Automated Detection and Automated Decision-Making Under the EU Digital Services Act: Evidence from 43.5 Billion Statements of Reasons (2023–2026)
This study analyses 43.46 billion published Statements of Reasons across 38 monthly releases of the EU DSA Transparency Database (September 2023 to October 2026; 278 platforms). While 93.86% report automated detection, only 54.50% report fully automated decisions, 9.60% partially automated decisions and 35.90% non-automated decisions (a 39.35 percentage-point gap). Google Shopping alone accounts for 76.95% of all records; across the other 277 platforms (10.02 billion records), a 34.65 percentage-point gap persists, shifting predominantly into partially automated decisions (41.66%).
Two distinct automation stages and a dominant concentration effect
Across 43.46B DSA filings, automated detection is near-universal (93.86%), whereas fully automated decision-making is far lower (54.50%)—and heavily shaped by Google Shopping (76.95% of all records).
Abstract
This research note examines the empirical relationship between reported automated detection and reported automated decision-making across the first 38 monthly aggregate releases of the European Union Digital Services Act (DSA) Transparency Database (25 September 2023 to 31 October 2026; 43,463,346,962 published Statements of Reasons across 278 named platforms). Across the cumulative aggregate, 93.86% of statements report automated detection, whereas 54.50% report fully automated decision-making, 9.60% partially automated decision-making and 35.90% non-automated decision-making—a 39.35 percentage-point gap between automated intake flagging and full decision automation. In the joint 2x3 contingency table, 34.50% of automatically detected statements (14.07 billion records) are recorded as non-automated at the decision stage. Provider-level decomposition shows that unweighted EU totals are heavily shaped by extreme concentration: Google Shopping accounts for 76.95% of all published statements and 94.96% of all non-automated decision records. Across the remaining 277 platforms (10.02 billion statements), a 34.65 percentage-point gap persists between automated detection (85.14%) and fully automated decisions (50.49%), while partially automated decisions account for 41.66% and non-automated decisions for 7.85%. All results are descriptive accounting identities over provider-submitted administrative records; statutory 'automated means' does not establish artificial-intelligence use, review quality or DSA compliance.
Original abstract of the English-language paper.This issue remains frozen
The live monitor ingests new monthly DSA releases as they publish. This Research Note, its PDF and its analysis CSV permanently freeze the exact 38-month snapshot (revision 88, through 2026-10) supporting the paper.
Files
Sources and identity
- Frozen source projection
- 2026-10-31 · 544 observations
- SHA-256 fingerprint
5e437e3c2de4667ca561fbe97aefc8f2fc38b9d18c6a0b46c6aef70452b2d7b2- Cambridge Open Engage
- 10.33774/coe-2026-t4pz0 ↗ · 18 November 2026
- Harvard Dataverse
- 10.7910/DVN/CGIMZ4 ↗ · 18 November 2026