Recommender system parameter disclosure and opt-out

Description

Recommender-transparency regimes are unusual among AI obligations because they assume the technology already works and focus instead on user agency. DSA Article 27 does not ask whether a ranking system is well-built; it asks whether the user is told what the main parameters are and given a non-personalized alternative. China's algorithmic-recommendation provisions, Colorado's consumer-notice rules, and California SB 942's ranking-disclosure overlay all converge on the same posture: the platform decides how to rank, but the user gets to see why and to switch the ranking off. The compliance work is operational plumbing rather than algorithmic redesign. The regime decomposes into three loosely-coupled surfaces. The disclosure surface is a help-center or settings page that names the main parameters of each recommender system in plain language: what signals it weighs, what the dominant signals are, how user behavior changes the output, and how the user can adjust the inputs. The DSA expects this in the terms of service; China's regime expects it within the platform; Colorado expects it as part of consumer notice for consequential-decision systems. The alternative-mode surface is the non-personalized ranking that lets the user opt out of profiling-based ranking entirely; very large online platforms under the DSA must offer this and the alternative must be reachable from each recommender's UI rather than buried in settings. The opt-out surface is the per-user toggle that disables personalization within a category (feed, search, suggestions, advertising) and persists across sessions; the implementation work is usually larger than it looks because most product surfaces have implicit personalization that has to be unwired piece by piece rather than gated by a single flag. The thresholds and timing matter. DSA Article 27 has applied to all platforms since 2024-02-17, with Article 38's non-profiling-alternative duty restricted to designated very large online platforms and very large online search engines (the 45-million-monthly-EU-user threshold). China's algorithmic-recommendation provisions have been operative since 2022-03-01 and apply to providers with public-opinion or social-mobilization properties; the threshold is qualitative rather than user-count. Colorado AI Act consumer-notice obligations apply to deployers of high-risk AI systems making consequential decisions and become operative on 2026-02-01. California SB 942's covered-provider threshold is one million California users and operative date is 2026-01-01. The four regimes do not share a single trigger; an operator above the DSA VLOP threshold is almost certainly above the others, but the converse is not true, and several platforms will land in scope for one regime years before the others reach them. The operationally interesting piece is the gap between disclosure and intelligibility. The DSA requires the main parameters of the ranking system to be set out in plain and intelligible language in the terms of service; the standard most operators ship initially is a list of signals ("engagement, recency, relevance") that satisfies the literal text and tells the user almost nothing. Enforcement signals from the European Commission's early investigations have pushed toward something closer to ranking explanations the user can actually act on, which is harder to write than the engineering team's instinct on what the model does. The non-personalized mode raises the parallel question: alternatives that are technically present but materially unusable (degraded feed quality, missing content, broken navigation) are likely to be treated by regulators as not-really-an-alternative, which is the unsettled enforcement question on this Control.

Applicability

Applies when: features include ai-recommendations or algo-feeds.

How predicates are evaluated

Required by (6 regulations)

• DSA

  Article 27 requires online platforms to set out in their terms and conditions, in plain and intelligible language, the main parameters used in their recommender systems and any options for the user to modify them; Article 38 requires very large online platforms and search engines to offer at least one recommender option not based on profiling.

  Regulation (EU) 2022/2065 of the European Parliament and of the Council (Digital Services Act); Article 27 in application since 2024-02-17

• Algorithm Provisions

  Articles 16-17 require providers of algorithmic-recommendation services to inform users of the operating mechanism of the service in conspicuous terms and to offer a convenient option to switch off algorithmic recommendation or to select non-tailored content.

  Provisions on the Management of Algorithmic Recommendations in Internet Information Services (CAC, MIIT, MPS, SAMR; effective 2022-03-01)

• Colorado AI Act

  SB 26-189 narrowed scope to ADMT consumer-facing disclosure and adverse-decision notices in seven covered consequential-decision domains. Recommendation systems that do not materially influence consequential decisions in covered domains are out of scope. AG rulemaking will define the operational mechanics.

  Colorado AI Act (SB 26-189, repealing and replacing SB 24-205); effective 2027-01-01; codification sections pending Governor signature and AG rulemaking

  Source →

• California SB 942

  The California AI Transparency Act's ranking-disclosure overlay (read together with adjacent consumer-protection provisions) requires covered generative AI providers to disclose how AI-generated content is surfaced and ranked alongside organic content within California-facing surfaces.

  California AI Transparency Act, SB 942 (2024); operative 2026-01-01

  Source →

• California SB 976

  Governs the recommendation engine that determines whether a feed is an addictive feed for a minor under SB 976.

  California SB 976 (2024), Protecting Our Kids from Social Media Addiction Act

  Source →

• EU AI Act

  EU AI Act Article 50 transparency obligations extend to recommendation and content-personalization systems interacting with users.

  Regulation (EU) 2024/1689 of the European Parliament and of the Council

Evidence formats

• Ranking-parameter help-center page (per recommender system)
• Non-personalized alternative UI surface (reachable from each recommender)
• Per-user personalization opt-out flow + persistence test
• Recommender-system Terms of Service section naming main parameters
• Consumer-notice copy for consequential-decision recommenders (Colorado AI Act)
• Internal mapping of implicit-personalization surfaces requiring opt-out wiring