You watched one video about a topic. The next ten appeared automatically. You did not search for them.
Literal meaning: Recommender systems are algorithms that determine which content, products, or connections to surface to users — filtering the vast supply of available content down to a personalised feed optimised for a platform-defined objective, typically engagement, dwell time, or conversion.
Origin: Collaborative filtering — the foundational technique — was developed in the early 1990s at Xerox PARC. Amazon popularised recommendation at consumer scale with “customers who bought this also bought” from 1998. Netflix’s $1 million algorithm prize (2006–2009) accelerated research. YouTube’s recommendation system became the subject of radicalisation research from 2018, when Zeynep Tufekci and later researchers documented pathways from mainstream to extreme content through recommendation chains.
Algorithms that decide what you see next — optimised not for what is accurate or good for you, but for what keeps you on the platform longest.
The Appeal: Recommender systems solve a real problem: infinite content requires filtering. A good recommendation genuinely serves the user — surfacing relevant music, useful products, or interesting articles they would not have found otherwise. The engineering challenge is substantial and the success cases are real.
The Friction: The objective function is the problem. Recommender systems optimise for what they are instructed to optimise for — and platforms typically optimise for engagement, not for accuracy, balance, or user wellbeing. Filter Bubble — algorithmic enclosure of similar viewpoints — is one documented consequence. Radicalisation pathways are another: Ribeiro et al. (2020) document how users migrate from mainstream YouTube content through the Alt-lite to Alt-right channels through recommendation chains — the algorithm recommends more extreme content because extreme content generates more engagement. Bail et al. (2023) demonstrate a counterintuitive corollary: exposure to opposing views through recommendations does not reduce polarisation but increases it, because the emotional response to disagreement is itself an engagement signal. Anger and fear produce the strongest engagement signals of any emotional state — Brady et al. (2017) documented that moral-emotional language spreads furthest on social platforms, which means Ragebaiting is not an abuse of the system but its logical output. Virtual Influencer shows a different dimension of the same logic: by repeatedly surfacing a fabricated persona, the algorithm builds the familiarity that makes a parasocial bond feel earned — with a figure that does not exist. SMV discourse is amplified the same way: manosphere and looksmaxxing content generates the same anger/status engagement signals that drive ragebaiting and radicalisation, so the recommender surfaces more of it, not less. Microtargeting extends recommendation into advertising: the same optimisation logic applied to persuasion. The EU’s Digital Services Act (2022) introduced obligations for major platforms to provide non-algorithmic feed options and to disclose recommendation parameters. Van Iperen (2026) captures the full scope of this logic: feed, notification, and recommendation algorithm function as a virtual fence — not merely steering users through available content, but determining how much they consume and defining which territories exist for them at all. The system directs; the user moves. The pastures never reached do not exist for them.
Why This Matters: Recommender systems are the infrastructure of the contemporary information environment. Understanding that your feed is an optimisation output — not a neutral window on the world — is the first step in reading it differently.
Related terms: Filter Bubble · Echo Chamber · Microtargeting · Dopamine Feedback Loops · Attention Harvesting · BUMMER · Surveillance Capitalism · Virtual Influencer · Ragebaiting · SMV (Sexual Market Value)
Read more:
Primary:
- Ribeiro et al. (2020) — Auditing Radicalization Pathways on YouTube. FAccT
- Bail et al. (2023) — Exposure to opposing views on social media can increase political polarization. PNAS
Secondary:
- Tufekci (2019) · Tufekci (2018) · Ledwich & Zaitsev (2019) · Brady et al. (2017) · Kang et al. (2025) · Hope, Fear, or Anger? (2023) · Van Iperen (2026) — journalism, counter-evidence, emotion research and theoretical frameworks