DSA: Recommender System Transparency
Recommender System Transparency [Art 27, 38]
Rule: Online platforms using recommender systems must explain the main parameters used for recommendations and offer users options to modify or influence them. VLOPs must offer a non-profiling option.
What is a Recommender System? [Art 3(sa)]
A recommender system means:
Any fully or partially automated system used by an online platform to suggest in its online interface specific information to recipients of the service or prioritise that information, including as a result of a search, or as a result of determining the relative order or prominence of information.
Examples:
- Social media feeds (Facebook, Instagram, TikTok)
- Search result rankings
- Product recommendations (Amazon, Netflix)
- “For You” pages
- Related content suggestions
- Trending/popular content sections
Basic Transparency Requirements [Art 27]
All online platforms using recommender systems must set out in their terms and conditions:
| Required Information | Description |
|---|---|
| Main parameters | The main parameters used in the recommender system |
| Modifiability | Any options for users to modify or influence those parameters |
What Are “Main Parameters”?
Parameters that significantly affect what content users see:
| Parameter Type | Examples |
|---|---|
| Engagement signals | Likes, shares, comments, watch time |
| User profile | Interests, demographics, history |
| Content signals | Recency, popularity, format |
| Social signals | Friends’ activity, network |
| Commercial signals | Paid promotion, advertising |
Plain Language Requirement [Art 27(1)]
Information must be presented in:
- Plain and intelligible language
- Clearly visible in terms and conditions
User Options [Art 27(3)]
Platforms must provide easily accessible functionality for users to:
- Select and modify preferences
- Influence information presented
- At any time, adjust these preferences
VLOP Enhanced Obligations [Art 38]
VLOPs and VLOSEs have additional requirements:
Non-Profiling Option [Art 38(1)]
Must provide at least one option for each recommender system that is:
Not based on profiling as defined in Article 4(4) of Regulation (EU) 2016/679 (GDPR)
This means offering a recommendation option that doesn’t use:
- Personal data inference
- Behavior analysis
- Preference prediction
Examples of compliant options:
- Chronological feed
- Random selection
- Popularity-based only (no personalization)
Accessibility [Art 38(2)]
Non-profiling option must be:
- Directly and easily accessible from the recommender system interface
- Available where recommendations are made
Practical Implementation
For all online platforms:
- Document all main parameters in T&Cs
- Explain in plain language how recommendations work
- Provide user preference controls
- Allow modification at any time
For VLOPs additionally:
- Offer at least one non-profiling option (e.g., chronological)
- Make non-profiling option easy to access
- Don’t bury it in settings — surface in the interface
Examples of Compliance
| Platform Type | Main Parameters to Disclose | Non-Profiling Option |
|---|---|---|
| Social media | Engagement, recency, network, interests | Chronological feed |
| E-commerce | Purchase history, browsing, popularity | Bestsellers/newest only |
| Video streaming | Watch history, completion rate, genre | New releases/trending |
| News aggregator | Reading history, sources, topics | Latest news only |