Chapter 3|Due Diligence Obligations - Online Platforms|📖 7 min read
1. Providers of online platforms that use recommender systems shall set out in their terms and conditions, in plain and intelligible language, the main parameters used in their recommender systems, as well as any options for the recipients of the service to modify or influence those main parameters that they may have made available, including at least one option which is not based on profiling as defined in Article 4, point (4), of Regulation (EU) 2016/679.
2. The main parameters referred to in paragraph 1 shall explain why certain information is suggested to the recipient of the service. They shall include, at least:
(a) the criteria which are most significant in determining the information suggested to the recipient of the service, including where applicable how other relevant information such as the way the recipient uses the online platform, engagement levels or geographical location, affects the selection of the information suggested;
(b) the reasons for the relative importance of those parameters.
3. Where several options are available pursuant to paragraph 1, providers of online platforms shall provide an easily accessible functionality on their online interface that allows the recipient to select and to modify at any time their preferred option for each of the recommender systems concerned. That functionality shall be directly accessible from the specific section of the online platform's online interface where the information is being prioritised.
Understanding This Article
Article 27 addresses THE defining feature of modern platforms: recommendation algorithms that determine what content users see. TikTok's For You page, YouTube's recommended videos, Facebook's News Feed, Twitter's algorithmic timeline - these systems shape user experience and information access. Article 27 makes these 'black box' algorithms at least somewhat transparent and gives users control over them.
Paragraph 1 requires disclosure of 'main parameters' - the key factors that determine content prioritization. Platforms need not reveal full algorithmic details (protecting trade secrets and preventing gaming), but must explain the primary factors: engagement signals (likes, shares, watch time), recency, relevance to user interests, content type, source credibility, etc. This explanation must be in 'plain and intelligible language' - no technical jargon, accessible to average users.
Critically, paragraph 1 requires 'at least one option which is not based on profiling' - a non-personalized feed. If TikTok shows algorithmically-personalized For You feed by default, it must offer chronological Following feed. If Twitter defaults to algorithmic timeline, it must offer chronological option. This non-profiling alternative prevents platforms from forcing personalized curation on unwilling users.
Paragraph 2 elaborates on 'main parameters' - not just listing factors, but explaining their significance. If YouTube recommends videos, it must explain: 'We prioritize videos similar to those you previously watched (40% weight), videos from channels you're subscribed to (30%), popular videos in your region (20%), and recent uploads from trending topics (10%).' The percentages/weights show RELATIVE importance, enabling users to understand what most influences their recommendations.
Paragraph 2(a) requires explaining how user behavior affects recommendations: 'Your watch history, search queries, liked videos, engagement patterns (watch time, completion rates), and location all influence recommended videos.' This transparency helps users understand how their actions shape their algorithmic experience.
Paragraph 3 is THE user control provision. Selection functionality must be 'easily accessible' 'directly' from where recommendations appear. If YouTube's homepage shows recommended videos, there should be visible toggle/menu right there ('Switch to: Subscriptions chronological' / 'Trending' / 'Algorithmic recommendations'). Users shouldn't need to dig through settings - control must be immediate and contextual.
The requirement to modify preferences 'at any time' means users can switch between recommendation modes freely. Today algorithmic, tomorrow chronological, back to algorithmic - users control their experience dynamically based on current needs and preferences.
Key Points
Platforms using recommendation algorithms must disclose main parameters in terms
Must explain in plain language how recommendations work and why content is shown
Must disclose criteria determining suggested content and their relative importance
Must provide at least one non-personalized recommendation option (no profiling)
Users must be able to select and modify recommendation preferences anytime
Selection functionality must be easily accessible where content prioritization occurs
Empowers user agency over algorithmic content curation
Mandatory algorithmic transparency for platforms using recommendation systems
Practical Application
For TikTok For You Page: TikTok's terms must explain: 'For You recommendations are based on: (1) Videos you've liked, shared, or watched completely (50% importance); (2) Video information including captions, sounds, hashtags (25%); (3) Device and account settings like language preference and location (15%); (4) Accounts you interact with (10%). We also offer Following feed showing only accounts you follow in chronological order without personalization.' The app must have easily accessible toggle: [For You] / [Following] directly on main feed interface.
For YouTube Homepage: YouTube must disclose: 'Recommended videos prioritize: (1) Similar videos to your watch history (40%); (2) Subscription feed uploads (30%); (3) Trending/popular videos in your country (20%); (4) Topics you've searched (10%). Your engagement (watch time, likes, comments) heavily influences recommendations.' YouTube must offer: [Home] (algorithmic) / [Subscriptions] (chronological) / [Trending] (non-personalized) options, switchable with one click from homepage.
For Instagram Feed: Instagram must explain: 'Your feed shows posts from accounts you follow, prioritized by: (1) How often you interact with the account (likes, comments, DMs) - 45% weight; (2) Recency of posts - 25%; (3) Post engagement from others - 20%; (4) Post format matching your preferences (photos vs videos vs reels) - 10%.' Must provide: [Algorithmic Feed] / [Chronological Following] toggle accessible directly on main feed, enabling instant switching.
For Twitter/X Timeline: Twitter must disclose: 'Your timeline mixes tweets from followed accounts with recommended content based on: (1) Accounts you interact with most (35%); (2) Tweets similar to those you've liked (30%); (3) Popular tweets from accounts you don't follow (20%); (4) Trending topics relevant to your interests (15%). You can switch to chronological feed showing only followed accounts in order posted.' Provide: [For You] / [Following] tabs prominently displayed.
For Facebook News Feed: Facebook must explain: 'Your News Feed prioritizes: (1) Posts from friends/pages you frequently interact with - 40% weight; (2) Post type matching your typical engagement (videos vs photos vs links) - 25%; (3) Recency - 20%; (4) Post popularity among your network - 15%. We also offer Most Recent feed showing posts chronologically without ranking.' Users can toggle between [News Feed] and [Most Recent] views.
For LinkedIn Feed: LinkedIn must disclose: 'Your feed balances: (1) Posts from connections you engage with (40%); (2) Posts about topics you follow or engage with (30%); (3) Sponsored content (15%); (4) Posts from second-degree connections (15%). We prioritize content your connections have engaged with and content LinkedIn deems professionally relevant to your profile.' Must offer chronological option from connections only.
For Spotify Recommendations: Spotify must explain: 'Discover Weekly playlist is generated based on: (1) Songs in playlists you've created or saved (40%); (2) Songs you've played repeatedly (30%); (3) Songs similar to your listening history by audio features (20%); (4) Songs other users with similar taste enjoy (10%).' While Discover Weekly is inherently algorithmic, Spotify must ensure users can browse music without algorithmic filtering (search, browse by genre/artist, etc.).
For Reddit Homepage: Reddit must disclose: 'Your homepage shows: (1) Posts from subreddits you're subscribed to (60%); (2) Upvotes and engagement signals (30%); (3) Recency (10%). Hot/Rising/Top sorting further prioritizes by different engagement metrics. You can switch to subscription chronological feed.' Users can select: [Best] (algorithmic) / [Hot] (trending) / [New] (chronological) / [Top] sorting options.
For Netflix Recommendations: Netflix must explain: 'We recommend titles based on: (1) Shows/movies you've watched and rated (50%); (2) Viewing patterns (genres, actors, completion rates) - 30%; (3) Popular content among users with similar tastes (15%); (4) New releases in your preferred categories (5%). You can browse all content categorically without personalization through Browse menu.' Must enable non-algorithmic browsing alongside recommendations.
For Non-Compliant Platforms: If TikTok offers only algorithmic For You feed with no chronological option, violates Article 27(1) requirement for non-profiling alternative. If YouTube explains algorithms vaguely ('We show videos you might like based on various factors') without detailing specific parameters and weights, violates Article 27(2) transparency requirements. If Instagram buries feed preference toggle deep in settings rather than making it directly accessible from feed interface, violates Article 27(3) accessibility requirement.
For User Benefits: Algorithmic transparency enables informed decisions: If users realize TikTok heavily weights watch time, they understand why they see long videos. If they know Instagram prioritizes recent interactions, they can decide whether to switch to chronological to see friends they haven't interacted with recently. Transparency transforms algorithm from mysterious force to understood tool users can manage.
For Platform Benefits: Clear algorithm explanations can build trust - users understand platforms aren't arbitrarily censoring content but applying stated criteria. Non-personalized options reduce filter bubble concerns - users can escape algorithmic curation when desired. This transparency and control may reduce user alienation and regulatory scrutiny.