Recommender system transparency

On this page 8

Recommender system transparency is the set of legal obligations requiring online platforms to disclose the main parameters that decide which content, products or accounts users see, and to let users choose or modify how their feeds are ranked. In platform regulation, it addresses the fact that algorithmic curation shapes attention, purchasing and public debate, yet usually operates as an invisible black box controlled by the platform alone.

The most developed rules come from the EU Digital Services Act (DSA), which sets a baseline duty for all online platforms and stricter duties for very large services. Related ranking disclosure rules exist in EU consumer and business-to-platform law, and other jurisdictions, notably China, regulate algorithmic recommendation directly.

What the DSA requires from recommender systems

The DSA defines a recommender system as a fully or partially automated system that a platform uses to suggest or prioritize information in its interface. That covers news feeds, video autoplay queues, search result ordering on social platforms and “for you” pages, but not paid advertising, which has its own transparency rules and an ad repository.

Article 27 applies to every online platform. Its terms and conditions must set out, in plain and intelligible language, the main parameters used in recommendations and any options to change them. Article 38 adds that very large online platforms and search engines (VLOPs and VLOSEs, with at least 45 million monthly EU users) must offer at least one option not based on profiling, as defined in the GDPR.

ProvisionWho it applies toCore obligation
Article 27All online platformsDisclose main parameters and user options in terms
Article 27(3)Platforms offering several optionsDirect, easily accessible control to switch options
Article 38VLOPs and VLOSEsAt least one option not based on profiling
Articles 34 and 35VLOPs and VLOSEsAssess and mitigate systemic risks from recommenders
Article 40VLOPs and VLOSEsData access for vetted researchers and regulators
Recommender duties by type of service under the DSA

What counts as a main parameter

A main parameter is a criterion that is most significant in determining what is suggested, together with the reasons for its relative importance. Disclosure must be meaningful to an ordinary user: listing every model feature is neither required nor helpful, but vague statements such as “we show content you might like” do not meet the standard.

In practice, platforms describe the signal categories their ranking models weigh. The table below shows typical parameters and the kind of explanation regulators expect.

ParameterWhat it measuresWhat users should learn
Engagement historyPast likes, watch time, sharesThat prior behavior strongly drives the feed
Social graphAccounts followed or interacted withHow much followed accounts outweigh others
Content featuresTopic, language, format, recencyWhy fresh or certain formats are favored
Inferred interestsProfiling from activity and device dataThat inferences exist and can be reset
Integrity signalsPolicy violations, low-quality labelsThat some content is demoted, not removed
Location and contextRegion, time, session behaviorHow local or trending items are boosted
Typical ranking parameters and what disclosure should explain

User control and the non-profiled option

Transparency without choice changes little, so the DSA pairs disclosure with user control. Where a platform offers alternative orderings, users must be able to select and change their preference at any time through a functionality that is directly accessible from the part of the interface where recommendations appear.

The non-profiled option required of very large platforms is typically a chronological feed, a feed of followed accounts only, or ranking by popularity within a region. The option must be genuinely usable: the European Commission has scrutinized designs that bury the toggle in settings, revert it after each session or degrade the experience so much that users give up, since such patterns can also breach the DSA ban on deceptive interface design in Article 25.

How it relates to other platform rules

Recommender system transparency overlaps with several neighboring obligations, and confusing them is a common source of compliance gaps. Each regime targets a different audience and a different kind of ranking.

RuleAudience protectedFocus
DSA recommender dutiesPlatform usersFeed parameters and user choice
P2B Regulation ranking rulesBusiness users of platformsMain parameters of marketplace and search ranking
EU consumer law ranking rulesConsumers searching for productsRanking parameters and paid placement disclosure
DMA self-preferencing banCompeting businessesFair ranking of gatekeeper services
DSA advertising transparencyUsers and researchersWhy a specific ad is shown, ad repository
Recommender transparency compared with related ranking rules

Outside the EU, China’s algorithmic recommendation provisions, in force since 2022, require providers to publicize basic principles, offer an option to switch off personalized recommendations and file algorithms with the regulator. The UK Online Safety Act approaches feeds mainly through child safety risk assessments rather than a general disclosure duty.

Systemic risk and independent scrutiny

For very large services, recommender systems are treated as a potential source of systemic risk, for example amplification of illegal content, harms to minors, electoral manipulation or effects on mental health. Platforms must assess how design choices in ranking contribute to such risks and adjust them where needed, which can mean changing the algorithm itself, not only the explanation of it.

These assessments are reviewed in annual independent audits, and vetted researchers can request data under Article 40 to test claims empirically. Through this layered model, transparency extends from user-facing summaries to expert-level verification.

Limitations and frequent compliance failures

Recommender system transparency has inherent limits. Modern ranking relies on deep learning models with thousands of signals, so even engineers cannot always explain a single recommendation, and simplified disclosures inevitably leave out detail. Platforms also argue that full disclosure would help spammers game the ranking, a tension regulators accept to a point.

FailureWhy it is a problemBetter practice
Generic wording in termsDoes not identify main parametersName signal categories and their weight
Hidden feed toggleControl is not easily accessiblePlace the switch on the feed itself
Option resets each sessionUndermines user choicePersist the selection until changed
Profiled signals in the alternative feedFails the Article 38 requirementExclude personal inferences entirely
Disclosure never updatedOut of date after model changesReview text with each major ranking change
Common failures and how to address them

The idea in brief

Every time you open a social app, a computer program decides which posts you see first. Recommender system transparency means the platform must tell you, in simple words, what that program pays attention to. Examples are what you liked before or who you follow.

On the biggest platforms in Europe, you must also get a way to turn off the personal tailoring. One example is a feed that simply shows the newest posts first. Think of a shop. It must explain why certain products sit at eye level. It must also let you browse the shelves in plain order.

A practical scenario

A short-video platform designated as a very large online platform reviews its feed before a Digital Services Coordinator inquiry. Its legal and product teams find that the terms describe ranking as “based on your interests” and that the chronological option sits four menus deep in settings and resets after every app update.

The platform rewrites its disclosure to name five main parameters, including watch time, follows, content recency, language and integrity demotions, and explains that watch time carries the greatest weight. It adds a feed switcher at the top of the home screen offering a following-only chronological view that uses no inferred interests, and it stores the choice until the user changes it.

In the next independent audit, the auditor confirms that the Article 27 and Article 38 obligations are met, and internal data shows that about 8 percent of users use the non-profiled feed at least weekly, evidence that the option is genuinely accessible.