Industries / Media, arts and recreation / Media and entertainment / Streaming
Recommendations
Recommendation is the work of choosing which titles, songs and videos to put in front of each viewer or listener, from catalogs too large for anyone to browse.
7
Milestones achieved, from 1992 to 2023.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1992
Filtering by other people
At Xerox PARC, David Goldberg, David Nichols, Brian Oki and Douglas Terry described Tapestry, an email system built on what they called collaborative filtering: "people collaborate to help each other perform filtering by recording their reactions to documents." Ringo, from the MIT Media Lab, applied the idea to music from 1 July 1994. It compared each user's ratings of artists with other users' and suggested what similar people liked, and by September 1994 it had 2,100 users.
LessonIn Tapestry, as Ringo's authors put it, "the user still needs to know which other people may have similar tastes"; Ringo found them automatically.
Credit
Goldberg, Nichols, Oki and Terry (Xerox PARC); Upendra Shardanand and Pattie Maes (MIT Media Lab)
2009
A million-dollar contest
In October 2006 Netflix offered $1 million for a 10% improvement in its predictions of how members would rate films, and released 100 million ratings for contestants to work with. BellKor's Pragmatic Chaos, a team of seven, won on 21 September 2009 against "more than 40,000 teams from 186 countries." Netflix wrote in 2012 that it had put two methods from an earlier stage of the contest into use, but not the winning blend, because "the additional accuracy gains that we measured did not seem to justify the engineering effort needed to bring them into a production environment."
LessonThe best score was not the best system. In April 2017 Netflix replaced the five-star ratings the contest was built on with thumbs up and down, and said ratings activity rose by 200%.
Credit
Netflix; BellKor's Pragmatic Chaos (Robert Bell, Martin Chabbert, Michael Jahrer, Yehuda Koren, Martin Piotte, Andreas Töscher and Chris Volinsky)
2010
Anonymous data, identified
Arvind Narayanan and Vitaly Shmatikov of the University of Texas at Austin showed that the contest's ratings, released without names, could be linked back to people. By their count, with 8 ratings and dates up to 14 days off, 99% of records could be uniquely identified. After a lawsuit and inquiries from the US Federal Trade Commission, Netflix dropped a planned second contest in March 2010. The FTC's closing letter records that Netflix would release such data in future only "to researchers who contractually agree to specific limitations on its use."
LessonA viewing history with the name removed can still identify the person who made it.
Credit
Narayanan and Shmatikov; the FTC
2015
A weekly personal playlist
In July 2015 Spotify began Discover Weekly, which Engadget described as "a two-hour playlist of suggestions based on your listening habits and those of folks with similar preferences," renewed every Monday. Spotify said in June 2025 that listeners had streamed "more than 100 billion tracks" from it, and that each week it leads to "more than 56 million new artist discoveries, with 77% coming from emerging artists."
Credit
Spotify
2016
Ranked by watch time
Paul Covington, Jay Adams and Emre Sargin of Google described YouTube's recommender at the RecSys conference in September 2016. One neural network narrows millions of videos to a few hundred candidates, and a second ranks them by expected watch time rather than clicks, because "ranking by click-through rate often promotes deceptive videos that the user does not complete." The models "learn approximately one billion parameters and are trained on hundreds of billions of examples."
LessonWhat a system is told to maximize decides what it shows.
Credit
Covington, Adams and Sargin (Google)
2019
Fewer borderline videos
In January 2019 YouTube said it would recommend less "borderline content," videos that come close to breaking its rules. In December 2019 it said more than 30 changes had cut watch time of such videos coming from recommendations to non-subscribers in the US by 70% on average. Mozilla's 2021 YouTube Regrets study, drawn from 37,380 volunteers, found that 71% of the videos they reported regretting had been recommended to them. Its 2022 study, from 22,722 volunteers, found that the "Don't recommend channel" button prevented only 43% of unwanted recommendations, and "Not interested" only 11%.
LessonA platform's own count of what it changed needs an outside check of what users see.
blog.youtubeblog.youtubefoundation.mozilla.orgmozillafoundation.org
Credit
YouTube; the Mozilla Foundation
2023
A feed without profiling
Article 38 of the EU's Digital Services Act requires the largest platforms to "provide at least one option for each of their recommender systems which is not based on profiling," that is, not based on a record of each user's behaviour. The European Commission named the first 17 very large online platforms on 25 April 2023, YouTube among them. Its current list puts YouTube at 416.6 million average monthly users in the EU and does not include Spotify or Netflix. In March 2026 the digital rights network EDRi complained to Belgium's regulator that YouTube users who switch profiling off "are left with a blank screen and a deleted YouTube watch history."
LessonA right to a feed without profiling helps only if the feed it gives can be used.
eur-lex.europa.euec.europa.eudigital-strategy.ec.europa.euedri.org
Credit
the European Parliament and Council; the European Commission
Watch for
Steering your own feed
Spotify's Gustav Söderström wrote in December 2025, "For the first time, Spotify is giving users the power to steer the algorithm." The Prompted Playlist feature, which builds a playlist from a typed request and the listener's history, went into a test in New Zealand on 11 December 2025 and reached the US and Canada in January 2026 and Australia, Ireland, Sweden and the UK in February. In March 2026 Spotify began testing Taste Profile, which lets listeners "steer what you see on the Spotify homepage."
Achieved whenAn outside test with test accounts shows that prompts or Taste Profile edits change what the home screen and algorithmic playlists recommend, by a measured amount and for longer than a few days.
newsroom.spotify.comnewsroom.spotify.comnewsroom.spotify.com
Claimed by
Spotify, December 2025 to March 2026
Recommendations from Gemini
In October 2025 Google researchers wrote that "The PLUM framework is in production for YouTube recommendations." PLUM starts from a Gemini language model and adapts it to choose videos. In their live tests it raised click-through on the Shorts home panel by 4.96%, views of Shorts by 0.39% and views of long-form videos by 0.80%.
Achieved whenAn outside party, such as an auditor under the Digital Services Act or a vetted researcher with data access under its Article 40, confirms the system is in use and measures its effect.
Claimed by
Google DeepMind and YouTube researchers, October 2025
How much recommendations matter
In 2015 two Netflix executives wrote that its recommender "influences choice for about 80% of hours streamed at Netflix" and that personalization and recommendations "save us more than $1B per year." In November 2025 Netflix said its new foundation model, one large model trained on members' viewing that feeds many recommendation features, was "now used in production." A paper first posted the same month, five of whose authors are Netflix employees, estimates that switching to two simpler methods "would lead to a 4% and 12% reduction in engagement," far below the 80% figure for influence.
Achieved whenAn independent study, for example one using viewing histories donated by members, measures how much viewing the recommender adds.
Claimed by
Netflix and Netflix researchers, 2015 and November 2025
Open
01
Count AI artists in playlists
A repeated, public count of how many artists in a music service's algorithmic playlists, such as Release Radar and Discover Weekly, make music generated by AI, across genres and months, with the raw lists published.
The briefCredit · unclaimed02
Audit the unprofiled feed
A public tally of what YouTube shows users in the EU with profiling switched off, set against what the same users see by default: the videos on each surface (home, the "up next" list and Shorts), the overlap between the two modes, the steps it takes to switch, and what happens to watch history.
The briefCredit · unclaimed
Open 01 · the brief
Count AI artists in playlists
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A repeated, public count of how many artists in a music service's algorithmic playlists, such as Release Radar and Discover Weekly, make music generated by AI, across genres and months, with the raw lists published.
Why it is still open
The Duke University Tech Policy team ran the only count found, once, in October 2026. Five Premium test accounts over 28 days met 728 artists. Of the artists its scoring flagged, "only 14%" were confirmed as AI-generated by Deezer's labels, and all of those had come through Release Radar. The team asked platforms to "regularly publish … how frequently AI-generated tracks appear in recommendation systems." It also notes that Spotify's terms bar users from running detection software directly. Spotify's Web API has blocked new apps from reading "Algorithmic and Spotify-owned editorial playlists" since 27 November 2024, so the work has to be done by hand.
Built with what exists
Ordinary Premium test accounts, each trained by listening to one genre; screenshots of each account's own playlists; a vision-language model such as Claude Fable 5.2 (see the Note "Three working devices") to turn the screenshots into lists of artist names; Deezer's public AI-content labels, checked by hand; a shared spreadsheet.
First step
Set up two test accounts in different genres, listen normally for two weeks, then list every artist in their Release Radar and Discover Weekly and check each against Deezer's labels.
Done looks like
Six monthly snapshots from at least 10 accounts across five genres, with per-playlist counts and shares of AI-labelled artists, and the raw artist lists published, so that a second team can check any month.
Open 02 · the brief
Audit the unprofiled feed
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public tally of what YouTube shows users in the EU with profiling switched off, set against what the same users see by default: the videos on each surface (home, the "up next" list and Shorts), the overlap between the two modes, the steps it takes to switch, and what happens to watch history.
Why it is still open
A January 2026 study of the first Digital Services Act audits found that auditors "can verify that alternative recommendation modes exist but cannot assess whether these truly eliminate profiling." EDRi's March 2026 complaint describes a blank home screen, and this research found no published outside tally. YouTube's terms bar access "using any automated means" without its written permission, so the work must be done by people with real accounts, or through the Act's data access for vetted researchers.
Built with what exists
Volunteers in at least two EU countries with fresh YouTube accounts; screen recordings and screenshots; a vision-language model such as Claude Fable 5.2 (see the Note "Three working devices") to list the videos in each screenshot; an application for vetted-researcher data access under Article 40 of the Digital Services Act as a later stage.
First step
Make two accounts in one EU country, watch the same 20 videos on each, switch one to the option without profiling, and screenshot both home screens every day for a week.
Done looks like
10 accounts in each mode for four weeks, with published counts per surface, the share of videos the two modes have in common, the number of steps to switch, and whether watch history survives the switch.