Recount finds 13 of 21 AI providers publish the EU training-data summary

The authors split the gap two ways: some of it is real drift, and some is their own shallow search ^[2]. They do not say how much of each. The paths the post records show the search half. OpenAI indexes six or more per-model PDFs at help.openai.com ^[9]. xAI lists its Grok summaries on x.ai/legal ^[14], and ByteDance serves its from seed.bytedance.com/transparency ^[13]. Anthropic keeps ten per-model documents in a "Training Data Summaries" section of its Trust Center ^[11]. Google hosts its 8-page Gemini 3 family summary on the AI Transparency Report's GCS bucket ^[10].

One line of the count needs reconciling. The first sweep found 2 providers ^[1]. Ten promotions on that base make 12 ^[21]. The post reaches 13 by adding the ten to three providers it lists as already at T3 before the correction: DeepSeek, Mistral and NVIDIA ^[8][19]. DeepSeek, whose summary file the authors could not resolve the week before ^[17], appears to be the difference.

T3 is the census's tier for a dedicated summary that follows the template's structure ^[7]. It tests shape. Meta's report is, in the authors' own caveat, a thin 2-pager that a third-party tracker grades C/D+ ^[12]. Aleph Alpha's Kolibri 1 summary runs 8 pages, with all three sections substantive, 8 named public datasets and 7 named synthetic-generator models ^[15].

The obligation has applied since 2 August 2025, and pre-existing models have until 2 August 2027 ^[4]. The EU market-placement date sits in the template's general-information section ^[3]. That date decides which deadline a model is under. OpenAI's GPT-5.2 was placed on 11 December 2025 ^[9], inside the live obligation. For the eight providers outside T3 ^[22], a missing summary is overdue only for models placed after 2 August 2025 ^[4].

Scope needs reading as well. The GPT-5.2 document, dated 30 July 2026, says it covers GPT-5.2 and all subsequent releases in the lifecycle ^[9]. Matching a model ID to a summary means reading that clause, since one document can stand for releases that shipped after it was written.

DeepSeek's file was the hardest to pin down. Its CDN is an S3 bucket whose index returns 404 and whose filenames appear to rotate, and the authors found the current file through a third-party tracker's link list ^[17]. For one of the 13, then, the working index to a compliance document was somebody else's list of links. "Lesson: the bucket isn't unstable for the current filename, just for the old ones," the authors wrote ^[18].