The Rabbit Hole Is Five Claims, Not One: Why Audits, Browsing Panels, Escalation Studies, Exposure Experiments and Discourse Maps of Online Conspiracy Content Keep Disagreeing

Abstract

Life of Research

The claim that online platforms draw people down conspiracy "rabbit holes" is among the most repeated in public debate about recommender systems, and the published evidence on it appears to contradict itself. Sock-puppet audits find recommendations that narrow and drift toward problematic channels; browsing panels and counterfactual-bot experiments find that consumption tracks prior preference and that the recommender, if anything, moderates it; a 2025 study of Singapore Telegram groups reports that conspiratorial talk is woven into ordinary conversation rather than sealed in enclaves. In this analysis the author reviews and synthesises published results to show that the metaphor bundles five separable claims: that a recommender followed blindly drifts toward such content (drift), that real consumption of it is caused by recommendation (uptake), that individuals escalate over time (escalation), that exposure changes belief (effect), and that the content lives in isolated spaces (location). Each design in the literature identifies at most two of the five, and most of the apparent conflict comes from reading a result about one link as evidence about another, or across YouTube's 2019 recommender change. The paper contributes a link-by-design map, a consolidated table of more than thirty published results indexed by link and period, a link-attribution procedure for reading any new study, and a re-reading of the Telegram study's own tables, which show conspiratorial share ranging from 5 to 70 percent across its clusters in a corpus 95 percent drawn from one group. The evidence is strong that post-2019 YouTube drift is weak and that exposure is concentrated in a small, motivated tail; it is weak on long-horizon belief effects, on platforms other than YouTube, and on location claims without an unselected comparison set.

About the author

Pranay Mahendrakar is an ai specialist and large language model engineer based in Bengaluru, India. He builds production artificial intelligence systems and publishes open-access research on how those systems fail. See all 111 papers by Pranay Mahendrakar, or his ORCID record.