Recommender Systems
The software that decides what you see next on a feed, video site or shop, ranked by what is most likely to keep you watching rather than by what you asked for. On one big video platform it chooses over 70% of what people watch. When researchers reduced how many like-minded posts thousands of users saw, their political views did not measurably change. It shapes what billions of people see, and so what seems normal.
Maturity
Where it feeds in
Where you have already met this
During the 2020 US election, researchers cut the share of posts that 23,377 Facebook users saw from like-minded sources by about a third. Like-minded sources made up half of what US Facebook users saw. Across eight measures - polarisation, extremity, belief in false claims - nothing measurable moved. The algorithm is not the whole story. It is not no story either.
If this moved
If non-profiled feeds became the default rather than an option, the map would not move: the attention economy sits at 1 on consumption norms, epistemic fragmentation at 2 on epistemic trust, polarisation at 2 on generational turnover. What would move is what people see.
What to watch for
How many EU users choose, and keep, the non-profiled feed the DSA requires. Researcher access to platform data granted by law rather than negotiated.
Not to be confused with
Search. A search engine answers a question you typed; a recommender decides what to show you when you asked for nothing.
What it is
Algorithms that decide what a person sees next, optimised against engagement signals rather than against stated preference.
How it works
- 01Narrow the catalogueFrom hundreds of millions of items, a fast model picks a few thousand that roughly match your history.
- 02Gather the contextThe system adds what you just did - what you watched to the end, what you skipped, where you are, what time it is.
- 03Predict your reactionsA large model estimates, for each item, how likely you are to click, watch, like, share or hide it.
- 04Weigh them into one scoreThe platform decides how much each reaction is worth. Those weights are business choices, not user choices.
- 05Apply the rules and serveSafety filters, duplicates and paid placements reshape the list before it reaches your screen, in a fraction of a second.
State of play
- Meta's apps averaged 3.19 billion daily active people in December 2023 (Meta, Q4 2023 results, 1 Feb 2024).
- YouTube's chief product officer said its recommendation algorithm 'drives over 70 percent of content watched' on the platform (Neal Mohan, CES, Jan 2018, reported by NBC News).
- The Commission designated 17 very large online platforms and 2 very large search engines, giving them four months to comply with the Digital Services Act (European Commission, 25 Apr 2023).
- Under Article 38 of the DSA, those services must offer at least one option for each recommender system that is not based on profiling (Regulation (EU) 2022/2065).
- Switching Facebook and Instagram users to a chronological feed for three months substantially cut their time on the platforms and did not significantly change their polarisation (Guess et al., Science 381, Jul 2023).
- Reducing exposure to like-minded sources by about a third for 23,377 Facebook users had no measurable effect on eight attitude measures (Nyhan et al., Nature 620, 27 Jul 2023).
Why it matters
It is the mechanism by which attention is allocated at population scale. It shapes what is seen, and therefore what is believed to be normal - which is a claim about social reality, not about software.
- It feeds three megatrends heavily - the attention economy, platform-mediated social life, epistemic fragmentation - and holds none. Each is held by something human.
- Mechanism is established; magnitude is disputed. The big field experiments found smaller effects than the public argument assumes.
- It is measured continuously by its operators and barely at all by anyone else. That asymmetry is the diagnostics gap in its cleanest form.
What holds it back
Regulation
The EU requires a non-profiled option and audits systemic risk. It constrains design, not use.
Data access
Independent researchers depend on what platforms release. The best experiments so far ran with the platform's cooperation.
Where liability sits
Rules focus on what platforms host rather than on what their rankers amplify. Moving that line is one of the few changes that would alter the incentive to optimise for engagement.
Terms
- Candidate generation
- The first stage of a recommender, which cuts a huge catalogue down to a few thousand items.
- Ranking model
- The model that scores each candidate by predicted engagement.
- Profiling
- Using personal data to predict a person's interests or behaviour - the thing DSA Article 38 requires an alternative to.
- Chronological feed
- A feed ordered by time rather than by predicted interest.
- Engagement
- Measurable reactions - clicks, watch time, likes, shares - that ranking systems optimise for.
Steepens
| Megatrend | Weight | Role |
|---|---|---|
| The Attention Economy | large | |
| Platform Mediation of Social Life | large | |
| Epistemic Fragmentation | large | |
| Polarisation and Radicalisation | moderate | |
| Time Migration into Digital Environments | moderate |
The gradient behaves like a sticky rubber band: it is pulled hardest toward the weakest enabler. Marginal links are shown but do not pull.
If this moved one step
Scored 5 of 5: there is no step above it.
What if it moved?What would change the gradient
Conditional, not predictive. No dates, no probabilities.
- 01Researcher data access mandated rather than negotiated
- 02Interface defaults - chronological or user-set ranking as an option that persists
- 03Liability attaching to amplification rather than to hosting
Common misconception
That it reflects preference. It optimises a proxy for engagement, and the gap between what people engage with and what they would choose on reflection is where the harm sits. The effect on polarisation is genuinely contested and must be presented as such.
Talin, B. (2026). Recommender Systems. The Future Explorer, dataset v0.4.0 (2026-11-16). https://explorer.talin.digital/e/recommender-systems/