive.
Implications for Collective Wisdom
The study suggests that the "collective wisdom" of online communities is often a skewed representation of reality. Because the system incentivizes those with dissenting views to speak up, the resulting average may reflect a motivated minority rather than the silent majority.
For platform designers and consumers, these findings serve as a warning: the perceived quality of a product or the sentiment of a public discourse is heavily influenced by the order and frequency of early views. Understanding "costly expression bias" is essential for accurately interpreting the temporal dynamics of public opinion in the digital age.
Summary updated