TikTok's "Rabbit Hole" Effect: What the Commission's DSA Finding Really Says, and Why It Anticipates the AI Act

On 6 November 2026, the European Commission announced its preliminary finding: TikTok's "For You" recommender system would breach the Digital Services Act (DSA) due to an insufficient risk assessment and inadequate mitigation measures against the "rabbit hole" effect — repeated exposure to content on themes such as sadness, isolation and body image, reinforced by an algorithm that follows engagement more than users' explicit preferences. This is still a finding, not a final penalty. But the way Brussels takes apart the risk assessment of a recommender algorithm is already a useful reference for anyone who will have to manage the same type of risk under the AI Act and ISO 42001.
The preliminary finding: what the Commission is contesting
The official press release (IP/26/2361, DG CONNECT) identifies two distinct shortcomings in how TikTok manages its recommender system: an insufficient risk assessment and inadequate mitigation measures regarding the "rabbit hole" effect. The Commission's text never literally cites "Article 34" or "Article 35" of the DSA — it describes the two shortcomings in plain language, as "risk assessment" and "risk mitigation measures". Attributing them to those two articles (systemic risk assessment and mitigation obligations) is substantively correct and confirmed by multiple independent sources, but it should be treated as an interpretive mapping, not a direct quotation from the Commission. No reference to Article 26 DSA (recommender systems and advertising) appears in the release: it is not the subject of this finding.
The "rabbit hole" effect
The Commission defines the effect as a mechanism whereby the algorithm, by optimising for views and interactions, ends up intensifying exposure to content on sensitive themes — sadness, isolation, body image — regardless of what the user has explicitly chosen to see. This gap between behavioural signal (engagement) and stated preference is precisely at the core of the systemic risk assessment required of VLOPs (Very Large Online Platforms) under the DSA.
Where it sits in the proceeding
This is a preliminary finding, not a final decision. TikTok has the right to access the file and to reply in writing; in parallel, the Commission will consult the European Board for Digital Services. Should the preliminary findings be confirmed, the Commission could adopt a non-compliance decision with a fine of up to 6% of the group's total worldwide annual turnover.
The formal proceeding against TikTok did not start with this finding: it was opened on 19 February 2024 and covers several fronts. The same file includes preliminary findings on addictive design features (February 2026) and minors' protection (July 2026), a preliminary finding on researcher data access (October 2025), and an advertising transparency track closed with binding commitments (December 2025). The "For You" feed contestation is thus the latest piece of a multi-year investigation, not an isolated case.
The common thread with the AI Act and ISO 42001
The DSA and the AI Act remain distinct regulations with different scopes: there is no regulatory overlap between the two. But the method by which the Commission assesses TikTok's risk assessment — risk identification, proportionate mitigation measures, supporting documentation — mirrors the same conceptual framework required by Article 9 of the AI Act (risk management system) and by ISO 42001 controls for AI management systems.
It is important not to overstretch the analogy: AI Act Article 9 applies only to AI systems classified as high-risk, and TikTok's "For You" feed does not necessarily fall into that category — it is not listed in Annex III of the Regulation. The link to the AI Act is therefore methodological, not normative: the Commission is applying to a recommender algorithm the same kind of scrutiny — identified risk, proportionate measures, documentary evidence — that AI Act oversight will apply to high-risk systems, and that ISO 42001 requires to be formalised within a management system.
What it means for algorithm governance, beyond VLOPs
The risk assessment and mitigation obligations under DSA Articles 34-35 apply only to designated VLOPs and VLOSEs — not to every online platform. But anyone managing recommender systems or other high-impact algorithms, even outside that perimeter, can read this preliminary finding as a concrete case study: this is what a European regulator considers an adequate — or inadequate — risk assessment for a system optimising on behavioural signals. It is the same standard that AI Act Article 9 and ISO 42001 are converging toward. It is worth documenting risk identification and mitigation measures on your own algorithmic systems today, before it arrives as a regulator's finding.
Would your algorithmic risk assessment survive this kind of scrutiny?
Identified risk, proportionate measures, documentary evidence: this is the framework the Commission is using to take apart a recommender system, and it is the same one behind AI Act Article 9 and ISO 42001. We can help you document it before a regulator asks.
Let's talk about your algorithm governance →