5 Redefining amplification

This chapter reframes algorithmic amplification as a problem of measurement, power, and interpretation rather than a single fixed quantity. Using analogies like Flatland and the blind men and the elephant, it argues that amplification is a complex phenomenon that can only be understood through multiple perspectives and metrics. The chapter stresses that defining what to measure is itself an act of power, since controlling the metric also shapes the story platforms can tell about their impact.

The text explains why amplification should still be measured despite uncertainty: it helps researchers understand recommender systems, compare different kinds of users and platforms, and detect unusual changes over time. Rather than seeking one perfect number, the goal is to monitor trends and spot anomalies, such as sudden shifts in exposure during an election or other sensitive events. In this sense, amplification becomes a practical tool for oversight and investigation, especially when platform behavior may be distorted by manipulation or hidden incentives.

To make the concept usable, the chapter introduces several operational definitions from both the producer and reader sides. On the producer side, it discusses measures such as A/B test-based ratios, normalized impressions, and the out-of-network ratio to see how much algorithmic distribution boosts content beyond direct followers. On the reader side, it proposes feed analysis to check whether what users see reflects what they follow or whether the system is overemphasizing certain voices. Across all of these, the chapter emphasizes multi-stakeholder input, privacy safeguards, and the need for real-time monitoring so amplification can support transparency without becoming a surveillance tool.

When a user U produces some content C, that content can be delivered to others in two different ways. First, it will be added to the pool of in-network content for users who follow U directly. At the same time, it might be algorithmically flagged as content with high relevance (or high probability of engagement) for users who don’t follow U directly. In that case, it will be inserted in the out-of-network bucket and be ranked accordingly. In neither case is there a guarantee that other users will see C, as it will compete for a spot with all the other content.

The distribution of U’s followers. Let’s assume that an A/B test is being run, in which A is the baseline—receiving 99% of traffic—and B is the new model, receiving only 1%. If U has enough followers, we can assume that 1% of them are in bucket B, and 99% are in A. This also implies that 99% of impressions are generated by users in bucket A and only 1% from B. Any deviation from this 1/99 ratio can be attributed to algorithmic amplification.

• Algorithmic amplification can be used as a tool to measure the extent to which a recommender system deviates from its baseline. In that respect, it can be used as a tool to detect anomalies (including attempts at manipulation) for online platforms.
• Algorithmic amplification lacks a single, definitive definition. Each definition is a new measuring tool that helps to better understand how the recommender system works. It’s important to strike a balance between usefulness and information.
• Multiple stakeholders, especially those from historically marginalized communities, should be involved in defining amplifications.
• A piece of content can be seen by either people who follow its producer directly or by others if the content is deemed interesting enough to be broadcast. The latter is due to a higher intervention by the algorithm, and it should be taken into account when measuring amplification.
• Amplification can (and should) be defined both on the author's and on the reader’s side, as they highlight different concerns.
• From the author's side, amplification can be measured via an A/B test, using normalized impressions, or using the Out-of-Network ratio of impressions.
• From the reader's side, amplification can highlight whether the recommendations in the feed reflect the distribution of possible content.

FAQ