Do not recommend? Reduction as a form of content moderation

Gillespie, T. (2022). Do not recommend? Reduction as a form of content moderation. Social Media + Society, 8, 205630512211175. https://doi.org/10.1177/20563051221117552

View paper

Summary

Gillespie argues that debate about content moderation has been overly fixated on removal — deleting posts and banning users — while ignoring reduction, the now-pervasive practice of demoting the visibility or reach of problematic content without taking it down. Platforms use machine-learning classifiers to identify content deemed misleading, risky, or offensive enough to warrant demotion in rankings, recommendations, and search results, yet not removal. The essay documents how eight major platforms implement and justify these techniques, offers a typology of reduction, and insists that reduction be recognized as a core form of platform governance. Far from a gentler alternative to removal, reduction avoids none of removal’s problems (arbiter power, bias, inequitable impact) while concentrating even more curatorial power in private, profit-oriented intermediaries by keeping that power invisible to transparency and accountability regimes.

Key Contributions

• Names and defines reduction as a distinct, under-theorized category of content moderation, deliberately chosen over “borderline content,” “shadowbanning,” or “suppression.”
• Offers a concrete typology grounded in how recommender systems work: do not recommend at all, do not recommend as much, do not recommend to some users — implemented at the inventory/candidate-generation or ranking stages.
• Documents and compares reduction policies and justifications across eight major platforms, surfacing practices firms have publicly obscured.
• Reframes content moderation and algorithmic recommendation as a single governance problem rather than two separate platform functions.
• Identifies that reduction is effectively invisible to transparency/accountability mechanisms and may be inherently unmeasurable, with implications for regulation and law.
• Provides a reflexive methodological account of “studying up” in secretive Silicon Valley firms under NDA regimes.

Methods

Qualitative analysis of company policies, corporate communications (blog posts, transparency documents), and technical literature from academia and industry. Gillespie closely reads platform announcements and justifications — YouTube’s 2019 “borderline content” post and “Four Rs” framework, Facebook/Instagram policies, and statements from Twitter, LinkedIn, TikTok, Tumblr, and Reddit. He supplements this with informal background conversations with a dozen-plus Trust & Safety and adjacent employees across five platforms, treated as background (not formal interviews) to respect NDAs; confidential details were cross-verified against published policy before inclusion. A reflexive discussion of access, NDAs, IRB, and insider positionality accompanies the analysis.

Findings

• YouTube announced borderline-content reduction in January 2019, framed as improving “recommendation quality” rather than as Trust & Safety, later claiming a 70% average drop in watch time from non-subscribed recommendations — a claim nearly impossible to independently verify.
• Facebook/Instagram announced reduction policies in May 2018 (“reduce its spread in News Feed using ranking”), later publishing detailed demotion guidelines covering borderline content, harmful misinformation, and low-quality junk.
• Twitter, LinkedIn, TikTok, Tumblr (hashtag/search blocking), and Reddit (“quarantine”) all use family-resemblance strategies, mostly less transparent than their removal policies; Reddit is comparatively explicit.
• Platforms justify not-removing via user preferences, Zuckerberg’s “natural engagement pattern” theory (engagement rises near any line, naturalizing away responsibility), and a consent-based lower standard for recommended vs. self-chosen content.
• Unstated reasons include lower political risk (dodging “censorship” charges), continued financial benefit from demand for such content, and flexibility to police fast-evolving harms while evading accountability.
• Technically, reduction inserts negative signals into recommenders at inventory or ranking stages; Facebook appears to reuse existing removal classifiers by demoting borderline scores, while YouTube built a separate human-rated classifier using Google’s search-quality evaluator guidelines.

Connections

This essay extends Gillespie’s own foundational work on platform gatekeeping and the politics of algorithmic visibility Gillespie2010-as, reframing the debate that empirical work on algorithmic curation and news exposure has largely operationalized around ranking and recommendation Bakshy2015-rn, Gonzalez-Bailon2023-uy. Its core insight — that demotion rather than deletion is the frontier of moderation — speaks directly to the register’s concerns with political-content visibility and to critical anniversary reflections on platform power Baym2026-tr, while its typology of reduction offers a conceptual scaffold for empirical shadowbanning and reach-measurement studies elsewhere in this topic.